Classification Method for Motor Imagery EEG Signals

Abstract

The present invention provides a classification method for motor imagery electroencephalogram (EEG) signals, which includes: acquiring EEG signal data of a subject and performing preprocessing to obtain a training set; inputting the training set into a feature extraction model constructed based on a siamese neural network for pre-training to generate a feature set; inputting the feature set into a pre-classification model constructed based on a convolutional neural network, and performing pre-training and transfer learning respectively to obtain a classification model; after inputting the EEG signal data of a person to be tested into the classification model, the classification model executes a classification task and outputs a classification result, where the classification result at least includes a class label of the EEG signal. By using a triplet loss function for parameter learning, the present invention enables the extracted features to have the maximum inter-class distance and the minimum intra-class distance. The classification model greatly reduces the dependence on prior data, and optimizes the network according to the distance relationship between feature vectors for various motor imagery tasks, ensuring that the features between different classes have the maximum distinguishability.

Description

Technical Field

[0001] The present invention relates to the technical field of signal processing, and in particular to a classification method for motor imagery electroencephalogram signals. Background Art

[0002] Brain-computer interface technology designs motor imagery, uses an electroencephalogram cap to obtain brain waves, processes the brain waves with an algorithm to decode and identify the patient's motor intention, and transmits the identified signal to an external auxiliary medical device to realize limb movement, helping patients with limb movement disorders to achieve motor imagery training for rehabilitation of disabled limbs; current electroencephalogram signal decoding and identification algorithms usually use machine learning methods to extract features, and then send the extracted features to a classifier for classification and decoding. They require a large amount of prior data to extract effective features, which is time-consuming. Since existing algorithms are all based on a single subject, the data volume is complex and cannot represent the real data distribution, resulting in limited generalization performance of the trained model and affecting the authenticity and efficiency of classification.

[0003] CN113705324A discloses a multi-class motor imagery electroencephalogram signal feature extraction and classification method based on a decision tree and CSP-SVM. Aiming at the problems of complex feature space distribution and low discrimination degree of CSP in multi-class motor imagery feature extraction, it uses the classification result of SVM to select the optimal spatial filter constructed by "one-versus-rest" CSP, and uses the decision tree idea to transform the multi-classification problem into multiple binary classification problems to improve the accuracy of multi-class motor imagery electroencephalogram signal classification; however, this method cannot solve the problems of efficiently extracting features from a large amount of complex data and limited generalization performance of the model.

[0004] CN115630314A discloses an improved Inception network-based motor imagery electroencephalogram signal classification method, which solves the problem of insufficient feature extraction and improves the influence of overfitting on the classification result by improving the Inception network, but cannot improve the efficiency of effective feature extraction and the robustness of the model is not high. Summary of the Invention

[0005] The present invention proposes a classification method for motor imagery electroencephalogram signals, which solves the problems of low efficiency, inability to represent the real data distribution, poor generalization performance of the model, and low robustness caused by manually extracting features in traditional machine learning models.

[0006] According to the first aspect of the object of the present invention, a classification method for motor imagery electroencephalogram signals is provided, including:

[0007] Obtain the electroencephalogram signal data of the subject and perform preprocessing to obtain a training set;

[0008] Input the training set into the feature extraction model constructed based on the Siamese neural network for pre-training to generate a feature set;

[0009] Input the feature set into the pre-classification model constructed based on the convolutional neural network for pre-training and transfer learning respectively to obtain a classification model;

[0010] After inputting the electroencephalogram (EEG) signal data of the subject to be measured into the classification model, the classification model performs a classification task and outputs a classification result, which at least includes the class label of the EEG signal.

[0011] Furthermore, obtaining the training set includes:

[0012] Collect the EEG signal data of the subject's motor imagery. The EEG signal data of the subject's motor imagery at least includes N EEG signals performing the c-class motor imagery task; N is the number of signals in each category, and c is the category of the EEG signal performing the motor imagery task;

[0013] Perform normalization processing and EEG signal segment segmentation on the EEG signal data of the subject's motor imagery;

[0014] Combine the segmented EEG signal segments, and use the set formed by the combined training samples as the training set.

[0015] Furthermore, the normalization processing includes:

[0016]

[0017] Among them, represents the input EEG signal, is the mean value, is the standard deviation. Through normalization processing, the mean value of the EEG signals collected by each channel is 0.

[0018] Furthermore, the EEG signal segment segmentation includes:

[0019] Randomly sample a segment of EEG signal from the N EEG signals performing the c-class motor imagery task and define it as an anchor point;

[0020] Randomly sample a segment of EEG signal from the EEG data with the same category as the anchor point and define it as a positive example;

[0021] Randomly sample a segment of EEG signal from the data with a different category from the anchor point and define it as a negative example;

[0022] Then combine the anchor point, the positive example, and the negative example as a training sample;

[0023] Repeat the above steps of segmenting EEG signal segments until all the EEG signal data of the subject's motor imagery are combined into training samples.

[0024] Furthermore, the scale of the training set is C * N * (N - 1) * (N * (C - 1));

[0025] where C is the number of categories of EEG signals for performing motor imagery tasks, and N is the number of signals in each category.

[0026] Furthermore, obtaining the feature set includes:

[0027] Input the training set into the feature extraction model, and map the training set into feature vectors through a siamese neural network;

[0028] Use the triplet loss function to perform learning iterations on the input feature vectors. After the iteration ends, generate the feature set.

[0029] Furthermore, the expression form of the triplet loss function is:

[0030]

[0031] where represents the feature vector after mapping of the anchor signal, represents the feature vector after mapping of the positive example signal, represents the feature vector after mapping of the negative example, represents the minimum interval of the set feature vectors.

[0032] Furthermore, outputting the classification result includes:

[0033] Input the feature set into the pre-classification model, and map the feature set into class vectors through a convolutional neural network;

[0034] Use the binary cross-entropy function to perform transfer learning on the input class vectors for the binary classification task. After the learning iteration ends, output the class label of the EEG signal, which is the classification result.

[0035] Compared with the prior art, the beneficial effects achieved by the present invention:

[0036] 1. The feature extraction network in the present invention uses the triplet loss function for parameter learning, so that the features extracted by this network have the largest inter-class distance and the smallest intra-class distance, enabling the classification model to classify the features extracted by the above feature extraction network, realizing independent training of the two network models respectively, greatly reducing the dependence on prior data. For various motor imagery tasks, the network can be optimized according to the distance relationship between feature vectors, always ensuring the maximum distinguishability between features of different categories.

[0037] 2. The feature extraction network in the present invention uses its own data and public data sets to form a large-scale training set for training. The extracted feature dimensions can be set according to the model complexity of the classification network. The classification network is designed for specific classification tasks. According to the data volume of the specific subjects, corresponding machine learning classifiers (such as support vector machines, perceptrons) or more complex deep learning networks (such as MLP, CNN, Transformer) can be designed to increase the size of the training set through the combination of different EEG signals to improve the generalization performance of the model.

[0038] 3. The feature extraction network trained by the present invention on a large-scale data set has stronger robustness than the features extracted by traditional machine learning methods. The customized classification network can take into account both the task characteristics and the individual differences of the subjects, and adopts the transfer learning method to improve the adaptability to the differences in EEG signals, thereby improving the true accuracy of classification.

[0039] It should be understood that all combinations of the foregoing concepts and the additional concepts described in more detail below, as long as such concepts are not mutually inconsistent, can be considered part of the inventive subject matter of this disclosure. In addition, all combinations of the claimed subject matter are considered part of the inventive subject matter of this disclosure.

[0040] The foregoing and other aspects, embodiments, and features of the present invention will be more fully understood from the following description in conjunction with the accompanying drawings. Other additional aspects of the present invention, such as features and / or beneficial effects of the exemplary embodiments, will become apparent from the following description or through practice of specific embodiments according to the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings are not intended to be drawn to scale. In the accompanying drawings, each identical or nearly identical component shown in various figures may be represented by the same reference numeral. For clarity, not every component is labeled in every figure. Embodiments of various aspects of the present invention will now be described by way of example and with reference to the accompanying drawings.

[0042] Figure 1 Schematic diagram of the steps of the method for classifying motor imagery EEG signals shown in the present invention;

[0043] Figure 2 1 is a flow chart of a method for classifying motor imagery EEG signals according to the present invention;

[0044] Figure 3 This is a schematic diagram of data segmentation using a sliding window as shown in the present invention;

[0045] Figure 4 Schematic diagram of the feature extraction network shown in the present invention;

[0046] Figure 5 It is a schematic diagram of the classification network based on LENET-5 shown in the present invention. Detailed implementation manners

[0047] In order to better understand the technical content of the present invention, specific embodiments are hereby given and described in conjunction with the accompanying drawings as follows.

[0048] In the present disclosure, aspects of the present invention are described with reference to the accompanying drawings, in which many illustrative embodiments are shown. The embodiments of the present disclosure are not necessarily intended to cover all aspects of the present invention. It should be understood that the various concepts and embodiments introduced above, as well as those concepts and embodiments described in more detail below, can be implemented in any of a number of ways, because the concepts and embodiments disclosed in the present invention are not limited to any embodiment. Additionally, some aspects of the present invention can be used alone, or in any suitable combination with any other aspects of the present invention.

[0049] According to an embodiment of the present invention, in combination with Figure 1 as shown, a classification method for motor imagery electroencephalogram (EEG) signals includes the following steps:

[0050] S1: Obtain the EEG signal data of the subject and perform preprocessing to obtain a training set;

[0051] S2: Input the training set into a feature extraction model constructed based on a siamese neural network for pre-training to generate a feature set;

[0052] S3: Input the feature set into a pre-classification model constructed based on a convolutional neural network, and perform pre-training and transfer learning respectively to obtain a classification model;

[0053] S4: After inputting the EEG signal data of the person to be tested into the classification model, the classification model performs a classification task and outputs a classification result, and the classification result at least includes the class label of the EEG signal.

[0054] It should be noted that, in order to improve the operation accuracy of the feature extraction model and the classification model, the embodiments of the present invention perform pre-training on the above two models respectively. Among them, the classification model also adjusts the weight parameters through the method of transfer learning, thereby improving the execution accuracy of the binary classification task and avoiding the influence of model parameter errors.

[0055] Among them, as Figure 2 shown, the detailed processing steps are as follows:

[0056] (1) Collect the publicly available motor imagery EEG signal data of the subjects (rehabilitation training patients and healthy normal persons), and use it as the first EEG signal data, and obtain a first training set through the preprocessing method in the embodiments of the present invention;

[0057] (2) Inputting the first training set into the feature extraction model constructed based on the twin neural network for pre-training to obtain the first feature set;

[0058] (3) Inputting the first feature set into a pre-classification model built based on a convolutional neural network for pre-training, and outputting the category label of the EEG signal;

[0059] (4) Collecting the motor imagery EEG signals of healthy subjects and using them as the second EEG signal data, and using the preprocessing method in the embodiment of the present invention to obtain the second feature set;

[0060] (5) Inputting the second feature set into the pre-trained pre-classification model for transfer learning to adjust the weight parameters and obtain a classification model;

[0061] (6) collecting the motor imagery EEG signals of the subject and using them as the third EEG signal data, and using the preprocessing method in the embodiment of the present invention to obtain a third training set;

[0062] (7) Inputting the third training set into the feature extraction model to obtain a third feature set;

[0063] (8) Input the third feature set into the classification model to perform the classification task and output the classification result, which at least includes the category label of the EEG signal.

[0064] In one embodiment, the first EEG signal data is a data set in public literature or data published by all users (patients undergoing rehabilitation training, normal subjects), so as to be used for large-scale data training of the model in the embodiment of the present invention.

[0065] In one embodiment, the second EEG signal data is data provided by some subjects (healthy subjects, ten subjects in a group, at least three groups collected), which is used to verify and fine-tune the parameters of the model in the embodiment of the present invention to obtain a high-accuracy classification model.

[0066] In one embodiment, the third EEG signal data is data provided by a person who needs rehabilitation training, and is used in the method provided in the embodiment of the present invention to obtain the rehabilitation status of the person who needs rehabilitation training.

[0067] Preferably, the embodiment of the present invention obtains the data features of the EEG signal through the first model constructed by the twin neural network, which improves the accuracy of feature extraction and processing efficiency, and performs pre-training and transfer training respectively on the second model constructed by the convolutional neural network, and adjusts the appropriate weight parameters to ensure that the generated classification model accurately classifies and identifies the running imagination task of the EEG signal of the test person.

[0068] The following combinationFigures 3 to 5 The schematic diagram shown and some preferred or optional examples of the present invention will more specifically describe the implementation process and / or effects of certain examples of the present invention.

[0069]

Pretreatment

[0070] Collect the EEG signal data of the subject's motor imagery. The EEG signal data of the subject's motor imagery contains at least N EEG signals performing C-type motor imagery tasks.

[0071] Perform normalization processing and EEG signal segmentation on the EEG signal data of the subject's motor imagery.

[0072] Combine the segmented EEG signal segments, and use the set formed by the combined training samples as the training set.

[0073] Further, the EEG signal segmentation includes the following steps:

[0074] Randomly sample a segment of EEG signal from the N EEG signals performing c-type motor imagery tasks, and define it as the anchor.

[0075] Randomly sample a segment of EEG signal from the EEG data with the same category as the anchor, and define it as the positive example.

[0076] Randomly sample a segment of EEG signal from the data with a different category from the anchor, and define it as the negative example.

[0077] Then combine the anchor, positive example, and negative example as a training sample.

[0078] Repeat the above EEG signal segmentation steps until all the EEG signal data of the subject's motor imagery is completed for the combination of training samples.

[0079] In an optional embodiment, the above normalization processing is calculated by the following formula:

[0080]

[0081] Where, represents the input EEG signal, is the mean value, is the standard deviation. Through normalization, the mean value of the EEG signals collected by each channel is 0.

[0082] It should be noted that the amplitude range of the normalization processing in the embodiment of the present invention is the set error threshold (which can be customized according to actual application requirements), so as to ensure that each channel has the same weight after inputting into the first model, making the features extracted by the first model more balanced and effective, and effectively improving the learning speed of the model based on backpropagation.

[0083] In an alternative embodiment, the EEG signal segmentation is achieved by a sliding time window: set the sliding time window parameters, which at least include the window size and the sliding step parameter, and then segment the EEG signal in the way of Figure 2 .

[0084] For example: the original data RAW of 4 seconds (i.e., the EEG signal to be segmented), set the window size to 3.2s and the step to 0.2s, slide the time window, and segment out 5 time segments of 0 - 3.2s, 0.2 - 3.4s, 0.4 - 3.6s, 0.6 - 3.8s, and 0.8 - 4s to increase the data volume.

[0085] It should be noted that the embodiment of the present invention realizes the pre - processing method of segmenting the data on the entire time axis by setting the size of the time window and the sliding step to increase the data volume. This method can not only solve the problem that the deep learning model cannot be trained due to the small data volume, but also improve the generalization performance of the model.

[0086] In an alternative embodiment, the scale of the first training set is C * N * (N - 1) * (N * (C - 1)), where C is the number of categories of EEG signals for performing motor imagery tasks, and N is the number of signals for each category.

[0087]

Feature Extraction

[0088] Input the first training set into the feature extraction model, and map the first training set to feature vectors through a siamese neural network;

[0089] Use the triplet loss function to learn and iterate the input feature vectors. After the iteration ends, output the first feature set.

[0090] As an example, after the first training set is pre - processed, it is combined into training samples recognizable by the feature extraction model to train the feature extraction model constructed based on the siamese neural network. Each training sample is composed of three data, namely Anchor, Positive, and Negative, which form a training sample. It should be noted that the embodiment of the present invention further expands the scale of the training set samples through the combination of training samples, improving the generalization performance of the model.

[0091] Taking the two - class task of left and right hands as an example, the first training set includes M groups of left - hand and right - hand motor imagery EEG signals each. By randomly sampling a group of data as Anchor, sampling a signal of the same category as Anchor from the remaining signals as Positive, and a signal of a different category from Anchor as Negative, the combined training samples are (2 * M * (M - 1) * M) groups.

[0092] Reference Figure 4 In an optional embodiment, the twin neural network model adopted in the embodiment of the present invention uses the VGG16 network. The network structure is composed of 5 convolutional layer groups, 2 fully connected layers and an output layer. The layers are separated by max pooling layers, and the activation function of all hidden layers uses the ReLU function:

[0093] (1) Network structure: Conv – Conv - Max pooling - Conv – Conv - Max pooling -Conv - Conv – Conv - Max pooling - Conv - Conv – Conv - Max pooling - Conv -Conv – Conv - Max pooling – FC – FC – FC – Output;

[0094] (2) The convolutional kernel uses a one-dimensional convolutional kernel

[0095] In each max pooling layer, the number of channels is 64, 128, 256, 512, 512 (the number of channels in the convolutional layer doubles until it reaches 512 and then stops increasing);

[0096] The number of channels in the fully connected layers is 1024, 1024, 1024;

[0097] The number of channels in the output layer is 128;

[0098] (3) The pooling layer does not involve weight coefficients. The network has a total of 16 hidden layers. Among them, the convolutional layer and the pooling layer are responsible for feature extraction, and the fully connected layer is responsible for feature combination and screening;

[0099] (4) The loss function uses the Triplet function:

[0100]

[0101] Among them, represents the feature vector after the Anchor signal is mapped, represents the feature vector after the Positive signal is mapped, represents the feature vector after the Negative is mapped, represents the minimum interval of the set feature vector.

[0102]

Classification

[0103] Input the first feature set into the pre-classification model, and map the first feature set to a category vector through a convolutional neural network;

[0104] Perform transfer learning on the input category vector using the binary cross-entropy function for a binary classification task. After the learning iteration ends, output the category label of the EEG signal.

[0105] As an example, to better improve the classification accuracy of the pre-classification model, after the pre-training of the pre-classification model in the embodiment of the present invention, the EEG signal of the subject (the second EEG signal data in this embodiment) is obtained through an EEG cap as the training data for transfer learning. For example, when performing a motor imagery EEG signal classification task, the EEG signal data of the subject performing the left and right hand motor imagery tasks 15 times each in the training phase, after the learning iteration ends, a classification model is output.

[0106] In the embodiment of the present invention, the EEG signal data of the subject during the motor imagery task, through the operation of the input category vector by the classification model after transfer training, outputs the category label to which the EEG signal of the subject belongs through the output layer that fine-tunes the weight parameters.

[0107] In an alternative embodiment, the category label can be set in the form of numbers. For example, performing the left hand task is 0 and performing the right hand task is 1.

[0108] In one implementation, the transfer training in the embodiment of the present invention initializes the weight parameters of the output layer in the pre-classification model after pre-training, fixes the weight parameters of other layers, and after n learning iterations, obtains a classification model with the weight parameter transfer learning completed.

[0109] Refer to Figure 5 , in an alternative embodiment, the convolutional neural network model adopted in the embodiment of the present invention uses the LeNet-5 network:

[0110] (1) Network structure: Convolution - Pooling - Convolution - Pooling - Fully Connected

[0111] Obtain a feature map of 6 channels 28*28 (6@28*28) through the C1 convolutional layer, obtain a feature map of 6 channels 14*14 (6@14*14) through the S2 pooling layer, obtain a feature map of 16 channels 10*10 (16@10*10) through the C3 convolutional layer, obtain a feature map of  16 channels 5*5 (16@5*5) through the S4 pooling layer, obtain a feature vector of dimension 84 (Flatten@84) through the F5 and F6 fully connected layers, and through the output layer, obtain a probability vector corresponding to the number of categories to be recognized. The dimension with the highest probability is the classification result;

[0112] (2) The convolutional kernel uses a one-dimensional convolutional kernel

[0113] The number of channels in the convolutional layer is 6, 16, 120;

[0114] The number of channels in the fully connected layer is 84;

[0115] The number of channels in the output layer is 1;

[0116] (3) The loss function uses the binary cross - entropy function (binary classification task):

[0117]

[0118] where, is the task label of the i - th signal and is the probability that the category is after being judged by the classification model.

[0119] It should be noted that through the training of transfer learning in the embodiments of the present invention, the adaptability to different individuals is realized, the classification authenticity and accuracy of the classification model are improved, and the model robustness is increased.

[0120] As an optional embodiment, the pre - training in the embodiments of the present invention can be carried out in the pre - training manner of the existing technology, and it is not uniquely limited.

[0121] Preferably, the feature extraction network in the present invention is trained using its own data and public data sets to form a large - scale training set. The extracted feature dimension can be set according to the model complexity of the classification network. The classification network is designed for specific classification tasks and can design corresponding machine learning classifiers (such as support vector machines, perceptrons) or more complex deep - learning networks according to the data volume of specific subjects, so as to increase the scale of the training set through the combination of different EEG signals and improve the generalization performance of the model.

[0122] Preferably, the feature extraction network trained by the large - scale data set in the present invention has stronger robustness compared with the features extracted by traditional machine - learning methods. The customized classification network can take into account the task characteristics and individual differences of the subjects, and adopts the transfer learning method to improve the adaptability to EEG signal differences, thereby improving the true accuracy of classification.

[0123] Although the present invention has been disclosed above with preferred embodiments, it is not intended to limit the present invention. Those with ordinary knowledge in the technical field to which the present invention pertains can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be subject to what is defined by the claims.

Claims

1. A classification method for motor imagery electroencephalogram signals, characterized in that, include: Obtaining EEG signal data of subjects: Collecting publicly available motor imagery EEG signal data from rehabilitation training patients and healthy individuals to form a large-scale data set, which will be used as the first EEG signal data; Collecting the motor imagery EEG signals of healthy subjects and using them as the second EEG signal data; Collecting the motor imagery EEG signal data of the subject and using it as the third EEG signal data; Preprocessing the subject's EEG signal data to obtain a training set: preprocessing the first EEG signal data to obtain a first training set; Preprocessing the third EEG signal data to obtain a third training set; Inputting the first training set into a feature extraction model built based on a twin neural network for pre-training, mapping the first training set into a feature vector, and performing learning iterations on the substituted feature vector using a ternary loss function. After the iterations are completed, a first feature set is generated; Input the first feature set into a pre-classification model built based on a convolutional neural network for pre-training, and output the category label of the EEG signal; Preprocessing the second EEG signal data to obtain a second feature set, inputting the second feature set into a pre-trained pre-classification model for transfer learning to adjust weight parameters to obtain a classification model; Inputting the third training set into the feature extraction model to obtain a third feature set; For the third feature set, the classification model after transfer training operates on the input category vector, and outputs a classification result by fine-tuning the output layer of the weight parameters. The classification result at least includes the category label of the EEG signal.

2. The classification method for motor imagery EEG signals according to claim 1, wherein The training set is obtained, including: Collecting motor imagery EEG signal data of the subject, wherein the motor imagery EEG signal data of the subject includes at least N EEG signals of performing c-type motor imagery tasks, where N is the number of signals of each category, and c is the category of the EEG signals performing the motor imagery task; performing standardization processing and EEG signal segmentation on the subject's motor imagery EEG signal data; The segmented EEG signal segments are combined into training samples, and a set formed by the combined training samples is used as the training set.

3. The classification method for motor imagery EEG signals according to claim 2, wherein The standardization process includes: ； Among them, represents the input electroencephalogram (EEG) signal, is the mean value, is the standard deviation. Through normalization processing, the mean value of the EEG signals collected by each channel is 0.

4. The classification method for motor imagery EEG signals according to claim 2, wherein The EEG signal segmentation includes: Randomly sample a segment of EEG signal from N EEG signals performing c-type motor imagery tasks, which is defined as an anchor point; Randomly sample a segment of EEG signal from the EEG data of the same category as the anchor point, which is defined as a positive example; Randomly sample a segment of EEG signal from data that is different from the anchor point category, which is defined as a negative example; Then, the anchor point, the positive example, and the negative example are combined as a training sample; Repeat the above EEG signal segmentation steps until all the training samples of the subject's motor imagery EEG signal data are combined.

5. The classification method for motor imagery EEG signals according to claim 2, characterized in that, The scale of the training set is C*N*(N-1)*(N*(C-1)); Where C is the number of categories of EEG signals performing the motor imagery task, and N is the number of signals in each category.

6. The classification method for motor imagery EEG signals according to claim 1, wherein The expression of the ternary loss function is: ； Among them, represents the feature vector after the anchor signal is mapped, represents the feature vector after the positive example signal is mapped, represents the feature vector after the negative example is mapped, represents the minimum interval of the set feature vector.

7. The classification method for motor imagery EEG signals according to claim 1, characterized in that, Outputting the classification results includes: Inputting the feature set into the pre-classification model, and mapping the feature set into a category vector through a convolutional neural network; Use the binary cross-entropy function to perform transfer learning on the input category vectors for a binary classification task. After the learning iteration ends, output the category label of the EEG signal, which is the classification result.

Images