A method, system, medium and device for electroencephalogram emotion recognition based on Mamba

Abstract

The application discloses a kind of EEG emotion recognition method, system, medium and equipment based on Mamba, it is related to EEG emotion recognition field, including the feature extraction of EEG electroencephalogram signal feature by differential entropy, corresponding EEG graph node feature is generated;EEG graph node feature is input into pre-trained LEBiGCM_FDA network for electroencephalogram emotion recognition, and emotion recognition result is generated;Wherein, the LEBiGCM_FDA network feature extractor and label classifier;The feature extractor includes LEBiGAT module and Cross-Channel Mamba module;LEBiGAT module carries out information aggregation to the EEG graph node feature of adjacent channel, generates structure perception's local feature sequence, carries out dynamic weighting to local feature sequence, determines node feature sequence;Cross-Channel Mamba module determines the long-distance dependence relationship and multi-scale dynamic change of node feature sequence, generates fused global feature, global feature is converted into fixed dimension global feature vector and input into label classifier, and emotion recognition result is generated.

Description

Technical Field

[0001] This invention relates to the field of EEG emotion recognition, and in particular to a method, system, medium, and device for EEG emotion recognition based on Mamba. Background Technology

[0002] Emotions have a profound impact on human cognition, decision-making, and behavior, reflecting both subjective feelings and objective physiological responses. The remarkable advancements in machine learning herald a new era in sensing, interpreting, simulating, and stimulating human emotions.

[0003] Currently, there are two main types of emotion recognition methods: one based on non-physiological signals, such as facial expressions, body posture, blinking, and speech signals; and the other based on physiological signals, such as electroencephalography (EEG), electrocardiography (ECG), electromyography (EMG), electrooculography (EOG), skin conductance, and respiratory rate. Recognizing emotions through facial expressions among non-physiological signals has always been a key task in emotion computing. For many years, researchers have been dedicated to improving the performance of facial emotion recognition (FER) by designing models with more powerful feature extraction capabilities. For example, MAWitherow proposed the FACE-BE-SELF method, which achieves cross-domain classification of adult-child facial expressions through domain adaptation and beta distribution mixed features. However, neuroscience research shows that physiological signals can more directly measure emotional states than behavioral signals. EEG, as a key physiological signal reflecting brain electrophysiological activity, is widely used in emotion recognition research because it can provide more direct and objective clues for understanding and estimating emotional states.

[0004] Many researchers have proposed various deep learning-based frameworks and methods for EEG emotion recognition. For example, Wu's Graph Orthogonal Cleansing Network (Grop) captures emotion-related and irrelevant features through dual graphs, and then eliminates redundancy through feature projection and cleansing. However, it lacks a dedicated module to mine local dependencies between adjacent channels, relying solely on the global graph structure for feature extraction. This leads to technical problems such as losing fine-grained local discriminative features and weak noise resistance, resulting in low cross-subject domain recognition efficiency for EEG emotion recognition. Additionally, Cui's RACNN captures regional information between physically adjacent channels through a region feature extractor, focusing only on features of local channel combinations without integrating global association information from all channels (including non-adjacent and cross-brain region channels). However, this method neglects cross-regional global collaborative information, has limited feature generalization ability, and cannot capture long-range dynamic dependencies, thus affecting the cross-subject domain recognition efficiency of EEG emotion recognition. Furthermore, these methods all share a key limitation: none explicitly addresses the cross-subject domain adaptation problem. Specifically, in cross-subject experiments, the source and target domain data are independently and identically distributed. Due to individual physiological differences in EEG signals (such as brain structure and impedance), models trained on the source domain may not be applicable to the target domain, leading to a decline in the model's recognition performance. Chen proposed an EEG emotion recognition framework based on brain functional connectivity features and domain adaptive residual convolution (DARC). In his proposed method, maximum mean difference (MMD) loss is applied between the source and target domains before the fully connected layer to minimize the distribution difference between the two and solve the domain shift problem. However, the performance of the MMD used is highly dependent on the choice of kernel function (such as Gaussian kernel or linear kernel), and the kernel type and its parameters (such as bandwidth) usually need to be manually set, lacking adaptability and potentially leading to inaccurate measurement of distribution difference. Summary of the Invention

[0005] This invention provides a method, system, medium, and device for EEG emotion recognition based on Mamba, to solve the aforementioned problems in the prior art, namely, how to improve the efficiency of cross-subject domain recognition in EEG emotion recognition. This invention provides a method for EEG emotion recognition based on Mamba, which includes: Acquire multi-channel EEG signals to be tested, extract features from the EEG signals using differential entropy (DE), and generate EEG graph node features corresponding to each channel. EEG graph node features are input into a pre-trained LEBiGCM_FDA network for EEG emotion recognition to generate emotion recognition results. The LEBiGCM_FDA network includes a feature extractor and a label classifier connected in sequence. The feature extractor includes a LEBiGAT module and a Cross-Channel Mamba module connected in sequence. The LEBiGAT module aggregates information from EEG graph node features of adjacent channels to generate a structure-aware local feature sequence, dynamically weights the local feature sequence, and determines the node feature sequence. The Cross-Channel Mamba module determines the long-distance dependencies and multi-scale dynamic changes of the node feature sequence, generates a fused global feature, and transforms the global feature into a fixed-dimensional global feature vector through max pooling and average pooling. The fixed-dimensional global feature vector is then input into the label classifier to generate the emotion recognition result.

[0006] Optionally, the Cross-Channel Mamba module includes a State Space Model (SSM) and a pooling layer; wherein, the State Space Model (SSM) adopts a state space mechanism to determine the long-distance dependencies and multi-scale dynamic changes of node feature sequences, and generates fused global features; the pooling layer transforms the global features into a fixed-dimensional global feature vector through max pooling and average pooling operations.

[0007] Optionally, a training set can be constructed based on the acquired DEAP or SEED dataset, combined with adversarial examples generated using the Fast Gradient Sign Method (FGSM), and the LEBiGCM_FDA network can be trained using the training set.

[0008] Optionally, the generation of the adversarial examples specifically includes: Based on the acquired DEAP or SEED dataset, extract the target domain input sample x. Then, generate adversarial examples based on the target domain input sample x using the following formula. x adv : ; Where E is the disturbance amplitude coefficient, It is a symbolic function. It is the cross-entropy loss of the domain classifier on the samples. Provide the true labels for the domain classification task.

[0009] Optionally, when training the LEBiGCM_FDA network, the training set is input into the LEBiGCM_FDA network, and the domain discrimination result is obtained through a domain discriminator connected in parallel with the label classifier; the domain discriminator is a Domain Adversarial Neural Network (DANN).

[0010] Optionally, the domain discrimination result includes the source domain identification result and the target domain identification result.

[0011] Optionally, the EEG brain signals can be standardized.

[0012] This invention provides a Mamba-based EEG emotion recognition system, comprising: The acquisition module is used to acquire the multi-channel EEG brain signals to be tested, and to extract features from the EEG brain signals by differential entropy (DE) to generate EEG graph node features corresponding to each channel. A prediction module is used to input EEG graph node features into a pre-trained LEBiGCM_FDA network for EEG emotion recognition to generate emotion recognition results. The LEBiGCM_FDA network includes a feature extractor and a label classifier connected in sequence. The feature extractor includes a LEBiGAT module and a Cross-Channel Mamba module connected in sequence. The LEBiGAT module aggregates information from EEG graph node features of adjacent channels to generate a structure-aware local feature sequence, dynamically weights the local feature sequence, and determines the node feature sequence. The Cross-Channel Mamba module determines the long-distance dependencies and multi-scale dynamic changes of the node feature sequence, generates a fused global feature, and transforms the global feature into a fixed-dimensional global feature vector through max pooling and average pooling. The fixed-dimensional global feature vector is then input into the label classifier to generate the emotion recognition result.

[0013] The present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described Mamba-based EEG emotion recognition method.

[0014] The present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described Mamba-based EEG emotion recognition method.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention provides a Mamba-based EEG emotion recognition method. This method uses differential entropy (DE) to extract features from EEG signals, generating EEG graph node features corresponding to each channel. Then, guided graph representation learning is achieved through local feature embedding of the LEBiGAT module, capturing fine-grained local features in the EEG graph node features. Higher-order aggregation is then used to strengthen the focus on key nodes, upgrading from fixed topology aggregation to adaptive weight allocation, thereby improving the accuracy of local channel combination feature modeling. Simultaneously, the Cross-ChannelMamba module utilizes the State Space Mechanism (SSM) to enhance the dynamic correlation between channels, highlighting the role of key channels and adapting to the dynamic characteristics of multi-channel data, thus effectively improving the cross-subject domain recognition efficiency of EEG emotion recognition. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0017] Figure 1 A flowchart illustrating a Mamba-based EEG emotion recognition method provided in an embodiment of the present invention; Figure 2 The LEBiGCM_FDA network structure diagram provided in this embodiment of the invention; Figure 3 A one-dimensional bidirectional chain diagram of brain regions provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the composition and local convolution of the LEBiGAT module provided in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the working principle of the Cross-Channel Mamba module provided in an embodiment of the present invention. Figure 6 This is a schematic diagram of the FGSM and DANN collaborative training process provided in an embodiment of the present invention; Figure 7 The LOSO experimental results accuracy curve of the DEAP dataset obtained in the embodiments of the present invention and the experimental data of Valence and Arousal for each Subject are shown in the figure. Figure 8 The accuracy curve of the LOSO experimental results of the SEED dataset obtained in the embodiments of the present invention and the experimental data of each Subject are shown. Figure 9 The DE scalp topology electroencephalogram of a 32-bit Subject in the DEAP dataset provided in this embodiment of the invention; Figure 10A bar chart showing the accuracy of ablation experiments using the DEAP / SEED dataset provided in this embodiment of the invention; Figure 11 A schematic diagram of a computer device for a Mamba-based EEG emotion recognition method provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0019] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0020] Figure 1 This is a flowchart of a Mamba-based EEG emotion recognition method provided in an embodiment of the present invention, as shown below. Figure 1 As shown in the figure, this embodiment illustrates a Mamba-based EEG emotion recognition method, including: S1: Acquire the multi-channel EEG signals to be tested, extract features from the EEG signals using differential entropy (DE), and generate EEG graph node features corresponding to each channel.

[0021] For example, differential entropy (DE) is an entropy-based EEG feature extraction method used to quantify the energy distribution complexity of signals within a specific frequency band. Since low-frequency energy in EEG data is higher than high-frequency energy, DE has the ability to distinguish between low-frequency and high-frequency EEG patterns. The original formula for calculating differential entropy is defined as:

[0022] ; in, S Let X be the interval of integration for the random variable X (the domain of the sample values). Let X be the continuous probability density function of the random variable X; In EEG signal processing, it is assumed that the filtered signal... X It follows a Gaussian distribution within a short time window. Its probability density is: ; Substituting this into the differential entropy formula, we can derive the simplified expression for DE: ; It has been proven that for a fixed-length EEG band, the differential entropy is equivalent to the logarithmic energy spectrum of a certain frequency band. Therefore, the differential entropy can be calculated in five frequency bands (delta: 1-4 Hz, theta: 4-8 Hz, alpha: 8-13 Hz, beta: 13-30 Hz, gamma: 30-45 Hz).

[0023] S2: Input the EEG graph node features into a pre-trained LEBiGCM_FDA network for EEG emotion recognition to generate emotion recognition results; wherein, the LEBiGCM_FDA network includes a feature extractor and a label classifier connected in sequence; the feature extractor includes a LEBiGAT module and a Cross-Channel Mamba module connected in sequence; the LEBiGAT module aggregates information from the EEG graph node features of adjacent channels to generate a structure-aware local feature sequence, dynamically weights the local feature sequence, and determines the node feature sequence; the Cross-Channel Mamba module determines the long-distance dependencies and multi-scale dynamic changes of the node feature sequence, generates fused global features, and transforms the global features into a fixed-dimensional global feature vector through max pooling and average pooling; the fixed-dimensional global feature vector is input into the label classifier to generate emotion recognition results.

[0024] For example, Figure 2 The experimental framework includes a feature extractor combining LEBiGAT and Cross-Channel Mamba, and domain adaptation using FDA (Fast Gradient Signed Method-Domain Adversarial Neural Network Co-optimization Module). Input is EEG data from the DEAP or SEED dataset. Data preprocessing includes EEG segmentation, DE value calculation, splitting the training and test sets, standardization, and FGSM generation of adversarial examples. After the source and target domains are processed by the LEBiGAT and Cross-Channel Mamba feature extractors, domain adaptation is performed using a DANN network, and finally, labels are classified into two classes.

[0025] For example, in spatial topology modeling, establishing edge connections between all channel nodes introduces a large number of redundant cross-channel interactions, increasing the computational complexity of the model and potentially introducing noise interference, thereby weakening the ability to model local dependencies. Figure 3 As shown, this invention employs a linear chain topology for graph construction. Specifically, it is assumed that the EEG data contains... N One channel Then, based on the spatial location of brain regions, adjacent pathway nodes are... and Connect using bidirectional directed edges, i.e.:

[0026] ; This structure forms a one-dimensional bidirectional linked graph, where each node is connected to only its two left and right neighbors, except for the nodes at both ends. The advantages of this linked graph are: (a) it maintains the connectivity of the graph structure, allowing information to propagate sequentially along the channels; (b) it explicitly limits the number of neighbors for each node, highlighting local spatial dependencies and reducing the noise impact of irrelevant edges; and (c) compared to fully connected graphs, the linked structure significantly reduces the number of edges, improving computational efficiency. It is particularly suitable for embedding local features in the early stages of the model, providing a cleaner and more structured input representation for subsequent global feature modeling.

[0027] like Figure 2 As shown, the local embedding module is implemented based on GCN. Its core idea is to utilize the adjacency relationships between EEG channels to aggregate and encode features of each channel and its local neighborhood in a graph structure space. The node input feature matrix... Where N is the number of channels and F is the feature dimension; the EEG channel connectivity is represented by the adjacency matrix A, and is obtained by introducing self-loops. The local embedding process is characterized by the following propagation formula:

[0028] ; in, for The degree matrix; H (l) This represents the feature representation of the l-th layer nodes; W (l) The weight matrix is ​​a learnable weight matrix; This is a non-linear activation function. The introduction of this normalized adjacency matrix can effectively alleviate the numerical instability caused by the difference in degree between different nodes, thereby ensuring the stability and robustness of feature propagation. In this process, the new features of each node not only retain their own original feature information, but also integrate semantic information from their first-order neighbor nodes, thus achieving efficient feature embedding within the local structure. After the local embedding module, this invention uses the ReLU activation function to suppress overfitting.

[0029] like Figure 3 As shown, the one-dimensional bidirectional linked graph (optimized 10 / 20 system layout) of the 32-channel EEG in the DEAP dataset is a channel association topology constructed based on the spatial location of brain regions, providing adjacency relationship input for the LEBIGAT module.

[0030] For example, for the 32 EEG channels of the DEAP dataset, this invention first divides them into five brain regions—frontal, central, parietal, occipital, and temporal—using a 10 / 20 system. Then, based on the physical spatial adjacency of the channels, a bidirectional chain association is established: Taking the frontal lobe channel Fp1 as an example, it forms a bidirectional connection with the adjacent frontal lobe channel Fp2 and the central lobe channel Cz in the chain diagram; while the temporal lobe channel T7 establishes a bidirectional connection with the adjacent central lobe channel C3 and the parietal lobe channel P3.

[0031] For the adjacency matrix A∈RN×N (N=32 is the number of DEAP channels) corresponding to the chain graph, if channel i and channel j are adjacent nodes, then Ai, j=1, otherwise Ai, j=0.

[0032] This adjacency matrix will serve as input to the local feature embedding in the LEBIGAT module, used for the initial aggregation of local features from adjacent channels: hi(0) = ∑j∈N(i)Ai, j·xj; Where xj is the original feature of channel j, N(i) is the set of neighboring nodes of channel i, and hi(0) is the initial local feature of the local feature embedding output.

[0033] For the 62 EEG channels in the SEED dataset, this invention divides brain regions using the same logic, establishes bidirectional chain associations between adjacent channels, and generates a corresponding adjacency matrix A∈R62×62, ensuring the consistency of channel topology under different datasets and providing a unified spatial association basis for subsequent feature extraction by LEBiGAT.

[0034] Generally, GAT is a neural network model specifically designed for graph-structured data, first proposed by Veličković et al. in 2018. Unlike traditional GCNs, which require pre-computing a normalized adjacency matrix, GAT introduces a self-attention mechanism, enabling each node to dynamically assign different weights to different neighbors when aggregating neighbor information.

[0035] This invention introduces a two-layer GAT network. On the one hand, it can capture second-order neighbor relationships by focusing attention layer by layer on the first-order neighborhood, thereby expanding the receptive field of nodes and improving the diversity of feature representation. On the other hand, the nonlinear transformation between layers enables the model to achieve adaptive modeling of complex neighborhood dependencies while maintaining local smoothness. Compared with a single-layer structure, it is more conducive to obtaining robust and discriminative node representations.

[0036] like Figure 4As shown, taking channel F7 as an example, when running the LEBiGAT module, channel F7 performs local convolution between adjacent channels with Fp2 and F3.

[0037] For example, for the input node feature matrix: , ,in, A collection of full-channel EEG features; Let F be the one-dimensional feature vector corresponding to the i-th electrode channel, i=1,2…N, where N is the number of channels. in The dimension of each node's feature (input dimension). In the first layer of the GAT network, this invention uses multi-head attention. Multi-head attention is an extension of the self-attention mechanism, used to enhance the model's expressive power across different "attention subspaces". Its core idea is to project the input feature vector onto multiple subspaces, calculate the attention separately, and then concatenate the outputs of each head to capture more diverse association patterns.

[0038] The formula for calculating each attention head k individually is as follows: ; in, This represents a learnable weight matrix; This represents the output dimension of each head. The core of GAT lies in the fact that, unlike traditional graph convolution which uses all adjacent neighbors of a node as input, it computes a learnable attention coefficient for each edge in the graph. , This is used to represent the "degree of attention" that node i pays to node j when aggregating neighbor information:

[0039] ; in, It is a learnable weight vector, and || represents feature concatenation. This is a non-linear activation function. Then, Softmax is used to normalize the attention coefficients.

[0040] ; in, For nodes i The set of neighboring nodes, Let be the feature transformation weight matrix of the k-th layer graph convolution. Let i be the index of any adjacent node in the neighborhood of node i.

[0041] Next, the present invention aggregates neighbor node information: ; in, For activation function, It refers to the first i The EEG electrode channel (graph network node) passes through the first k Layer graph attention aggregation and updated node features; This refers to the neighboring node j Point to target node i The k Normalized attention weight coefficients learned by the layered graph attention network; It refers to the first j The first neighboring EEG channel (node) corresponding to the first k The original input feature vector of the layer network.

[0042] Finally, the outputs of the multiple heads are concatenated to obtain the output features of the first layer: ; Figure 2 In the first layer of GAT convolution, overfitting is suppressed and convergence is accelerated through ELU and Dropout. The second layer of GAT structure is similar to the first layer, but adopts a single-head attention mechanism to simplify the network structure and reduce the feature dimensionality. Specifically, the second layer of GAT takes the high-dimensional features output from the first layer as input, and maps the node features to a high-dimensional space through linear transformation. Then, an attention mechanism is used to calculate the weights between nodes, aggregate the feature information of neighboring nodes, and further enhance the nonlinear expressive power through the ELU activation function. Unlike the multi-head parallel computation of the first layer, the single-head structure of the second layer effectively controls the model complexity and overfitting risk while maintaining the feature representation capability, generating a more compact and representative node feature representation.

[0043] For example, Cross-Channel Mamba is a deep learning architecture proposed in recent years specifically for efficient modeling of long sequences. Its core idea lies in combining structured state-space models with hardware-aware, efficient computation. The core idea of ​​the Cross-Channel Mamba module is based on continuous-discrete state-space equations, where the input sequence is... Where T represents the sequence length, d If the sequence dimension is used, then the state equation is: ;

[0044] For example, the Cross-Channel Mamba module includes a state space layer and an output layer; wherein, the state space layer uses a state space mechanism to dynamically weight local feature sequences to determine long-distance dependencies and multi-scale dynamic changes, and generates fused global features; the output layer generates a fixed-dimensional global feature vector through max pooling and average pooling.

[0045] Figure 5 In the Mamba principle, in the formula, Let be the input signal at time t (the current time step). Let t-1 be the state vector (the previous time step). Let be the updated state vector at time t (current time step). for t The output vector at time step (current time step). The Cross-Channel Mamba module innovatively utilizes discretization techniques, equating its computation process to a global convolutional kernel. This enables the model to possess both powerful long-range dependency modeling capabilities and excellent efficiency in large-scale sequence processing. Figure 5 In the Mamba principle, A is the state transition matrix, B is the input projection, C is the output projection, and D is the input direct-connection projection matrix. In the cross-channel Mamba workflow, the calculated contrastive loss is ultimately directly weighted into the total loss function. The cross-channel Mamba implementation illustrates how the experiment uses the EEG channel as the time step input to the Cross-Channel Mamba module, with each channel... express t The input vector at time (current time step).

[0046] In the previous LEBiGAT, this invention focused on the relationship between adjacent nodes. Figure 5 The cross-channel Mamba implementation in the model treats the feature sequences of each channel node as a one-dimensional temporal sequence. It captures long-distance dependencies and multi-scale dynamic changes in one go through efficient state-space convolution. This not only makes up for the problem of insufficient integration of remote node information in LEBiGAT, but also has a natural smoothing and suppression effect on band noise in EEG signals. At the same time, due to its linear complexity and parallel convolution implementation, it can model the global features of dozens of channels with low memory overhead while ensuring training and inference efficiency. This makes the model obtain a more robust and generalizable spatiotemporal fusion representation in cross-subject scenarios.

[0047] Once LEBiGAT has completed capturing the neighborhood information of each node, the node features are input into the Cross-Channel Mamba model in tensor form. Figure 5In the cross-channel Mamba workflow, the module first performs layer normalization on the vector of each node to ensure that features of different dimensions are distributed on the same scale, avoiding difficulties in convergence of the subsequent state space model. Next, the Cross-Channel Mamba state space model performs parallel scanning of the temporal features of the entire batch, feeding the normalized node sequence into its internal recursive equations. The SSM core captures the long-distance dependencies of node sequences in each graph structure through a learnable state transition matrix and input gating mechanism, and completes the computation in linear time complexity with hardware-friendly vectorized operations. Finally, let the output of this invention after passing through the Cross-Channel Mamba module be... ,in, B For batch size, N The number of channels in each sample graph and F represent the feature dimension of each node. After the model outputs, this invention applies max pooling and average pooling along the node dimension (dim=1) to generate a global representation of the graph, which is then passed to a fully connected layer for classification. After the Cross-Channel Mamba module generates the global representation of the graph, this invention adds a contrastive loss. The contrastive loss is used to enhance the discriminativeness of the feature representations learned by the model, bringing samples of the same class closer together and widening the distance between samples of different classes, thereby improving the model's discriminative and generalization abilities. The design goal of the Cross-Channel Mamba module is to leverage the long-range memory advantage of the state-space model to preserve sequence information, while continuously optimizing the graph-level representation through contrastive loss and normalization strategies. This allows it to take into account both static spatial structure and dynamic temporal evolution, ultimately providing highly compact and robust features for subsequent classifiers.

[0048] For example, the domain discriminator of the present invention specifically includes: a Domain Adversarial Neural Network (DANN); and performs domain determination on a fixed-dimensional global feature vector using the DANN.

[0049] Generally, in the field of deep learning, improving the generalization ability and adversarial robustness of models in cross-domain scenarios has always been a key challenge. DANNs, by introducing an adversarial training mechanism between the feature extractor and the domain discriminator, can learn domain-invariant feature representations, thereby enhancing the model's recognition performance in unknown target domains. However, when there are significant distributional differences between the source and target domains, traditional DANN methods still suffer from insufficient generalization performance. Therefore, this invention proposes combining FGSM with the DANN framework to further improve robustness and generalization performance in cross-subject emotion recognition tasks.

[0050] Illustratively, according to the loss function of the domain classifier, the present invention uses FGSM to apply perturbation to target domain samples to generate adversarial samples. The core of this strategy is to actively construct samples that are "most deceptive" to the domain classifier, accurately expose the regions with the most significant domain differences in the current feature space, and use this as a strong driving signal to force the feature extractor to extract more essential domain-invariant features. Compared with traditional distribution alignment methods, this fine training mechanism based on adversarial perturbation can achieve tighter domain alignment at the decision boundary, thereby significantly enhancing the generalization ability and robustness of the model in the target domain.

[0051] Specifically, given a target domain input sample x, its adversarial sample x adv is generated by the following formula: ; wherein E is the perturbation amplitude coefficient. In the present invention, the value range of the perturbation amplitude coefficient E is defined as 0 < E < 0.1, and 0.05 is a preferred value. When E is greater than 0.1, an excessively large perturbation amplitude will tamper with the original physiological features of EEG, losing the data validity of the adversarial sample; when E approaches 0, the perturbation fails, and effective adversarial samples cannot be generated; is a sign function; is the cross-entropy loss of the sample by the domain classifier, is the ground-truth label for the domain classification task, which is used to indicate the domain (source domain or target domain) that each data sample belongs to. In the experiment, the ground-truth label of the domain is set to 1.

[0052] The fused model framework comprises three core modules: a feature extractor, a label classifier and a domain discriminator, as shown in Figure 6 By adopting the FGSM algorithm to generate adversarial samples, both source domain and target domain samples are fed into the feature extractor. For the DANN network, the feature vectors obtained after the original data passes through the feature extractor are respectively sent to the label classification module to calculate classification loss and to the domain classification module to calculate domain loss. The gradient of the classification loss function with respect to the model output is calculated and back-propagated layer by layer via the chain rule, while the domain loss function is back-propagated through GRL. During back propagation, the gradient is normally back-propagated to the domain discriminator, and the inverted gradient (multiplied by a negative coefficient) is back-propagated to the feature extractor. Dropout is set to 0.5 in the experiment.

[0053] wherein, the feature extractor comprises a LEBiGAT module and a Cross-Channel Mamba module, which are configured to extract high-order spatial features; the label classifier is configured to complete emotion category discrimination; the domain discriminator This is achieved by using a GRL (Generative Adversarial Language) structure to create an adversarial relationship with the feature extractor, enabling cross-domain feature alignment. During training, source and target domain data are fed into the model separately to obtain their respective domain recognition results. These two sets of results are then merged, and a "target domain" label is constructed (0 for the source domain and 1 for the target domain). The difference between the predicted result and the "target domain" label is calculated using the cross-entropy loss function; this is the domain loss. In implementation, as shown... Figure 6 The GRL layer is used, which does not change the input features during forward propagation, but simply passes them as is to the domain classifier:

[0054] y = x; Here, x represents the input feature, which in the experiment is the output of the feature extractor; y is the output of GRL during forward propagation, which remains unchanged. During backpropagation, GRL reverses the sign of the gradients returned to the feature extractor and adjusts the gradient strength through dynamic weight coefficients, thereby enabling the model to learn domain-invariant features. The adjustments employ a dynamic strategy proposed by Ganin et al. in DANN to gradually increase the difficulty and effectiveness of domain adversarial training.

[0055] ; Where L is the total model loss ( ), For sentiment classification loss, Loss is determined by the domain; Representative feature extractor All learnable parameters; These are the domain-adversarial gradient reversal weight coefficients. Among them,

[0056] ; in, epoch This indicates the current training iteration number. epochs This is the total number of training rounds. In the initial phase of the experiment, A value close to 0 indicates a weaker adversarial task, with the model focusing more on task classification learning from the source domain data. As the training process progresses... The value of gradually increases until it approaches 1, which enhances the gradient reversal strength during the backpropagation stage. This makes it increasingly difficult for the domain classifier to distinguish the feature representations of the source and target domains, ultimately achieving a domain feature mixture effect, thereby improving the model's generalization performance in the target domain. By combining FGSM and DANN, this invention not only improves the feature extractor's generalization ability across subject scenarios but also enhances the diversity of the target domain using domain-level adversarial examples, thus significantly improving the robustness and performance of the emotion recognition model in cross-domain tasks.

[0057] For example, this invention designs a three-part loss function to simultaneously guide the model's collaborative optimization in emotion classification, domain adversarial learning, and contrastive learning. Its form can be written as:

[0058] ; in, It is the total loss function. It is classification loss. It is a domain loss. It is a comparative loss.

[0059] The classification loss is obtained using the following formula. : ; in, y i It is a sample i The true label (0 or 1). It is the model on the sample i The predicted probability, N This refers to the batch sample size.

[0060] The neighborhood loss is obtained using the following formula. : ; in, It is the number of samples in the source domain. It is the number of samples in the target domain. It is a sample i The probability of being predicted as the source domain. It is a sample j The probability of being predicted as the source domain.

[0061] The contrast loss is obtained using the following formula. : ; in, B It refers to the batch size. z i It is a sample i The normalized eigenvectors, This is the temperature parameter (set to 0.5).

[0062] Emotion Classification Loss This term is the standard cross-entropy loss, used to measure the model's accuracy in classifying sentiment on samples in the source domain (training set). By comparing the probability distribution output by the model with the true labels, the gradient feedback allows the classifier to continuously adjust its weights. Domain loss It is mainly used to reduce the difference in feature distribution between the source and target domains. The weighting coefficient for the domain adversarial loss controls the degree of influence of the domain adversarial module on the total loss; it is a configurable hyperparameter. (Comparative Loss) The aim is to enhance the discriminative power of the feature space. By narrowing the distance between samples of the same class and widening the distance between samples of different classes, the contrastive loss further improves the model's ability to distinguish emotion categories. It is the weighting coefficient of the contrastive loss, which controls the degree of influence of the contrastive learning loss on the total loss.

[0063] For example, the present invention was validated on the DEAP and SEED datasets. The DEAP dataset contains 32-channel EEG signals (sampling rate 128Hz) and 8-channel peripheral physiological signals from 32 subjects watching 40 emotion-inducing videos. The baseline signal for the first 3 seconds was removed. For the experimental signals of the last 60 seconds of the 32 EEG channels, the signals were segmented according to a 10-second sliding window and a bandpass filter of 1.0–45.0Hz was applied to preserve emotion-related frequency bands. Emotion labels were based on four dimensions of subjective rating (pleasure, arousal, dominance, and liking, 1–9 points). The present invention focuses on binary classification of Valence and Arousal: a rating >5 is "high" (1), and ≤5 is "low" (0), constructing two independent classification tasks. The SEED dataset contains 62-channel EEG signals (sampling rate 200Hz) from 15 subjects watching 15 emotion-inducing videos (4 minutes each) three times at different times. For the SEED dataset, the experimental data of approximately 4 minutes of the 62 channels were segmented according to a 10-second sliding window. Unlike DEAP, SEED's emotion labels are based on fixed video content (not subjective ratings), and this invention only uses data from the first experiment with labels of negative and positive. By comparing the two datasets, the model's emotion recognition performance under different data distributions (subjective ratings vs. fixed labels, different number of channels and sampling rates) can be evaluated.

[0064] Based on the dataset partitioning method, this invention obtained 7680 32×5 samples from the DEAP dataset and 3345 62×5 samples from the SEED dataset. The experiments of this invention were implemented using the PyTorch framework and the Cross-Channel Mamba model, trained on an NVIDIA 3060 using the Adam optimizer. The epochs and learning rate were 50 and 0.001, respectively. The batch size for the DEAP dataset was 128, and the batch size for the SEED dataset was 32. When generating adversarial examples using FGSM, the perturbation amplitude coefficient E used in this invention can be 0.1. When calculating the total loss function, the weight coefficients for both the domain adversarial loss and the contrastive loss in this invention are 0.1.

[0065] For example, this invention can evaluate the model's emotion recognition performance across subject scenarios by employing a rigorous experimental paradigm of Leave-One-Side Cross-Validation (LOSO). LOSO cross-validation effectively avoids leakage of samples from the same subject, rigorously evaluates the model's cross-subject generalization ability, and better reflects its robustness and adaptability in real-world application scenarios. This validation method sequentially uses all EEG data of each subject as an independent test set, and the data of all other subjects as a training set, thereby ensuring that the test results are completely unaffected by individual subject data and can truly reflect the model's generalization ability. Tables 1 and 2 show the experimental results of the Fast Gradient Symbolic Method-Domain Adversarial Neural Network (FAR) jointly optimized by a locally embedded two-layer graph attention network and cross-channel Mamba optimization on the DEAP and SEED datasets, respectively.

[0066] In emotion recognition experiments on the DEAP dataset (32 channels, 32 subjects), the method of this invention performed excellently: achieving a classification accuracy of 62.16% in the Valence dimension (only 0.34% lower than the best method, Variational Mode Decomposition + Deep Neural Network), and ranking first in the Arousal dimension with an accuracy of 66.94% (0.1% higher than the second-place Residual Learning Capsule Network). Figure 7 As shown in Table 1, the accuracy curves and individual data of 32 subjects cross-validated using LOSO demonstrate that the method of this invention exhibits stable cross-subject recognition performance. Compared with five traditional methods (including convolutional neural networks on continuous wavelet transform scale maps, convolutional neural networks-recurrent neural networks-generative adversarial networks, spatiotemporal feature maps + capsule networks, etc.), the model of this invention performs best in the Arousal classification task and also achieves highly competitive results (ranked second) in the Valence classification task. This verifies the effectiveness and robustness of the method in emotion dimension recognition, as shown in Table 1. Particularly noteworthy is that while maintaining high accuracy, the performance fluctuations among subjects are small, indicating that the model has good generalization ability.

[0067] Table 1. LOSO accuracy of the DEAP dataset In emotion recognition on the SEED dataset (62 channels, 15 subjects), this invention achieved an average recognition accuracy of 80.39%. Figure 8 The table shows the accuracy curves for emotion recognition in the LOSO experiment for 15 subjects and the accuracy data for each subject. For the SEED dataset, 10 different methods were compared, and the results are shown in Table 2. The locally embedded two-layer graph attention network combined with the cross-channel Mamba-optimized fast gradient symbolic method-domain adversarial neural network of this invention achieved the highest accuracy and the best standard deviation. The second best result was achieved by the Emotion Transformer network, with an accuracy of 80.2% and a standard deviation of 9.30%, which is 0.19% lower than the accuracy of this invention. However, the method of this invention improved the standard deviation by 1.38% (the standard deviation of the method of this invention is 7.92%). The experimental results show that the locally embedded two-layer graph attention network combined with the cross-channel Mamba-optimized fast gradient symbolic method-domain adversarial neural network not only has a higher recognition accuracy but also the smallest standard deviation, indicating that the framework of this invention is more stable in experimental results on different subjects, has stronger generalization ability, and can be more reliably applied to cross-subject emotion recognition tasks in real-world scenarios. This advantage may stem from the better adaptability to individual differences and robustness of feature extraction of the fast gradient symbolic method-domain adversarial neural network model jointly optimized by the locally embedded two-layer graph attention network and cross-channel Mamba optimization, thereby significantly improving the stability and accuracy of emotion recognition.

[0068] Table 2 LOSO accuracy of the SEED dataset like Figure 9As shown, the visualization displays the spatial distribution patterns of normalized differential entropy at different electrode locations for 32 subjects in the DEAP dataset, presenting the overall neural activity characteristics of the emotional dimension (such as arousal or valence). Significant differences in EEG activity among different subjects can be observed in the figure. Some subjects exhibited higher entropy values ​​in the frontal lobe regions (such as Fp1 and Fp2), with entropy values ​​close to 1.0 in the figure, possibly reflecting prefrontal activation in emotional processing. The entropy distribution in the central regions (such as Cz, C3, and C4) and parietal regions (such as Pz) showed greater variability, suggesting heterogeneity in emotion regulation strategies among individuals. The color bars indicate that the entropy values ​​were normalized to 0 to 1. Low entropy values ​​(the darker areas with lower entropy values ​​in the figure) may correspond to synchronized neural activity or a low-complexity state, while high entropy values ​​(the darker areas with higher entropy values ​​in the figure) may represent more complex neural dynamics. It is worth noting that the heatmaps of some subjects showed asymmetry between the left and right hemispheres (such as the difference in entropy between T7 in the left temporal lobe and T8 in the right temporal lobe), which may be related to the lateralization processing mechanism of emotional valence.

[0069] For example, to evaluate the contribution of each key module in the proposed model to the overall performance, this invention conducted systematic ablation experiments on the DEAP and SEED datasets, progressively removing three types of core components (LEBiGAT module, Cross-Channel Mamba module, and DA module), and comparing the results with the complete model. On the DEAP dataset (e.g., ... Figure 10 As shown in the figure, the complete model achieved accuracies of 62.16% and 66.94% in the Valence and Arousal tasks, respectively. Removing the LEBiGAT module reduced the accuracy to 57.31% (a decrease of 4.85%) and 61.45% (a decrease of 5.49%), respectively; removing the Cross-Channel Mamba module further reduced the accuracy to 55.52% (a decrease of 6.64%) and 58.85% (a decrease of 8.09%); removing the DA module resulted in accuracy drops to 56.41% (a decrease of 5.75%) and 52.68% (a decrease of 14.26%).

[0070] Figure 10 The paper also presents the accuracy of binary classification on the SEED dataset obtained by removing the LEBiGAT, Cross-Channel Mamba, and DA modules in sequence. It can be seen that compared to the complete model, the accuracy of the three ablation experiments of this invention decreased by 18.55%, 11.86%, and 5.87%, respectively.

[0071] Ablation experiments fully validate the importance and irreplaceability of each component module in the proposed model for emotion recognition tasks. In the binary classification task on the DEAP dataset, removing any module significantly reduces the accuracy of both Valence and Arousal classifications. The DA module is particularly crucial in improving cross-subject generalization performance, demonstrating its significant role in suppressing distributional biases caused by individual differences and enhancing model robustness. Furthermore, from... Figure 10 As can be seen, in the binary classification task of the SEED dataset, the complete model consistently achieves the highest classification accuracy, while the performance degrades to varying degrees after removing any module. These results not only verify the effectiveness of each module in feature modeling and information fusion, but also further illustrate the significant role of the synergistic effect between modules in improving overall performance.

[0072] Cross-validation using LOSO showed that the model achieved average classification accuracies of 62.16% and 66.94% for Valence and Arousal on the DEAP dataset (32 subjects, audiovisual emotion-induced scenarios), with standard deviations of 0.0545 and 0.0848, respectively. On the SEED dataset (15 subjects, complex emotion scenarios), the average accuracy was 80.39%, with a standard deviation of 0.0792. This indicates that the model can capture both the common EEG features generated by synchronized video stimuli in DEAP and the more significant individual differences in emotional responses in SEED, demonstrating good cross-dataset adaptability. The visualization results provide intuitive support for the distribution of EEG features. In the scalp topology heatmaps of the DEAP dataset, the spatial distribution of DE features among different subjects exhibits a clear pattern: under high arousal scenarios, the prefrontal cortex (e.g., Fp1, Fp2 channels) often shows high DE values, appearing as dark areas with entropy values ​​close to 1.0 in the graph, consistent with the physiological mechanism of the prefrontal cortex's involvement in emotion regulation; the occipital region (e.g., O1, O2 channels) maintains a consistently high DE value due to the active background of visual stimulation, reflecting the fundamental role of the visual cortex in inducing audiovisual emotions. Simultaneously, significant differences in the morphology of the heatmaps among the 32 subjects were observed, with some subjects showing asymmetry between the left and right hemispheres (e.g., the difference in DE values ​​between the left temporal lobe T7 and the right T8), suggesting a lateralization mechanism in emotional valence processing, providing visual evidence for the specificity of individual emotional responses. The channel-frequency matrix heatmap of the SEED dataset focuses on dimensional synergy: the Gamma band (30-45Hz) generally has higher DE values ​​in the frontotemporal channels (such as F3 and F4 channels), which may be related to the capture of dynamic emotional changes; the Alpha band (8-13Hz) shows a significant increase in DE values ​​in the occipital channels during periods of emotional calm, consistent with the classic conclusion of "enhanced Alpha waves in a relaxed state." Some subjects exhibited a characteristic extreme value combination of "frontal Beta band + occipital Alpha band," reflecting the synergistic effect of different brain regions and frequency bands in emotion processing, providing a basis for strengthening dimensional correlations in the model. Ablation experiments further revealed the importance and synergistic relationship of each module. The LEBiGAT module plays a core role in capturing local spatial correlations between EEG channels, highlighting key channel features related to emotion through weighted aggregation. After removing LEBiGAT, the model accuracy of DEAP on the Valence and Arousal datasets decreased by 4.85%, 5.49%, and 18.55%, respectively, compared to the SEED dataset. The Cross-Channel Mamba temporal modeling module focuses on capturing global correlations of EEG channels. The domain adaptation module effectively mitigates distributional differences across subjects and datasets. In DEAP, where individual differences are more significant, the accuracy decreased by 5.75% and 14.26% respectively after removing domain adaptation, indicating that its generalization ability to unfamiliar subject data is crucial.The synergistic optimization of these three technologies is particularly crucial: the spatial features extracted by LEBiGAT and Cross-Channel Mamba provide a more representative transfer basis for DA, making domain adaptation more efficient.

[0073] Inspired by the spatial topological characteristics and neural temporal dynamics of EEG signals, this invention proposes the LEBiGCM_FDA method. This method uses a locally embedded two-layer graph attention network to capture channel-specific spatial dependencies, combined with a cross-channel Cross-ChannelMamba module to model global temporal associations. It leverages domain adaptation mechanisms to mitigate cross-subject distribution differences, providing a novel solution to the challenge of cross-subject generalization in EEG emotion recognition. This method significantly improves recognition performance on the DEAP and SEED datasets. Furthermore, through EEG heatmap visualization, it visually presents brain region functional connectivity patterns and frequency band temporal feature distributions, further revealing the roles of key brain regions such as the prefrontal cortex and frequency bands such as Gamma and Alpha in emotion processing, providing empirical evidence for understanding the connection mechanism between EEG and emotion.

[0074] The above are one or more embodiments of the Mamba-based EEG emotion recognition method provided in this specification. Based on the same idea, this specification also provides a corresponding Mamba-based EEG emotion recognition system, including: The acquisition module is used to acquire the multi-channel EEG brain signals to be tested, and to extract features from the EEG brain signals by differential entropy (DE) to generate EEG graph node features corresponding to each channel. A prediction module is used to input EEG graph node features into a pre-trained LEBiGCM_FDA network for EEG emotion recognition to generate emotion recognition results. The LEBiGCM_FDA network includes a feature extractor and a label classifier connected in sequence. The feature extractor includes a LEBiGAT module and a Cross-Channel Mamba module connected in sequence. The LEBiGAT module aggregates information from EEG graph node features of adjacent channels to generate a structure-aware local feature sequence, dynamically weights the local feature sequence, and determines the node feature sequence. The Cross-Channel Mamba module determines the long-distance dependencies and multi-scale dynamic changes of the node feature sequence, generates a fused global feature, and transforms the global feature into a fixed-dimensional global feature vector through max pooling and average pooling. The fixed-dimensional global feature vector is then input into the label classifier to generate the emotion recognition result.

[0075] Specific limitations regarding the Mamba-based EEG emotion recognition system can be found in the above section on the limitations of the Mamba-based EEG emotion recognition method, and will not be repeated here. Each module in the aforementioned Mamba-based EEG emotion recognition system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0076] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the Mamba-based EEG emotion recognition method described above.

[0077] The present invention also provides Figure 11 The schematic diagram of the computer device shown is as follows: Figure 11 As shown, at the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to implement the Mamba-based EEG emotion recognition method provided in the above embodiments.

[0078] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.

Claims

1. A method for EEG emotion recognition based on Mamba, characterized in that, include: Acquire multi-channel EEG signals to be tested, extract features from the EEG signals using differential entropy (DE), and generate EEG graph node features corresponding to each channel. EEG graph node features are input into a pre-trained LEBiGCM_FDA network for EEG emotion recognition to generate emotion recognition results. The LEBiGCM_FDA network includes a feature extractor and a label classifier connected in sequence. The feature extractor includes a LEBiGAT module and a Cross-Channel Mamba module connected in sequence. The LEBiGAT module aggregates information from EEG graph node features of adjacent channels to generate a structure-aware local feature sequence, dynamically weights the local feature sequence, and determines the node feature sequence. The Cross-Channel Mamba module determines the long-distance dependencies and multi-scale dynamic changes of the node feature sequence, generates a fused global feature, and transforms the global feature into a fixed-dimensional global feature vector through max pooling and average pooling. The fixed-dimensional global feature vector is then input into the label classifier to generate the emotion recognition result.

2. The Mamba-based EEG emotion recognition method as described in claim 1, characterized in that, The Cross-ChannelMamba module includes a State-Space Model (SSM) and a pooling layer. The SSM uses a state-space mechanism to determine the long-distance dependencies and multi-scale dynamic changes of node feature sequences, generating fused global features. The pooling layer transforms the global features into a fixed-dimensional global feature vector through max pooling and average pooling operations.

3. The Mamba-based EEG emotion recognition method as described in claim 1, characterized in that, Based on the acquired DEAP or SEED dataset, and combined with adversarial examples generated using the Fast Gradient Sign Method (FGSM), a training set is constructed, and the LEBiGCM_FDA network is trained using the training set.

4. The Mamba-based EEG emotion recognition method as described in claim 3, characterized in that, The generation of the adversarial examples specifically includes: Based on the acquired DEAP or SEED dataset, extract the target domain input sample x. Then, generate adversarial examples based on the target domain input sample x using the following formula. x adv : ; Where E is the disturbance amplitude coefficient, It is a symbolic function. It is the cross-entropy loss of the domain classifier on the samples. Provide the true labels for the domain classification task.

5. The Mamba-based EEG emotion recognition method as described in claim 3, characterized in that, When training the LEBiGCM_FDA network, the training set is input into the LEBiGCM_FDA network, and the domain discrimination result is obtained through a domain discriminator connected in parallel with the label classifier; the domain discriminator is a Domain Adversarial Neural Network (DANN).

6. The Mamba-based EEG emotion recognition method as described in claim 5, characterized in that, The domain discrimination result includes the source domain identification result and the target domain identification result.

7. The Mamba-based EEG emotion recognition method as described in claim 1, characterized in that, The EEG brain signals were standardized.

8. A Mamba-based EEG emotion recognition system, characterized in that, include: The acquisition module is used to acquire the multi-channel EEG brain signals to be tested, and to extract features from the EEG brain signals by differential entropy (DE) to generate EEG graph node features corresponding to each channel. A prediction module is used to input EEG graph node features into a pre-trained LEBiGCM_FDA network for EEG emotion recognition to generate emotion recognition results. The LEBiGCM_FDA network includes a feature extractor and a label classifier connected in sequence. The feature extractor includes a LEBiGAT module and a Cross-Channel Mamba module connected in sequence. The LEBiGAT module aggregates information from EEG graph node features of adjacent channels to generate a structure-aware local feature sequence, dynamically weights the local feature sequence, and determines the node feature sequence. The Cross-Channel Mamba module determines the long-distance dependencies and multi-scale dynamic changes of the node feature sequence, generates a fused global feature, and transforms the global feature into a fixed-dimensional global feature vector through max pooling and average pooling. The fixed-dimensional global feature vector is then input into the label classifier to generate the emotion recognition result.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the Mamba-based EEG emotion recognition method according to any one of claims 1-7.

10. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the Mamba-based EEG emotion recognition method according to any one of claims 1-7.

Images