Electroencephalogram signal classification method based on multi-scale causal convolution and KAN attention

Abstract

The invention relates to an electroencephalogram signal classification method based on multi-scale causal convolution and KAN attention. The electroencephalogram signal classification method comprises the steps of 1, EEG data collection and preprocessing; 2, EEG data enhancement; and 3, motor imagery task classification based on EEG signals. According to the method, the spatial-temporal characteristics of the EEG signals under different frequencies can be effectively extracted, the multi-scale nonlinear characteristics are efficiently fused and weighted, and the problem that nonlinear fitting of the EEG signals in a motor imagery task is insufficient is solved, so that more accurate and reliable motor imagery classification is realized.

Description

Technical Field

[0001] This invention relates to a brain-computer interface (BCI) signal classification method based on multi-scale causal convolution and KAN attention, belonging to the field of brain-computer interface (BCI) technology. Background Technology

[0002] Brain-Computer Interface (BCI) technology enables direct information interaction between the human brain and external devices by decoding electroencephalogram (EEG) signals. Among the many BCI paradigms, the classification of Motor Imagery (MI) EEG signals is one of the most widely studied directions. MI generates different EEG signals by imagining limb movements, and by decoding these signals, different intention states can be identified and classified. Due to its non-invasiveness and intuitive interaction, MI is widely used in neurorehabilitation training, assisted motor control, and intelligent prosthetic manipulation. However, MI-EEG signals typically exhibit significant individual variability and spatiotemporal non-stationarity. Their characteristic distribution is affected by multiple factors such as the intensity of subjective imagination, brain region activity, and environmental noise, resulting in low signal decoding accuracy. Therefore, decoding MI-EEG signals remains a challenging task.

[0003] Existing deep learning methods mostly employ convolutional neural networks (CNNs) to extract local temporal features of EEG signals. Lawhern et al. proposed the EEGNet network, which significantly reduces the number of model parameters while still achieving excellent classification performance in various EEG signal decoding tasks by using depthwise separable convolutions in CNNs.

[0004] Recently, a type of network called Temporal Convolutional Network (TCN) has been specifically designed for time series modeling and classification. Ingolfsson et al. combined TCN with the EEGNet architecture to design the EEG-TCN model. Zhao et al. combined TCN with Transformer to improve the model's ability to capture key information. Although TCN has advantages in temporal modeling, its multi-layer concatenated and dilated convolution structure is prone to problems such as gradient vanishing, information redundancy, and feature loss, making it difficult to effectively represent complex EEG dynamic features. In addition, since TCN modules are usually located after the local feature extraction layer of CNN, this structure may cause leakage of future temporal information, which is not conducive to the real-time modeling and online application of EEG signals. At the same time, traditional attention mechanisms usually rely on a combination of linear mapping and fixed activation functions to calculate attention weights, which is difficult to achieve accurate modeling and effective fitting of complex EEG patterns for EEG signals with highly nonlinear and non-stationary characteristics. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention provides an EEG signal classification method based on multi-scale causal convolution and KAN attention, aiming to improve the accuracy of motor imagery EEG signal classification and achieve effective temporal modeling and nonlinear feature characterization of EEG signals. This invention innovatively combines multi-scale processing with TCN, eliminating the dependence on dilated convolution, preserving the temporal continuity of EEG signals, and enabling parallel capture of short-term and long-term temporally dependent features. This effectively alleviates the gradient vanishing and information redundancy problems commonly found in single-scale multi-layer causal convolution, improving the completeness and stability of feature representation. Simultaneously, this invention introduces the Kolmogorov-Arnold network (KAN) into the attention mechanism, constructing a KAN-based local and global fusion attention structure. This allows the model to simultaneously focus on local features and overall temporal patterns of EEG signals, thereby significantly improving the expressive power of nonlinear EEG features and the decoding performance of motor imagery EEG signals.

[0006] A brainwave signal classification method based on multi-scale causal convolution and KAN attention includes the following steps:

[0007] Step 1: EEG Data Acquisition and Preprocessing: The experiment was set up according to the motor imagery (MI) experimental paradigm, and EEG data of subjects were collected under motor imagery tasks. After the experiment, the collected EEG data were labeled and preprocessed to generate EEG data sets under different motor task conditions.

[0008] Step 2: EEG Data Augmentation: Take the EEG signal data from Step 1 and perform data augmentation by slicing and reconstructing the time series data to expand the diversity of the training samples;

[0009] Step 3: Classification of motor imagery task based on EEG signals: The enhanced samples from Step 2 are input into the feature extraction module to extract the spatiotemporal features of the EEG signals; the extracted features are further weighted by the attention module to highlight key information; the weighted features are classified by the classification module to achieve EEG classification of motor imagery tasks.

[0010] S3.1: Perform preliminary spatiotemporal feature extraction on the enhanced EEG data. Spatiotemporal features of EEG at different scales are extracted using a multi-layer, multi-scale network, and information from each scale branch is integrated through feature fusion and interaction mechanisms.

[0011] The multi-scale feature extraction module consists of two multi-scale networks, including a temporal causal convolutional layer and a depthwise temporal convolutional layer. Each branch structure is composed of temporal causal convolution, spatial convolution, and depthwise separable convolution, used to extract spatiotemporal features at different scales. Finally, the features extracted from each scale branch are fused and interacted to achieve collaborative modeling and comprehensive representation of multi-scale information.

[0012] S3.2: Input the EEG spatiotemporal features extracted by the multi-scale network into the local-global nonlinear attention module. Obtain the weighted local and global features through an adaptive weighting mechanism, and then perform adaptive fusion through learnable parameters. Finally, input the results into the nonlinear classifier to output the classification results.

[0013] The features extracted in S3.1 undergo local-global attention weighting to highlight key information and suppress redundant features. Subsequently, the weighted local and global features are fused and weighted to fully combine their complementary information. Finally, the fused features are input into the classification module, which outputs the final classification result of the motor imagery EEG signal.

[0014] Before the start of the above steps, a preliminary experiment is conducted to assess the potential impact of eye movement signals on EEG recordings. The formal experimental procedure includes: an attention cuing phase, a task cuing phase, a task execution phase, and a rest phase.

[0015] Specifically, a 5-minute pre-experiment was conducted before the formal experiment to assess the potential impact of eye movement signals on EEG recordings. The formal experimental procedure included: a 2-second attentional cueing phase, a 1.25-second task cueing phase, a 4-second task execution phase, and a 1.5-second rest phase. After the experiment, the collected EEG data were labeled and preprocessed to generate EEG signal datasets under different motor task conditions.

[0016] The experiment in step one, which is set up according to the motor imagery experiment paradigm, has the following procedure for each experiment: task preparation prompting stage, task prompting stage; participants perform the corresponding motor imagery task; and then enter the rest stage to restore the baseline state.

[0017] Specifically, the experiment in step one is set up according to the motor imagery experiment paradigm. The experimental procedure for each experiment is as follows: First, there is a 2-second task preparation prompting phase, followed by a 1.25-second task prompting phase; then, the participants perform the corresponding motor imagery task within 3 seconds; finally, there is a 2-3 second rest phase to restore the baseline state.

[0018] One specific preprocessing step involves applying a 0.5-100Hz bandpass filter to the EEG signal to preserve the effective frequency band relevant to the motion visualization task; followed by using a 50Hz notch filter to suppress power frequency noise interference.

[0019] During the classification phase, only the EEG data in the [2s, 6s] interval of each experiment were selected as the effective analysis segment to ensure that the complete motor imagery task execution process was included and to reduce interference during the preparation and recovery periods. The preprocessed EEG data consisted of dimensions (N, 288, 22, 1000), where N represents the number of participants, 288 represents the number of experiments for each participant, 22 represents the number of channels, and 1000 corresponds to the number of sampling points of the 4s data at a sampling rate of 250Hz.

[0020] Data augmentation was performed on the preprocessed EEG data. For the same subject's EEG data under the same motor imagery task, a segmentation-reconstruction strategy was adopted to generate more samples. In the segmentation stage, each EEG signal was divided into several time segments according to the time window. In the reconstruction stage, new artificial EEG samples were constructed by randomly recombining and splicing these time segments.

[0021] S3.1 specifically includes:

[0022] S3.11: Input EEG data is first modeled using multi-scale temporal causal convolution to capture short-term and long-term temporal features of EEG signals at different scales;

[0023] The input raw EEG signal is sequentially passed through multi-scale temporal causal convolutional layers for temporal feature modeling, capturing short-term and long-term dependencies at different time scales, thereby enhancing the temporal continuity and predictability of the EEG signal. The receptive field of view (RFS) of causal convolution is calculated as follows:

[0024] RFS = 2 × (k-1) + 1

[0025] Where k represents the kernel size.

[0026] S3.12: The preliminary EEG temporal features obtained in S3.11 are further input into the spatial and deep temporal feature extraction network. First, the features between each EEG channel are fused through the spatial feature extraction module to enhance the spatial correlation across channels. Then, the fused features are input into the multi-scale deep convolution module to extract high-frequency temporal features from different scales to obtain a more discriminative spatiotemporal representation.

[0027] S3.13: The spatiotemporal features at different scales extracted in S3.12 are integrated through a convolutional fusion module to integrate the preliminary spatiotemporal features extracted by the multi-scale network.

[0028] S3.2 specifically includes:

[0029] S3.21: Input the multi-scale fused spatiotemporal features output from S3.13 into the local-global nonlinear attention module, adaptively weight the features, and obtain the weighted local and global features respectively. The nonlinear attention module is designed based on the KAN (Kolmogorov-Arnold) network, and its form is as follows:

[0030]

[0031] Where x represents the input feature, d in Indicates the input dimension, d out φ represents the output dimension, and φ represents the linear combination of the SiLU function and the B-spline function.

[0032] φ(x)=ω b ·SiLu(x)+ω s ·splien(x)

[0033] in Let spline(x) represent the self-gated activation function. i c i B i (x) represents the B-spline basis function;

[0034] The KAN network is constructed by stacking multiple φ(x) composite transformations:

[0035]

[0036] The global attention module combines the KAN network and the Softmax function to calculate the attention weight of each feature and multiplies it element-wise with the input features to obtain the globally weighted features; while the local attention module establishes the correlation between adjacent features through local convolution operations before calculating the attention weights, performs attention calculation and weighting, and better captures local information.

[0037] Attention global = x0·Softmax(KAN(x0))+x0

[0038] Attention local =Conv(x0)·Softmax(KAN(Conv(x0)))+Conv(x0)S3.22: Adaptively fuse global attention features and local attention features through learnable parameters a to obtain a comprehensive EEG spatiotemporal feature representation; input the fused features into a nonlinear classifier to output the final classification result;

[0039] X att =α·Attention global +(1-α)Attentionlocal .

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

[0041] This invention relates to an EEG signal classification method based on multi-scale causal convolution and KAN attention. It fully leverages the complementary characteristics between different scales to jointly model the continuity of multi-scale temporal signals, innovatively constructing a hierarchical multi-scale network structure. This allows the model to simultaneously capture local and global dynamic features at different levels and time scales, significantly improving its ability to model and classify complex EEG signal temporal patterns. Furthermore, to further enhance feature discriminative power, this invention designs a nonlinear local-global attention fusion module. Through an adaptive mechanism, it fuses local and global attention features, enabling the model to consider both fine-grained features and overall patterns of EEG signals. This effectively improves the ability to capture complex nonlinear dynamic features in EEG signals, thereby enhancing the model's feature representation and decoding performance and solving problems such as future information leakage and feature information loss in existing EEG temporal convolutional networks. Attached Figure Description

[0042] 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.

[0043] Figure 1 This is a flowchart of the EEG signal preprocessing process for the EEG signal classification method based on multi-scale causal convolution and KAN attention in this invention.

[0044] Figure 2 This is a flowchart illustrating the data augmentation process of the EEG signal classification method based on multi-scale causal convolution and KAN attention in this invention.

[0045] Figure 3 This diagram illustrates the EEG signal classification method of the present invention, which is based on multi-scale causal convolution and KAN attention. Detailed Implementation

[0046] 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.

[0047] Reference Figure 3 A brainwave signal classification method based on multi-scale causal convolution and KAN attention includes the following steps:

[0048] Step 1: EEG Data Acquisition and Preprocessing: The experiment was set up according to the motor imagery (MI) experimental paradigm, and EEG data of subjects were collected under motor imagery tasks. After the experiment, the collected EEG data were labeled and preprocessed to generate EEG data sets under different motor task conditions.

[0049] Participants were guided through the experiment according to the paradigm of motor imagery experiments. During the experiment, participants performed corresponding motor imagery tasks based on visual cues presented on the screen and brief beeping signals. There were four types of tasks: left hand, right hand, tongue, and feet. Before each task, participants were required to focus on a fixed black cross in the center of the screen to concentrate and reduce eye movement artifacts. Each participant completed 72 trials in each of the four tasks. The experimental procedure was as follows: first, a 2-second task preparation cues phase, followed by a 1.25-second task cues phase; then, participants performed the corresponding motor imagery task within 3 seconds; finally, a 2–3-second rest phase to restore baseline state. The experiment used the international 10–20 system standard EEG cap setup, employing 22 electrode channels (e.g., Fz, Cz, Pz, C3, C4, P3, P4, O1, O2, etc.), and recorded EEG signals at a sampling frequency of 250 Hz.

[0050] Figure 1 The preprocessing workflow for EEG data is demonstrated. Since EEG signals are highly susceptible to external interference such as eye movements, electromyography (EMG), and power supply frequency, rigorous filtering and denoising of the raw signals are necessary during preprocessing. First, a bandpass filter of 0.5–100 Hz is applied to the EEG signals to retain the effective frequency bands relevant to the motor imagery task. Subsequently, a 50 Hz notch filter is used to suppress power frequency noise interference. In the classification stage, only the EEG data within the [2s, 6s] interval of each experiment is selected as the effective analysis segment to ensure the complete execution of the motor imagery task and reduce interference from the preparation and recovery periods. The preprocessed EEG data is structured in dimensions (N, 288, 22, 1000), where N represents the number of participants, 288 represents the number of experiments per participant, 22 represents the number of channels, and 1000 corresponds to the number of sampling points for 4s of data at a sampling rate of 250 Hz.

[0051] Step 2: EEG Data Augmentation: Take the EEG signal data from Step 1 and perform data augmentation by slicing and reconstructing the time series data to expand the diversity of the training samples;

[0052] Data augmentation is performed on the preprocessed EEG data, such as... Figure 2 As shown, a segmentation-reconstruction strategy was employed to generate more samples from the EEG data of the same subject performing the same motor imagery task. Specifically, in the segmentation stage, each EEG signal was divided into several time segments according to a time window; in the reconstruction stage, new artificial EEG samples were constructed by randomly recombining and splicing these time segments.

[0053] Step 3: Classification of motion imagery tasks based on EEG signals:

[0054] The enhanced sample from step two is input into the feature extraction module to extract the spatiotemporal features of the EEG signal; the extracted features are further weighted by the attention module to highlight key information; the weighted features are classified by the classification module to achieve EEG classification of the motor imagery task.

[0055] S3.1: Perform preliminary spatiotemporal feature extraction on the enhanced EEG data. Spatiotemporal features of EEG at different scales are extracted using a multi-layer, multi-scale network, and information from each scale branch is integrated through feature fusion and interaction mechanisms.

[0056] The multi-scale feature extraction module consists of two multi-scale networks, including a temporal causal convolutional layer and a depthwise temporal convolutional layer. Each branch structure is composed of temporal causal convolution, spatial convolution, and depthwise separable convolution, used to extract spatiotemporal features at different scales. Finally, the features extracted from each scale branch are fused and interacted to achieve collaborative modeling and comprehensive representation of multi-scale information.

[0057] S3.1 specifically includes:

[0058] S3.11: Input EEG data is first modeled using multi-scale temporal causal convolution to capture short-term and long-term temporal features of EEG signals at different scales;

[0059] The input raw EEG signal is sequentially passed through multi-scale temporal causal convolutional layers for temporal feature modeling, capturing short-term and long-term dependencies at different time scales, thereby enhancing the temporal continuity and predictability of the EEG signal. The receptive field of view (RFS) of causal convolution is calculated as follows:

[0060] RFS = 2 × (k-1) + 1

[0061] Where k represents the kernel size.

[0062] S3.12: The preliminary EEG temporal features obtained in S3.11 are further input into the spatial and deep temporal feature extraction network. First, the features between each EEG channel are fused through the spatial feature extraction module to enhance the spatial correlation across channels. Then, the fused features are input into the multi-scale deep convolution module to extract high-frequency temporal features from different scales to obtain a more discriminative spatiotemporal representation.

[0063] S3.13: The spatiotemporal features at different scales extracted in S3.12 are integrated through a convolutional fusion module to integrate the preliminary spatiotemporal features extracted by the multi-scale network.

[0064] S3.2: Input the EEG spatiotemporal features extracted by the multi-scale network into the local-global nonlinear attention module. Obtain the weighted local and global features through an adaptive weighting mechanism, and then perform adaptive fusion through learnable parameters. Finally, input the results into the nonlinear classifier to output the classification results.

[0065] The features extracted in S3.1 undergo local-global attention weighting to highlight key information and suppress redundant features. Subsequently, the weighted local and global features are fused and weighted to fully combine their complementary information. Finally, the fused features are input into the classification module, which outputs the final classification result of the motor imagery EEG signal.

[0066] S3.2 specifically includes:

[0067] S3.21: Input the multi-scale fused spatiotemporal features output from S3.13 into the local-global nonlinear attention module, adaptively weight the features, and obtain the weighted local and global features respectively. The nonlinear attention module is designed based on the KAN (Kolmogorov-Arnold) network, and its form is as follows:

[0068]

[0069] Where x represents the input feature, d in Indicates the input dimension, d out φ represents the output dimension, and φ represents the linear combination of the SiLU function and the B-spline function.

[0070] φ(x)=ω b ·SiLu(x)+ω s ·splien(x)

[0071] in Let spline(x) represent the self-gated activation function. i c i B i (x) represents the B-spline basis function;

[0072] The KAN network is constructed by stacking multiple φ(x) composite transformations:

[0073]

[0074] The global attention module combines the KAN network and the Softmax function to calculate the attention weight of each feature and multiplies it element-wise with the input features to obtain the globally weighted features; while the local attention module establishes the correlation between adjacent features through local convolution operations before calculating the attention weights, performs attention calculation and weighting, and better captures local information.

[0075] Attention global = x0·Softmax(KAN(x0))+x0

[0076] Attention local =Conv(x0)·Softmax(KAN(Conv(x0)))+Conv(x0)S3.22: Adaptively fuse global attention features and local attention features through learnable parameters a to obtain a comprehensive EEG spatiotemporal feature representation; input the fused features into a nonlinear classifier to output the final classification result;

[0077] X att =α·Attention global +(1-α)Attention local .

[0078] During training, the Adam optimizer was used with a learning rate of 0.001, employing the cross-entropy loss function:

[0079]

[0080] Where M represents the number of MI-EEG experiments, N represents the number of categories, and y ij It is the true label of the j-th class of the i-th sample. This represents the predicted probability.

[0081] Train for 1000 rounds (batch_size=64) on an NVIDIA RTX 5070.

[0082] Table 1 presents a comparison of the classification accuracy of the proposed parallel temporal convolutional network with other networks. The average accuracy on subjects A01-A09 reached 87.54%, significantly outperforming the comparison methods.

[0083] Table 1. Performance comparison of this invention with other network classification methods.

[0084]

[0085] Before the formal experiment, a 5-minute pre-experiment was conducted to assess the potential impact of eye movement signals on EEG recordings. The formal experimental procedure included: a 2-second attentional cueing phase, a 1.25-second task cueing phase, a 4-second task execution phase, and a 1.5-second rest phase. After the experiment, the collected EEG data were labeled and preprocessed to generate EEG signal datasets under different motor task conditions.

[0086] The experiment in step one, which is set up according to the motor imagery experiment paradigm, has the following procedure for each experiment: First, there is a 2-second task preparation prompting phase, followed by a 1.25-second task prompting phase; then, the participants perform the corresponding motor imagery task within 3 seconds; finally, there is a 2-3 second rest phase to restore the baseline state.

[0087] One specific preprocessing step involves applying a 0.5-100Hz bandpass filter to the EEG signal to preserve the effective frequency band relevant to the motion visualization task; followed by using a 50Hz notch filter to suppress power frequency noise interference.

[0088] During the classification phase, only the EEG data in the [2s, 6s] interval of each experiment were selected as the effective analysis segment to ensure that the complete motor imagery task execution process was included and to reduce interference during the preparation and recovery periods. The preprocessed EEG data consisted of dimensions (N, 288, 22, 1000), where N represents the number of participants, 288 represents the number of experiments for each participant, 22 represents the number of channels, and 1000 corresponds to the number of sampling points of the 4s data at a sampling rate of 250Hz.

[0089] Data augmentation was performed on the preprocessed EEG data. For the same subject's EEG data under the same motor imagery task, a segmentation-reconstruction strategy was adopted to generate more samples. In the segmentation stage, each EEG signal was divided into several time segments according to the time window. In the reconstruction stage, new artificial EEG samples were constructed by randomly recombining and splicing these time segments.

[0090] This invention is the first to realize a parallel temporal causal convolution structure, which not only effectively overcomes the problems of future information leakage and feature information loss in traditional causal convolution, but also combines a nonlinear local-global attention mechanism to solve the limitations of traditional attention mechanisms in EEG nonlinear feature modeling. This method significantly improves the accuracy and stability of EEG motor imagery classification, providing more efficient and reliable technical support for the practical clinical application of brain-computer interface systems.

[0091] This invention first preprocesses EEG signals acquired during a motion imagery task and augments the data through time-series slicing and reconstruction. Then, the augmented EEG data is input into a multi-scale feature extraction network to obtain spatiotemporal feature representations at different frequencies. Next, a feature fusion module interactively fuses the multi-scale EEG features and introduces a nonlinear attention mechanism to achieve weighted feature reconstruction. Finally, the fused EEG features are input into a classification module to output the classification results for the motion imagery task. This invention effectively extracts the spatiotemporal features of EEG signals at different frequencies, efficiently fuses and weights multi-scale nonlinear features, alleviating the problem of insufficient nonlinear fitting of EEG signals in motion imagery tasks, thereby achieving more accurate and reliable motion imagery classification.

[0092] 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. A brainwave signal classification method based on multi-scale causal convolution and KAN attention, characterized in that: Includes the following steps: Step 1: EEG Data Acquisition and Preprocessing: The experiment was set up according to the motor imagery experiment paradigm, and EEG data of the subjects were collected under the motor imagery task. After the experiment, the collected EEG data were labeled and preprocessed to generate EEG signal datasets under different motor task conditions. Step 2: EEG Data Augmentation: Take the EEG signal data from Step 1 and perform data augmentation by slicing and reconstructing the time series data to expand the diversity of the training samples; Step 3: Classification of motor imagery task based on EEG signals: Input the enhanced samples from Step 2 into the feature extraction module to extract the spatiotemporal features of the EEG signals; The extracted features are further weighted by the attention module to highlight key information; the weighted features are then classified by the classification module to achieve EEG classification of EEG signals under the motor imagery task. Specifically, it includes: S3.1: Perform preliminary spatiotemporal feature extraction on the enhanced EEG data. Extract spatiotemporal features of EEG at different scales through a multi-layer, multi-scale network, and integrate the information of each scale branch through feature fusion and interaction mechanisms. S3.2: Input the EEG spatiotemporal features extracted by the multi-scale network into the local-global nonlinear attention module. Obtain the weighted local and global features through an adaptive weighting mechanism, and then perform adaptive fusion through learnable parameters. Finally, input the results into the nonlinear classifier to output the classification results.

2. The EEG signal classification method based on multi-scale causal convolution and KAN attention as described in claim 1, characterized in that: Before the start of the above steps, a preliminary experiment is conducted to assess the potential impact of eye movement signals on EEG recordings. The formal experimental procedure includes: an attention cuing phase, a task cuing phase, a task execution phase, and a rest phase.

3. The EEG signal classification method based on multi-scale causal convolution and KAN attention as described in claim 1, characterized in that: In step one, the experiment is set up according to the motor imagery experiment paradigm. The procedure for each experiment is as follows: task preparation prompting stage, task prompting stage, participants perform the corresponding motor imagery task, and then enter the rest stage to restore the baseline state.

4. The EEG signal classification method based on multi-scale causal convolution and KAN attention as described in claim 1, characterized in that: The preprocessing in step one specifically involves: applying a 0.5-100Hz bandpass filter to the EEG signal to retain the effective frequency band relevant to the motion visualization task; followed by using a 50Hz notch filter to suppress power frequency noise interference. During the classification phase, only the EEG data in the [2s, 6s] interval of each experiment were selected as the effective analysis segment to ensure that the complete motor imagery task execution process was included and to reduce interference during the preparation and recovery periods. The preprocessed EEG data consisted of dimensions (N, 288, 22, 1000), where N represents the number of participants, 288 represents the number of experiments for each participant, 22 represents the number of channels, and 1000 corresponds to the number of sampling points of the 4s data at a sampling rate of 250Hz. Data augmentation was performed on the preprocessed EEG data. For the same subject's EEG data under the same motor imagery task, a segmentation-reconstruction strategy was adopted to generate more samples. In the segmentation stage, each EEG signal was divided into several time segments according to the time window. In the reconstruction stage, new artificial EEG samples were constructed by randomly recombining and splicing these time segments.

5. The EEG signal classification method based on multi-scale causal convolution and KAN attention as described in claim 1, characterized in that: S3.1 specifically includes: S3.11: Input EEG data is first modeled using multi-scale temporal causal convolution to capture short-term and long-term temporal features of EEG signals at different scales; S3.12: The preliminary EEG temporal features obtained in S3.11 are further input into the spatial and deep temporal feature extraction network. Spatial convolution is used to fuse the features between EEG channels. After multi-scale deep temporal convolution, more complex high-frequency temporal features are extracted from different scales. S3.13: The initial spatiotemporal features extracted by the multi-scale network are integrated through the convolutional fusion module.

6. The EEG signal classification method based on multi-scale causal convolution and KAN attention as described in claim 5, characterized in that: S3.2 specifically includes: S3.21: Input the multi-scale fused spatiotemporal features output from S3.13 into the local-global nonlinear attention module, adaptively weight the features, and obtain the weighted local and global features respectively. The nonlinear attention module is designed based on the KAN network, and its form is as follows: Where x represents the input feature, d in Indicates the input dimension, d out φ represents the output dimension, and φ represents the linear combination of the SiLU function and the B-spline function. φ(x)=ω b ·SiLu(x)+ω s ·spleen(x) in Let spline(x) represent the self-gated activation function. i c i B i (x) represents the B-spline basis function; The KAN network is constructed by stacking multiple φ(x) composite transformations: The global attention module combines the KAN network and the Softmax function to calculate the attention weight of each feature and multiplies it element-wise with the input features to obtain the globally weighted features; while the local attention module establishes the correlation between adjacent features through local convolution operations before calculating the attention weights, performs attention calculation and weighting, and better captures local information. Attention global =x0·Softmax(KAN(x0))+x0 Attention local =Conv(x0)·Softmax(KAN(Conv(x0)))+Conv(x0)S3.22: Adaptively fuse global attention features and local attention features through learnable parameters a to obtain a comprehensive EEG spatiotemporal feature representation; input the fused features into a nonlinear classifier to output the final classification result; X att =a·Attention global +(1-a)Attention local 。

Images