Motor imagery electroencephalogram signal classification method based on multi-scale convolutional neural network

Abstract

The invention discloses a motor imagery electroencephalogram signal classification method based on a multi-scale convolutional neural network, and belongs to the technical field of brain-computer interfaces, and the method comprises the following steps: carrying out band-pass filtering, channel selection and standardized preprocessing on an original EEG signal; a multi-scale convolutional neural network model is constructed, multi-resolution time features are extracted through a shallow-layer multi-scale convolutional block, spatial-temporal feature alignment and cross-scale splicing are realized through a middle-layer convolutional layer, and finally a classification probability is adaptively generated through a dynamic full-connection layer; an independent subject training strategy is adopted to optimize model parameters, and classification and recognition of motor imagery tasks are achieved through adaptive feature dimension processing. According to the method, the time-frequency features of EEG signals are effectively captured through multi-scale convolution kernel design, key brain region response is enhanced in combination with a feature fusion strategy of space attention guidance, individual difference influences are eliminated through dynamic parameter initialization, and the feature expression ability and cross-subject generalization performance of a classification model are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of brain-computer interface technology, and more specifically, to a motor imagery EEG signal classification method based on a multi-scale convolutional neural network. Background Art

[0002] Electroencephalogram (EEG) signals are the changes in electric waves formed by the sum of the postsynaptic potentials of a large number of neurons when the brain is active. They are the overall reflection of the electrophysiological activities of brain nerve cells on the cerebral cortex or scalp surface. People usually place electrodes on the human scalp to detect EEG signals and use related equipment to collect and process them.

[0003] A brain-computer interface (BCI) is a direct connection created between the human or animal brain and an external device to enable information exchange between the brain and the device. It bypasses peripheral nerves and muscles and directly establishes a new communication and control channel between the brain and external devices. It captures brain signals and converts them into electrical signals to achieve information transmission and control. Brain-computer interfaces can be divided into invasive and non-invasive based on the way they detect signals. Invasive ones require surgical implantation of electrodes, which is more risky. Non-invasive brain-computer interfaces are more popular and have a wider range of applications.

[0004] Motor imagery is a popular topic in brain-computer interfaces in recent years. The principle is that when people imagine motor movements, the EEG signals in specific areas and frequency bands of the brain will change, and event-related synchronization (ERS) and event-related desynchronization (ERD) phenomena will occur. By collecting EEG signals and using classification algorithms to capture the differences in signals corresponding to different actions, the user's motor imagery intention can be analyzed, and then corresponding control instructions can be generated.

[0005] In the study of motor imagery EEG signal classification, traditional methods mainly use traditional machine learning algorithms, including: Common Space Pattern (CSP) and its derived Filter Bank Common Space Pattern (FBCSP) are often used for feature extraction, and the extracted features will be sent to classifiers such as Linear Discriminant Analysis (LDA) and Support Vector Machine (SVM) to obtain classification results. However, these traditional machine learning algorithms rely on a large amount of data prior knowledge, it takes a long time to obtain prior knowledge, and the generalization ability of traditional classification models is limited.

[0006] With the rise of deep learning, more and more neural networks are being applied to feature extraction and classification of EEG signals, such as EEGNet, residual network, etc. With its powerful learning ability, deep learning models make the EEG feature extraction process no longer rely on prior knowledge of data, and usually have stronger generalization performance. Among them, convolutional neural network (CNN), as a representative of deep learning algorithms, can automatically learn useful features during training and build complex nonlinear classifiers.

[0007] In convolutional neural networks, there are two types: single-scale and multi-scale. Single-scale convolution extracts data features using a fixed scale, making it difficult to adapt to subject differences. Multi-scale convolution, on the other hand, can capture richer and more complex features by analyzing data at different scales and performs better, but there is a lack of existing related research.

[0008] Therefore, there is a need for a method for classifying motor imagery electroencephalogram signals based on a multi-scale convolutional neural network. Summary of the Invention

[0009] The present invention aims to provide a method for classifying motor imagery electroencephalogram signals based on a multi-scale convolutional neural network to solve the problems raised in the above background technology.

[0010] To achieve the above object, the present invention provides the following technical solution: A method for classifying motor imagery electroencephalogram signals based on a multi-scale convolutional neural network, specifically including the following steps:

[0011] Step S1, data preprocessing: Perform band-pass filtering, channel selection, data truncation, normalization, and standardization on the original EEG signal; specifically including:

[0012] Use a 4th-order Butterworth filter for 7 - 47 Hz band-pass filtering to eliminate electromyogram and power frequency interference;

[0013] Remove the electrooculogram artifact channels and retain 22 effective EEG channels;

[0014] Perform channel-level Z-score standardization on the truncated EEG channel data. The formula is:

[0015]

[0016] Where X is the original EEG channel data, that is, a specific value in the electroencephalogram data that has not been standardized after truncation; μ is the channel mean, used to measure the average level and central tendency of the channel data; σ is the standard deviation, used to reflect the dispersion of the data points relative to the mean.

[0017] Compared with the prior art, this data preprocessing method has the following advantages:

[0018] (1) In terms of filtering, it can accurately filter out interference. A 4th-order Butterworth filter is used for band-pass filtering in the range of 7 - 47 Hz, which can more accurately eliminate EMG and power frequency interference. Compared with some other types of filters, the Butterworth filter has a flat frequency response characteristic in the passband, which can retain the useful information in the original signal in the 7 - 47 Hz frequency band to the greatest extent, reduce signal distortion, and effectively suppress noise and interference signals outside this frequency band.

[0019] (2) In terms of channel selection, the electrooculogram (EOG) artifact channels are clearly removed, and 22 effective EEG channels are retained. This targeted channel selection method can effectively reduce the impact of artifacts such as EOG on subsequent analysis. Compared with some more general channel selection or removal methods, it is more targeted and scientific, can better highlight the real EEG signals related to brain activities, and improve the data quality.

[0020] (3) In terms of normalization and standardization, channel-level Z-score standardization is performed on the intercepted EEG channel data, which can unify the data to a standard scale with a mean of 0 and a standard deviation of 1. This makes the data of different channels comparable, eliminates data differences caused by factors such as different channel acquisition sensitivities, and is conducive to subsequent data analysis and processing.

[0021] Step S2: Construct a multi-scale convolutional neural network model: Extract temporal features through a shallow multi-scale convolutional block, fuse spatial features in the middle convolutional layer, and complete classification in the dynamic fully connected layer; specifically:

[0022] Shallow multi-scale convolutional block: It contains 3 groups of parallel convolutional layers and uses the ELU activation function; Parallelly extract temporal features using convolutional kernels of different scales. The receptive fields of convolutional kernels of different sizes are different, and they can capture the dynamic changes of EEG signals from different temporal resolutions; Provide rich and valuable temporal information for subsequent feature fusion and classification, and enhance the model's ability to capture the temporal characteristics of EEG signals;

[0023] Middle convolutional layer: It contains an independent convolutional layer with a small convolutional kernel in the middle, an independent convolutional layer with a large convolutional kernel in the middle, and a middle feature fusion layer; The first two layers perform feature extraction, and the middle feature fusion layer aligns the sizes and splices the channels of the multi-scale features output by the shallow layer to extract fused features;

[0024] Dynamic fully connected layer: Automatically initialize the fully connected parameters according to the input feature dimensions, and finally output various probability distributions; This layer can adapt to different input features, comprehensively analyze the features extracted by the previous layers, output the classification probabilities of the motor imagery tasks, and complete the classification and recognition of EEG signals.

[0025] Among them, each layer of the middle convolutional layer is specifically:

[0026] Middle - layer small convolutional kernel independent convolutional layer: Adjust the number of channels through 1×1 convolution to achieve linear combination and compression of features; perform batch normalization to accelerate model convergence; use the ELU activation function to enhance the model's expressive ability; use max - pooling operation for downsampling in the time dimension to further extract important features;

[0027] Middle - layer large convolutional kernel independent convolutional layer: Perform 1×3 convolution, batch normalization, ELU activation, and max - pooling operations to capture a wider range of context information, work in cooperation with the middle - layer small convolutional kernel independent convolutional layer, extract features from different scales and angles; integrate the different - scale time features output by the shallow - layer multi - scale convolutional block to improve the spatial resolution, enabling the model to learn more representative spatio - temporal fusion features, enhancing the model's expressive ability for EEG signal features, and providing more effective features for subsequent classification;

[0028] Middle - layer feature fusion layer: Calculate the minimum time dimension and channel dimension of each branch feature, intercept the aligned features; splice the multi - scale features along the channel dimension to form a fused feature tensor; then extract the fused features through 1×3 convolution.

[0029] Compared with the existing convolutional neural network models, this multi - scale convolutional neural network model has the following advantages:

[0030] (1) Using convolutional kernels with different receptive fields can capture the dynamic changes of EEG signals from different time resolutions, and the features extracted are richer and more comprehensive than those extracted by a single - scale convolutional kernel.

[0031] (2) Adopting a feature fusion mechanism can effectively integrate features of different scales and different aspects, make full use of the multi - scale time features extracted by the shallow layer, enable the model to learn more representative fused features, and provide more effective features for subsequent classification.

[0032] (3) In the dynamic fully - connected layer, adaptive feature processing is adopted. Automatically initialize the fully - connected parameters according to the input feature dimension, and be able to adapt to different input features. This enables the model to flexibly process feature data of different scales and dimensions, and has stronger adaptability and generalization ability compared with the fixed - structure fully - connected layer.

[0033] Step S3, Model training: Adopt an independent - subject training strategy to optimize the model parameters and minimize the loss function; specifically:

[0034] The first subject is trained for 1000 rounds, and subsequent subjects are trained for 800 rounds, with a batch size of 48;

[0035] Set the initial learning rate to 0.001, and use the Adam optimizer to adaptively adjust the learning rate to accelerate model convergence;

[0036] The loss function is the cross-entropy loss, and the formula is:

[0037]

[0038] where y ic is the true label, representing the true situation that the i-th sample belongs to the c-th class; p ic is the predicted probability, which is the probability value that the model predicts the i-th sample belongs to the c-th class.

[0039] Compared with the existing method, the training method of this model has the following advantages:

[0040] (1) Flexibly adjust the number of training rounds. By flexibly adjusting the number of training rounds according to different subjects, it can more fully consider the importance of the subject data for model initialization and exploration, and give more training rounds to fully learn the data features; while subsequent subjects can appropriately reduce the number of training rounds based on the previous training experience and model foundation, improving the training efficiency while ensuring the model performance, avoiding the problems of overtraining or under-training, and being more flexible and targeted than the method of fixed number of training rounds.

[0041] (2) For this classification problem, the cross-entropy loss is used as the loss function. Compared with some other loss functions, the cross-entropy loss has better optimization characteristics and convergence speed in dealing with classification problems, enabling the model to converge to the optimal solution faster, thereby improving the training effect and generalization ability of the model.

[0042] Step S4, classification output: Input the preprocessed EEG signal into the trained model to output the classification result of the motor imagery task; specifically:

[0043] Reshape the EEG signal preprocessed in step S1 into a four-dimensional tensor (batch × 1 × electrode channels × sampling points) format;

[0044] Input the reshaped EEG signal into the multi-scale convolutional neural network model trained in step S3: The model performs calculations in sequence according to the constructed network structure. The shallow multi-scale convolutional block extracts temporal features, the middle convolutional layer fuses spatial features, and the dynamic fully connected layer performs calculations based on these features;

[0045] The result output by the dynamic fully connected layer is the probability distribution of the motor imagery task, and the class with the largest probability value is selected as the final classification result for output, thus completing the classification recognition of the motor imagery task corresponding to the EEG signal.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] (1) In terms of method, the present invention adopts a multi-scale spatio-temporal feature fusion mechanism, which is different from the traditional single-scale convolution method. By parallelly extracting features through temporal convolution kernels of different scales, it breaks through the limitation of the limited receptive field of a single convolution kernel, and can capture rich temporal and spatial feature information in EEG signals more comprehensively and meticulously. For the classification of motor imagery EEG signals, it reduces the possibility of model overfitting, improves the accuracy and reliability of classification, and has good classification effects and generalizability.

[0048] (2) In terms of performance, the multi-scale convolutional neural network model combined with an optimized individual-subject training strategy overcomes many limitations of traditional machine learning algorithms in feature extraction and classification recognition tasks. Traditional algorithms rely on a large amount of prior knowledge, and the acquisition process is time-consuming and the generalization ability of the classification model is limited. However, the model of the present invention does not require a large number of manually designed features, can automatically learn data features, and reduces manual design work. At the same time, it is not easy to generate the phenomenon of gradient disappearance or explosion during the training process, can effectively improve the generalization ability of the model among different individuals, and still maintains a high classification accuracy and stability when facing diverse EEG signals of subjects.

[0049] (3) In terms of practicality, the method of the present invention can be applied to the classification of motor imagery EEG signals, provides reliable classification results for the brain-computer interface system, and has important practical value in practical applications. Its output results can be used to accurately control external devices, such as assisting users in rehabilitation training, controlling wheelchairs, etc., and can effectively improve the quality of life and mobility of users.

[0050] (4) In terms of scalability, the structure and method of the model have a certain degree of scalability. Subsequently, according to different application scenarios and requirements, the model can be further optimized and adjusted to adapt to a wider range of EEG signal processing tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is the system flowchart of the present invention,

[0052] Figure 2 is the technical roadmap of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0053] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0054] In addition, in the present invention, descriptions such as "first", "second", etc. are only for descriptive purposes and should not be construed as indicating or implying their relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first", "second" may explicitly or implicitly include at least one of such features. In the description of the present invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0055] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the invention.

[0056] Embodiment: Refer to Figure 1 、 Figure 2 , a method for classifying motor imagery electroencephalogram (EEG) signals based on a multi-scale convolutional neural network, specifically including the following steps:

[0057] (1) Data preprocessing:

[0058] First, a 4th-order Butterworth filter is used for band-pass filtering in the range of 7 - 47 Hz to eliminate electromyogram and power frequency interference; subsequently, the electrooculogram artifact channels are removed, and 22 effective EEG channels are retained; finally, channel-level Z-score normalization is performed on the intercepted EEG channel data, and the formula is:

[0059]

[0060] where μ is the channel mean and σ is the standard deviation.

[0061] (2) Model construction:

[0062] The present invention proposes an efficient multi-scale convolutional neural network model for classifying EEG signals. The model includes a shallow multi-scale convolutional block, a middle convolutional layer, and a dynamic fully connected layer.

[0063] Shallow multi-scale convolutional block: It contains 3 groups of parallel convolutional layers with convolutional kernel sizes of 1×3, 1×5, and 1×7 respectively, a stride of 2, and the ELU activation function is used; parallel extraction of temporal features is performed using convolutional kernels of different scales. The receptive fields of convolutional kernels of different sizes are different, and they can capture the dynamic changes of EEG signals from different temporal resolutions; it provides rich and valuable temporal information for subsequent feature fusion and classification, and enhances the model's ability to capture the temporal characteristics of EEG signals;

[0064] Middle convolutional layer: It includes an independent convolutional layer with small middle - sized convolutional kernels, an independent convolutional layer with large middle - sized convolutional kernels, and a middle - layer feature fusion layer; the first two layers are used for feature extraction, and the middle - layer feature fusion layer performs size alignment and channel splicing on the multi - scale features output by the shallow layer to extract fused features;

[0065] Dynamic fully - connected layer: Automatically initializes the fully - connected parameters according to the input feature dimension and finally outputs various probability distributions; this layer can adapt to different input features, comprehensively analyzes the features extracted by the previous layers, outputs the classification probabilities of the motor imagery task, and completes the classification and recognition of EEG signals.

[0066] Among them, the middle convolutional layer is specifically as follows:

[0067] Independent convolutional layer with small middle - sized convolutional kernels: Adjusts the number of channels through 1×1 convolution to achieve linear combination and compression of features; performs batch normalization to accelerate model convergence; uses the ELU activation function to enhance the model's expressive ability; uses max - pooling operation for down - sampling in the time dimension to further extract important features;

[0068] Independent convolutional layer with large middle - sized convolutional kernels: Performs 1×3 convolution, batch normalization, ELU activation, and max - pooling operations to capture more extensive context information, works in cooperation with the independent convolutional layer with small middle - sized convolutional kernels, extracts features from different scales and angles; integrates the different - scale time features output by the shallow - layer multi - scale convolutional blocks, improves the spatial resolution, enables the model to learn more representative spatio - temporal fusion features, enhances the model's expressive ability for EEG signal features, and provides more effective features for subsequent classification;

[0069] Middle - layer feature fusion layer: Calculates the minimum time dimension and channel dimension of each branch feature, intercepts the aligned features; splices the multi - scale features along the channel dimension to form a fused feature tensor; then extracts the fused features through 1×3 convolution.

[0070] (3) Model training and classification output:

[0071] The embodiment of the present invention uses the publicly available EEG dataset BCI Competition IV 2a for model training. The BCI Competition IV 2a dataset collected data of 22 EEG channels from 9 subjects in two different stages. Each subject participated in four different motor imagery tasks, including motor imagery of the left hand, right hand, both feet, and tongue. Each round of experiment repeated 6 groups of actions with a short break in the middle. Each group of experimental data contained 48 motor imagery data (four types of actions, each type repeated 12 times), and each round produced a total of 288 experimental data segments. At the same time, this article considered the experimental data between [2,6] seconds of each experiment in the experiment.

[0072] Set the initial learning rate to 0.001, and use the Adam optimizer to adaptively adjust the learning rate to accelerate the convergence of the model; the loss function is the cross-entropy loss, and the formula is:

[0073]

[0074] where y ic is the true label, representing the true situation that the i-th sample belongs to the c-th class; p ic is the predicted probability, which is the probability value that the model predicts the i-th sample belongs to the c-th class.

[0075] Randomly initialize the weights of the CNN model, and train the data of 9 subjects respectively:

[0076] The first subject is trained for 1000 rounds, and the subsequent subjects are trained for 800 rounds, with a batch size of 48; for the data of each batch, first adjust the data dimension, move the data and labels to the specified device; clear the gradients of the optimizer, perform forward propagation on the input data, and calculate the model output; calculate the cross-entropy loss according to the actual labels and predicted probabilities; then, calculate the gradients through the backpropagation algorithm and update the model parameters using the optimizer; obtain the prediction results, and calculate and record the accuracy and loss function values.

[0077] After independently training the model using the EEG signal data of each subject, save the trained model parameters to obtain the model corresponding to each subject. Furthermore, the test set data can be used to test and verify the corresponding model. In this way, the existing model can be further optimized in terms of parameters according to the test results to improve the model performance.

Claims

1. A motor imagery EEG signal classification method based on a multi-scale convolutional neural network, characterized in that: The following steps are involved: Step S1, data preprocessing: bandpass filtering, channel selection, data interception, normalization and standardization of the original EEG signal; Step S2, construct a multi-scale convolutional neural network model: extract temporal features through shallow multi-scale convolution blocks, fuse spatial features through middle convolution layers, and complete classification through dynamic fully connected layers; Step S3, model training: using independent subject training strategy to optimize model parameters and minimize loss value; Step S4, classification output: input the preprocessed EEG signal into the trained model and output the classification result of the motor imagery task.

2. The method for classifying motor imagery EEG signals based on a multi-scale convolutional neural network according to claim 1, characterized in that: Step S1 data preprocessing specifically includes: Use a 4th-order Butterworth filter to perform 7-47Hz bandpass filtering to eliminate myoelectric and power frequency interference; Remove the electrooculogram artifact channel and retain 22 valid EEG channels; Channel-level Z-score normalization is performed on the truncated EEG channel data, and the formula is: Where X is the original EEG channel data, that is, a specific value in the EEG data that has not been standardized after being intercepted; μ is the channel mean, which is used to measure the average level and central tendency of the channel data; σ is the standard deviation, which is used to reflect the dispersion of data points relative to the mean.

3. The method for classifying motor imagery EEG signals based on a multi-scale convolutional neural network according to claim 1, characterized in that: The multi-scale convolutional neural network model in step S2 includes: Shallow multi-scale convolution block: contains 3 groups of parallel convolution layers, with convolution kernel sizes of 1×3, 1×5, and 1×7, a step size of 2, and an ELU activation function. It uses convolution kernels of different scales to extract temporal features in parallel. Convolution kernels of different sizes have different receptive fields and can capture the dynamic changes of EEG signals from different temporal resolutions. It provides rich and valuable temporal information for subsequent feature fusion and classification, and enhances the model's ability to capture the temporal characteristics of EEG signals. Middle convolution layer: includes independent convolution layer with small convolution kernel, independent convolution layer with large convolution kernel and feature fusion layer; the first two layers perform feature extraction, and the feature fusion layer performs size alignment and channel splicing on the multi-scale features output by the shallow layer to extract fusion features; Dynamic fully connected layer: Automatically initializes fully connected parameters according to the input feature dimension, and finally outputs various probability distributions; this layer can adapt to different input features, comprehensively analyze the features extracted by the previous layers, output the classification probability of the motor imagery task, and complete the classification and recognition of EEG signals.

4. The method for classifying motor imagery EEG signals based on a multi-scale convolutional neural network according to claim 3, characterized in that: The middle convolutional layer is specifically: Middle-layer small convolution kernel independent convolution layer: adjust the number of channels through 1×1 convolution to achieve linear combination and compression of features; perform batch normalization to accelerate model convergence; use ELU activation function to enhance the expressiveness of the model; use maximum pooling operation to downsample in the time dimension to further extract important features; Middle-layer large convolution kernel independent convolution layer: performs 1×3 convolution, batch normalization, ELU activation and maximum pooling operations to capture a wider range of temporal information, and works with the middle-layer small convolution kernel independent convolution layer to extract features from different scales and angles; Integrate the different scale temporal features output by the shallow multi-scale convolutional blocks, so that the model can learn more representative fusion features, enhance the model's ability to express EEG signal features, and provide more effective features for subsequent classification; Middle feature fusion layer: calculate the minimum time dimension and channel dimension of each branch feature, extract the aligned features; splice multi-scale features along the channel dimension to form a fused feature tensor; and then extract the fused features through 1×3 convolution.

5. The method for classifying motor imagery EEG signals based on a multi-scale convolutional neural network according to claim 1, characterized in that: The independent subject training strategy of step S3 includes: The first subject was trained for 1000 rounds, and the subsequent subjects were trained for 800 rounds, with a batch size of 48; Set the initial learning rate to 0.001, and use the Adam optimizer to adaptively adjust the learning rate to accelerate model convergence; The loss function is cross entropy loss, and the formula is: where y ic is the true label, representing the true situation that the i-th sample belongs to the c-th class; p ic is the predicted probability, which is the probability value that the model predicts that the i-th sample belongs to the c-th class.

6. The method for classifying motor imagery EEG signals based on a multi-scale convolutional neural network according to claim 1, characterized in that: The classification output of step S4 is specifically: Reshape the EEG signal after preprocessing in step S1 into a four-dimensional tensor (batch × 1 × electrode channel × sampling point) format; The reshaped EEG signal is input into the multi-scale convolutional neural network model trained in step S3: the model performs calculations in sequence according to the constructed network structure. The main body of the network adopts a parallel branch structure. The shallow multi-scale convolutional blocks extract features. The middle convolutional layer fuses the features output by convolution kernels of different sizes. The dynamic fully connected layer performs calculations based on these features. The result output by the dynamic fully connected layer is the probability distribution of the motor imagery task. The category with the largest probability value is selected as the final classification result for output, thereby completing the classification and recognition of the motor imagery task corresponding to the EEG signal.

7. A motor imagery EEG signal classification system based on a multi-scale convolutional neural network, characterized in that: The system is used to implement the method for classifying motor imagery EEG signals based on a multi-scale convolutional neural network as described in any one of claims 1 to 6, and comprises: A data preprocessing module, used for preprocessing MI-EEG signals; Model building module, used to build multi-scale convolutional neural network models; Model training module, used to train the model using preprocessed data; Signal classification module, used to classify new MI-EEG signals using the trained model.

8. A machine-readable storage medium, characterized in that: The machine-readable storage medium stores machine-executable instructions. When the machine-executable instructions are called and executed by the processor, the machine-executable instructions prompt the processor to implement the motor imagery EEG signal classification method based on a multi-scale convolutional neural network as described in any one of claims 1-6.

9. An electronic device, characterized in that: It includes a processor and a memory, wherein the memory stores machine executable instructions that can be executed by the processor, and the processor executes the machine executable instructions to implement the motor imagery EEG signal classification method based on a multi-scale convolutional neural network as described in any one of claims 1-6.

Images