Electroencephalogram emotion recognition model based on dynamic electroencephalogram channel selection and Conformer structure

Abstract

The invention relates to an electroencephalogram emotion recognition model based on dynamic electroencephalogram channel selection and a Conformer structure, which adopts a series four-module architecture: firstly, performing Z-Score standardization preprocessing on an electroencephalogram (EEG) signal, and screening out an optimal channel subset which is 3 / 4 of the total channel number through an mRMR algorithm; extracting local space-time features through a convolution module, integrating global context information through a self-attention module, and finally completing emotion category mapping through a classification module. According to the model, 98.69% of average recognition precision and 98.47% of F1-score are achieved on a DEAP data set, 90.54% of average recognition precision and 90.48% of F1-score are achieved on an SEED data set, the operation time is shortened by 62.35% to the maximum compared with that of an original Conformer model, convergence is faster, and robustness is higher. The method can be used for EEG signal emotion recognition, can also be expanded to emotion recognition of physiological signals such as electrocardio and skin electricity, and is suitable for the fields of medical diagnosis and human-computer interaction.

Description

Technical Field

[0001] This invention relates to an EEG emotion recognition model based on dynamic EEG channel selection and the Conformer structure. The core algorithm can be directly extended to emotion recognition scenarios using other physiological signals such as electrocardiogram (ECG), electrodermal conductance (EDA), and electromyography (EMG). It is particularly suitable for fields with high requirements for recognition accuracy and computational efficiency, such as healthcare, human-computer interaction, and smart wearable devices. It belongs to the field of physiological signal processing and pattern recognition technology. Background Technology

[0002] Emotion recognition is one of the core technologies in fields such as human-computer interaction and medical diagnosis. As a direct physiological indicator reflecting brain neural activity, the effective decoding of electroencephalogram (EEG) signals is the key to improving the accuracy of emotion recognition.

[0003] In the development of EEG signal decoding technology, traditional manual feature extraction methods have been extensively studied, but these methods have significant limitations: on the one hand, they are highly task-dependent, and their design is severely constrained by specific experimental paradigms and the prior knowledge of the subjects, resulting in extremely poor model generalization and universality, making it difficult to adapt to the needs of emotion recognition in different scenarios; on the other hand, the traditional process separates and optimizes EEG decoding and classification models independently, which cannot achieve optimal global performance and restricts the improvement of recognition accuracy.

[0004] The rise of deep learning has driven the application of convolutional neural networks (CNNs) in EEG decoding. CNN-based models can automatically learn effective features, eliminating the need for manual feature extraction, and their generalization ability is better than traditional methods. However, the limited receptive field of CNNs makes it difficult for them to capture long-range dependencies in EEG time series, and downsampling operations can easily cause the loss of effective information, affecting the decoding effect.

[0005] The emergence of the Transformer algorithm in 2017, with its global information perception capability through self-attention, provided a new approach to solving the long-range dependency problem in EEG. The EEG-Conformer hybrid model proposed by Song et al. combines the advantages of CNNs in extracting local features with Transformers in capturing global information, achieving significant performance breakthroughs in EEG decoding tasks such as motion visualization and emotion recognition, demonstrating excellent generalization ability. However, the EEG-Conformer still has two major shortcomings:

[0006] 1. High computational complexity: The Transformer architecture it relies on is inefficient when processing high-dimensional EEG data, and the computational overhead increases non-linearly with the increase of sample size or number of channels, which can easily lead to the curse of dimensionality and make it difficult to meet the requirements of real-time applications.

[0007] 2. Insufficient signal robustness: EEG signals are inherently unstable and susceptible to noise contamination. Some channels contain redundant or noisy information, but this invalid information is repeatedly focused on and calculated by the self-attention mechanism, which seriously affects the classification accuracy and robustness of the model.

[0008] Therefore, there is a lack of an emotion recognition model in the current technology that can balance computational efficiency and recognition performance while adapting to the characteristics of EEG signals. It is urgent to solve the above problems through technological innovation. Summary of the Invention

[0009] To overcome the problems of low computational efficiency, limited recognition performance, and difficulty in balancing efficiency and performance in existing EEG-Conformer models and related EEG emotion recognition technologies, this invention provides an EEG emotion recognition model based on dynamic EEG channel selection and a Conformer structure. The aim is to reduce the computational complexity of the Conformer by adding an mRMR module before the Conformer module to select EEG signals, thereby improving the model's convergence and robustness, and achieving high-precision emotion recognition and classification tasks based on EEG signals. This model can be used in the healthcare field to help analyze the relationship between emotions and brain activity, and to aid in the diagnosis of diseases such as depression and mood disorders through emotion recognition.

[0010] A brainwave emotion recognition model based on dynamic EEG channel selection and Conformer structure (mRMR-Conformer) includes signal preprocessing and four cascaded functional modules.

[0011] Signal preprocessing: The input raw EEG signal is Z-score standardized to unify the amplitude dimensions of signals from different channels;

[0012] EEG channel selection module: The minimum redundancy maximum correlation (mRMR) algorithm is used to select channels for the standardized EEG time series. The incremental algorithm searches for a subset of channels that meet the optimization condition of "maximum correlation - minimum redundancy". The number of the subset is 3 / 4 of the total number of channels and then rounded down.

[0013] Specifically, the EEG channel selection module uses the mRMR algorithm to select EEG channels from the standardized EEG time series. This includes calculating the maximum correlation and minimum redundancy, and then combining these two factors into a single optimization problem. An incremental algorithm is used to search for a subset of EEG channels that meet the optimization conditions, which are then used for training and classification in subsequent modules. The selected subset is 3 / 4 of the total number of channels, rounded down.

[0014] Convolutional module: Extracts local spatiotemporal information of EEG signals through temporal convolutional layer, spatial convolutional layer and pooling layer in sequence, and outputs token sequence after dimensionality transformation;

[0015] Specifically, the convolutional module is used to extract local spatiotemporal information from EEG signals, including temporal convolutional layers, spatial convolutional layers, and pooling layers. The convolutional kernels of the temporal convolutional layers move only in the temporal dimension, used to learn the local temporal information of the EEG signal. The convolutional kernels of the spatial convolutional layers are used to capture the interaction relationships between different electrode channels, learning the spatial information of the EEG signal, and a batch normalization component is used to enhance convergence and generalization. The pooling layers use average pooling to smooth the extracted spatiotemporal features and reduce computational complexity.

[0016] Self-attention module: Transforms the token sequence into Query(Q), Key(K), and Value(V) vectors, integrates global context information through self-attention calculation and multi-head attention mechanism, and completes EEG signal decoding;

[0017] The self-attention module integrates the contextual relevance of spatiotemporal features extracted by the convolutional module, i.e., global information, to decode the EEG signal. The self-attention module establishes a globally comprehensive view by introducing three new vectors: Query(Q), Key(k), and Value(V). This invention employs a multi-head attention mechanism to enrich feature diversity, dividing the input sequence into multiple heads on an average basis. These heads are then fed into the self-attention module in parallel, and the results are concatenated and used as the module output.

[0018] Classification module: The decoded features are mapped to class probabilities through two cascaded fully connected layers to complete the emotion classification.

[0019] The classifier module maps the extracted features to class probabilities, thus completing the classification task. This module consists of two fully connected layers: the first layer is used for feature dimensionality reduction, and the second layer is used to project the features into the output space.

[0020] The Z-Score standardization calculation formula is as follows:

[0021]

[0022] Where, xi and x o μ and σ represent the input data and the standardized output data, respectively. 2 These represent the mean and variance, respectively.

[0023] The optimization objective of the minimum redundancy maximum correlation algorithm is:

[0024]

[0025] Where D(S,c) represents all EEG channel signals x i The average mutual information between the DEAP dataset category c and the target dataset; R(S) is the signal x of each EEG channel. i With x j Redundancy between them.

[0026] In the convolution module:

[0027] The temporal convolutional layer uses 40 convolutional kernels of size (1, 25) with a stride of (1, 1);

[0028] The spatial convolutional layer uses 40 convolutional kernels of size (ch, 1) with a stride of (1, 1), where ch is the number of channels in the EEG data;

[0029] The pooling layer is a one-dimensional average pooling layer with a kernel size of (1, 75) and a stride of (1, 15).

[0030] The spatial convolutional layer is configured with a batch normalization component and uses an exponential linear unit as the activation function.

[0031] The calculation formula for the self-attention module is as follows:

[0032]

[0033] Where Q represents the query value, from the output of the previous decoder layer; K and V represent the key and weight, respectively, from the final output of the decoder; and k is the length of the token. The self-attention mechanism introduces a multi-head attention mechanism to enrich feature diversity. The input sequence is divided into h heads on average, which are then fed into the self-attention module in parallel. The results are concatenated and used as the module output. This process can be represented as:

[0034] MHA(Q,K,V)=[head0;…;head h-1 ]

[0035] head l =Attention(Q) l ,K l V l )

[0036] MHA stands for Bullish Attention. h Let Q represent the h-th single-head attention. l ,K l V l These represent the query, key, and value obtained from the segmented token through a linear transformation in the l-th header.

[0037] The classification module uses cross-entropy as the loss function, and the formula is as follows:

[0038]

[0039] Where M represents the number of EEG categories, N b y represents the number of samples in the training batch. i,c It is the true label of sample i with respect to category c. This is the probability predicted by the model that sample i belongs to category c.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0041] This invention proposes a novel EEG emotion recognition model: mRMR-Conformer. This model uses the Conformer network, known for its high classification accuracy and strong generalization, as a benchmark. The Conformer network utilizes convolutional modules to extract temporal and spatial local information from EEG signals and self-attention modules to extract global contextual information from spatiotemporal features, achieving state-of-the-art high-precision classification results in the field of EEG emotion recognition. However, the Conformer network suffers from the common drawbacks of all Transformer-based neural networks: long computation time and low computational efficiency. This invention innovatively adds an mRMR module before the Transformer module to pre-select EEG channels, significantly reducing the computational efficiency of the original algorithm while further improving the model's convergence, classification accuracy, and generalization. The computational efficiency is significantly improved: the mRMR channel selection module filters out 25% of redundant and noisy channels, significantly reducing the computational complexity of the subsequent Conformer module. On the DEAP dataset, the average computation time per subject decreased by 83 seconds, a year-on-year decrease of 21.67%; on the SEED large sample dataset, the average computation time decreased by 6795 seconds, a year-on-year decrease of 62.35%, and the mRMR module itself has a very short computation time, with almost no additional overhead.

[0042] Significantly optimized recognition performance: The filtered channel subset combines high relevance and low redundancy, effectively improving feature quality. The model achieves an average recognition accuracy of 98.69% and an F1-score of 98.47% on the DEAP dataset, and an average recognition accuracy of 90.54% and an F1-score of 90.48% on the SEED dataset, both exceeding the current state-of-the-art network models. Furthermore, it exhibits the smallest standard deviations in accuracy and F1-score across both datasets, demonstrating excellent stability.

[0043] Significantly improved convergence characteristics: The preprocessing function of the mRMR module helps the model avoid interference from invalid information, resulting in faster and more stable convergence. It can achieve a better loss value with the same number of iterations, reducing the difficulty and cost of model training.

[0044] Enhanced generalization and robustness: The synergistic effect of batch normalization, multi-head attention mechanism and mRMR channel selection enables the model to adapt to the instability and noise interference of EEG signals, while showing good adaptability to different datasets and different subjects, and its generalization ability is better than existing models.

[0045] Wide range of applications: The model is not only applicable to emotion recognition of EEG signals, which can assist in the diagnosis of diseases such as depression and mood disorders, but can also be extended to emotion recognition of other physiological signals such as ECG and EDA. It can maintain high robustness in noisy real-world environments and is suitable for multiple fields such as healthcare, smart wearables, and human-computer interaction. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart of the mRMR-Conformer algorithm of an EEG emotion recognition model based on dynamic EEG channel selection and Conformer structure according to the present invention.

[0048] Figure 2 This is a labeled 2D emotion model diagram of the DEAP dataset, which is a brainwave emotion recognition model based on dynamic brainwave channel selection and Conformer structure according to the present invention.

[0049] Figure 3 This is a diagram showing the specific parameters of the convolutional layer in an EEG emotion recognition model based on dynamic EEG channel selection and Conformer structure according to the present invention.

[0050] Figure 4 This is a comparison chart showing the classification accuracy of the EEG emotion recognition model based on dynamic EEG channel selection and Conformer structure proposed in this invention, mRMR-Conformer, on the DEAP and SEED datasets, with various current state-of-the-art deep learning algorithms.

[0051] Figure 5 This is a comparison of the computation time of the core module in the mRMR-Conformer model of the DEAP and SEED datasets of the EEG emotion recognition model based on dynamic EEG channel selection and Conformer structure of the present invention.

[0052] Figure 6 This is a comparison of the loss curves of mRMR-Conformer and Conformer networks on the DEAP and SEED datasets of the EEG emotion recognition model based on dynamic EEG channel selection and Conformer structure proposed in this invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Example

[0055] This application proposes an EEG emotion recognition model. This model utilizes mRMR for EEG channel selection to reduce computation time and improve model generalization. It employs CNN to extract local temporal and spatial information from EEG signals, uses a Transformer module to integrate global contextual information from spatiotemporal local features, and finally completes the emotion classification problem through a fully connected classification layer. Examples will be provided below for illustration.

[0056] The algorithm flowchart for mRMR-Conformer can be found here. Figure 1 It visually illustrates all the core modules of the mRMR-Conformer model, and the algorithm can be executed by a computer.

[0057] The embodiments include the publicly available EEG emotion datasets DEAP and SEED. The operation on the two datasets is similar. The following will use the DEAP dataset as an example to introduce the usage process of the present invention, and at the end, the application results of mRMR-Conformer on the two datasets will be introduced.

[0058] We selected two dimensions, arousal and valence, from the DEAP dataset to construct a four-class classification dataset. The original dataset contained continuous data ranging from 0 to 10 for these two dimensions; this example uses 5 as the boundary pair to divide it into discrete data. See the detailed four-class classification model below. Figure 2 .

[0059] In this embodiment, five-fold cross-validation is used to verify the model's performance. Each subject in the DEAP dataset is divided into five parts. Each part is used as the test set, and the remaining four parts are used as the training set. Finally, the test results are averaged to measure the model's performance.

[0060] To suppress irrelevant noise and reduce non-stationarity and volatility in EEG data, Z-score normalization was performed. The normalization formula is as follows:

[0061]

[0062] Where, x i and x o These represent the input data and the standardized output data, respectively. μ and σ 2 This represents the mean and variance.

[0063] like Figure 1 As shown, the mRMR-Conformer consists of four parts: an mRMR feature selection module, a convolution module, a self-attention module, and a fully connected classifier module. The input is a batch of preprocessed two-dimensional EEG data with dimensions ch×sp (number of channels × number of time sampling points). The EEG data is then input to the EEG channel selection module for EEG channel selection.

[0064] EEG channel selection module: The minimum redundancy maximum correlation algorithm is used to select channels for the standardized EEG time series. The incremental algorithm searches for a subset of channels that meet the optimization condition of "maximum correlation - minimum redundancy". The number of the subset is 3 / 4 of the total number of channels and rounded down.

[0065] Minimal-Redundance-Maximum-Relevance (mRMR) is a typical feature selection algorithm. Its goal is to find a subset of features S such that the subset has the highest correlation with the output variable c, while minimizing the redundancy between features within the subset. The correlation metric used in this invention is mutual information, and the maximum correlation D(S,c) can be approximated by the sum of all EEG channel signals x. i The average mutual information between the DEAP dataset and category c:

[0066]

[0067] Where S is the feature subset, c is the output variable, and x is the output variable. i The variable to be evaluated is I(x). i c) represents feature x i Mutual information between class c and class c. The formula for calculating mutual information is as follows:

[0068]

[0069] Wherein, P(x i c) represents feature x i The joint probability with class c, P(x) i P(c) and P(c) represent the marginal probabilities of the feature and the category, respectively.

[0070] Features selected using Max-Relevance are often redundant, and the lack of independence between features can easily lead to data consistency issues. Removing one redundant feature from a subset does not significantly affect the classification results. A concise feature subset not only improves computational efficiency but also ensures the robustness and interpretability of the model. Therefore, Min-Redundancy is used to remove redundant features:

[0071]

[0072] Combining the maximum correlation D and minimum redundancy R yields the Minimal-Redundance-Maximum-Relevance (mRMR) algorithm, which can be expressed as:

[0073]

[0074] The mRMR algorithm not only considers the correlation between the feature subset and the output variable, but also the redundancy between feature variables, obtaining the most concise and efficient feature subset through feature selection. The subset search algorithm uses an incremental algorithm, and the mRMR calculation tool used in this invention is the Python mrmr library, with all parameters set to default. In this model, the mRMR algorithm is applied to select the feature subset with the highest relevance and lowest redundancy from the initial feature set as the input to the subsequent convolutional module. The DEAP dataset has a total of 32 EEG channels for the subjects, and the number of feature subsets after mRMR selection is 24.

[0075] Convolutional module: Extracts local spatiotemporal information of EEG signals through temporal convolutional layer, spatial convolutional layer and pooling layer in sequence, and outputs token sequence after dimensionality transformation;

[0076] The EEG data is expanded to one dimension and used as a convolution channel input into the convolution module. The data shape entering the Conformer module is (1, 24, sp). 24 is the number of EEG channels, and sp is the number of samples per subject in the DEAP dataset at a sampling frequency of 128Hz. Figure 3 The specific parameters of the convolutional module are shown, where k is a hyperparameter, and in this invention, k equals 40. The first layer of the convolutional layer has k convolutional kernels of size (1, 25) with a stride of (1, 1). This convolutional kernel slides only in the time dimension, without mixing signals from different electrodes, and is used to learn temporal features. The second layer uses k convolutional kernels of size (ch, 1) with a stride of (1, 1), where ch is the number of channels in the EEG data. This layer is used to capture the interaction between different electrode channels and learn spatial features.

[0077] Then, batch normalization is employed to improve convergence and generalization ability, and the exponential linear unit (ELU) is used as the activation function. The third layer is a pooling layer that performs one-dimensional average pooling in the time dimension, with a kernel size of (1, 75) and a stride of (1, 15), which not only reduces computational complexity but also smooths the temporal features and improves the robustness of feature extraction. The hyperparameter k is set to 40. Finally, through operations such as squeezing electrode channels, transposing electrode channel dimensions, and spatial dimensions, the feature maps of the convolutional module are reordered, thereby converting a segment of EEG signal into a series of information-rich token sequences, which are then fed into the next module.

[0078] Self-attention module: Transforms the token sequence into Query(Q), Key(K), and Value(V) vectors, integrates global context information through self-attention calculation and multi-head attention mechanism, and completes EEG signal decoding;

[0079] Due to the coherence of neural activity, contextual relevance is extracted from low-level spatiotemporal features to facilitate EEG decoding. Because the limited receptive field of convolutional modules makes it difficult to capture long-term temporal dependencies, a self-attention mechanism is used to learn the global temporal dependencies of EEG features. The reordered tokens from the convolutional modules undergo three independent linear transformations to generate three new vectors: Query(Q), Key(k), and Value(V). The dot product of Q and K is calculated to obtain the attention score matrix, which measures the relevance between different tokens. Simultaneously, to avoid gradient vanishing...

[0080] A scaling factor was used to ensure training stability. Then, the Softmax function was applied to the scaled score matrix to normalize it into an attention weight matrix, which was then weighted and summed with the V matrix. This calculation process can be represented as:

[0081]

[0082] Where k is the length of the token.

[0083] The self-attention mechanism introduces a multi-head attention mechanism to enrich feature diversity. The input sequence is divided into h heads on average, which are then fed into the self-attention module in parallel. The results are then concatenated and used as the module output. This process can be represented as:

[0084] MHA(Q,K,V)=[head0;…;head h-1 ]

[0085] head l =Attention(Q) l ,k l V l )

[0086] MHA stands for Bullish Attention. h Let Q represent the h-th single-head attention. l ,k l V l These represent the query, key, and value obtained from the segmented token through a linear transformation in the l-th header.

[0087] Subsequently, each self-attention sub-layer is followed by a feedforward network consisting of two fully connected layers to enhance the model's fitting ability; this sub-layer follows the identity mapping design to ensure that its input and output dimensions are consistent. The computation process of self-attention and the feedforward network is repeated N times in the self-attention module, where N is set to 6.

[0088] Classification module: The decoded features are mapped to class probabilities through two cascaded fully connected layers to complete the emotion classification.

[0089] The classifier module consists of two fully connected layers. The first layer performs non-linear transformations and reduces the dimensionality of the features, while the second layer projects the hidden features into an M-dimensional output space. The output of this layer is normalized to class probabilities using the Softmax function, and cross-entropy is used as the model's loss function.

[0090]

[0091] Where M represents the number of EEG categories, in this case M is 4, N b y represents the number of samples in the training batch. i,c It is the true label of sample i with respect to category c. This is the probability predicted by the model that sample i belongs to category c.

[0092] This invention uses the Adam optimizer to train the model, with the learning rate, β1 (exponential decay rate of the first moment estimate), and β2 (exponential decay rate of the second moment estimate) set to 0.0001, 0.5, and 0.999, respectively. The maximum number of training iterations is set to 300.

[0093] Compare the model's performance with the latest types of deep learning networks, such as... Figure 4 As shown, mRMR-Conformer achieves an average accuracy of 98.69% and an F1 score of 98.47% on the DEAP dataset, and an average accuracy of 90.54% and an F1 score of 90.48% on the SEED dataset. Furthermore, it exhibits the smallest standard deviations in accuracy and F1 score on both datasets, demonstrating excellent model stability and surpassing various other state-of-the-art network models.

[0094] Figure 5 The changes in average computation time for all subjects across two datasets before and after the removal of the mRMR module are shown. It can be seen that the computation time of the mRMR module differs significantly from that of the Conformer module, sometimes by orders of magnitude. In the large SEED dataset, the Conformer's computation time is more than 100 times that of mRMR. After introducing mRMR for feature selection, in the small sample data DEAP, the average computation time per subject decreased by 83 seconds, a year-on-year reduction of 21.67%. In the large sample dataset SEED, the average computation time per subject decreased by 6795 seconds, a year-on-year reduction of 62.35%. With the introduction of mRMR, which adds almost no computation time, the Conformer module's computation time was reduced by more than half. In conclusion, mRMR is significantly effective in reducing computation time.

[0095] Figure 6 The paper demonstrates the convergence characteristics of the network before and after removing mRMR, using the first subject in two datasets as an example. It can be observed that mRMR can enable the Conformer to achieve earlier and smoother convergence.

[0096] This invention uniquely adds a dynamic EEG channel selection module (mRMR) before the end-to-end EEG-Conformer module to address the issues of excessive computation time and redundant calculations in Transformer-based network models for EEG emotion recognition. This step steadily improves the convergence characteristics, classification accuracy, and generalization ability of the original Conformer model, while significantly reducing the computational complexity of Transformer-based network models.

[0097] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. An electroencephalogram (EEG) emotion recognition model based on dynamic EEG channel selection and conformer structure, characterized in that: The method comprises signal preprocessing and four serial functional modules, The signal preprocessing is Z-Score standardization processing of the input electroencephalogram (EEG) original signal, and unification of the amplitude dimension of different channel signals; The EEG channel selection module adopts a minimum redundancy maximum correlation algorithm to select channels of the standardized EEG time sequence, and searches for a channel subset satisfying the optimization condition of "maximum correlation-minimum redundancy" through an incremental algorithm, wherein the number of the subset is 3 / 4 of the total number of channels and is rounded down; The convolution module sequentially extracts local space-time domain information of the EEG signal through a time convolution layer, a space convolution layer and a pooling layer, and outputs a token sequence after dimension conversion; The self-attention module converts the token sequence into Query (Q), Key (K) and Value (V) vectors, integrates global context information through self-attention calculation and a multi-head attention mechanism, and completes EEG signal decoding; The classification module maps the decoded features to class probabilities through two serial full connection layers, and completes emotion classification.

2. The electroencephalogram emotion recognition model based on dynamic electroencephalogram channel selection and conformer structure according to claim 1, characterized in that: The calculation formula of the Z-Score standardization is as follows: where x i and x o represent the input data and the normalized output data, respectively, and μ and σ 2 represent the mean and variance, respectively.

3. The electroencephalogram emotion recognition model based on dynamic electroencephalogram channel selection and conformer structure according to claim 1, characterized in that: The optimization objective of the minimum redundancy maximum correlation algorithm is as follows: where D(S, c) is the entropy of all EEG channel signals x i the average mutual information between the DEAP dataset class c; R(S) is the redundancy between the individual EEG channel signals x i j and x​ 4. The electroencephalogram emotion recognition model based on dynamic electroencephalogram channel selection and conformer structure according to claim 1, characterized in that: In the convolution module, The time convolution layer adopts 40 convolution kernels with a size of (1, 25) and a step of (1, 1); The space convolution layer adopts 40 convolution kernels with a size of (ch, 1) and a step of (1, 1), wherein ch is the number of channels of the EEG data; The pooling layer is one-dimensional average pooling, the convolution kernel size is (1, 75), and the step is (1, 15).

5. The electroencephalogram emotion recognition model based on dynamic electroencephalogram channel selection and conformer structure according to claim 1, characterized in that: The space convolution layer is configured with a batch normalization component, and an exponential linear unit is used as an activation function.

6. The electroencephalogram emotion recognition model based on dynamic electroencephalogram channel selection and conformer structure according to claim 1, characterized in that: The calculation formula of the self-attention module is as follows: Wherein, Q represents a query value, which is output from the previous layer decoder, K and V represent a key value and a weight, respectively, which are output from the last output of the decoder, and k is the length of the token; The multi-head attention mechanism is introduced into the self-attention mechanism to enrich the diversity of features, the input sequence is evenly divided into h heads, and the results are connected and used as the output of the module after parallel input into the self-attention module, and this process is represented as follows: MHA(Q,K,V) = [head0;... ; head h-1 ] head l = Attention(Q l , K l , V l ) where MHA denotes multi-head attention, head h represents the h-th single-head attention, Q l ,K l ,V l denote the query, key and value obtained by linear transformation of the segmented tokens in the l-th head, respectively.

7. The electroencephalogram emotion recognition model based on dynamic electroencephalogram channel selection and conformer structure according to claim 1, characterized in that: The classification module adopts cross entropy as a loss function, and the formula is as follows: where M represents the number of EEG classes, N b represents the number of samples in a training batch, y i,c is the true label of sample i for class c, is the probability that the model predicts that sample i belongs to class c.

Images