CN110796175A - An online classification method of EEG data based on lightweight convolutional neural network - Google Patents

Abstract

The invention discloses an electroencephalogram data online classification method based on a light-weight convolutional neural network, which is applied to a cloud service platform, wherein the cloud service platform comprises a sensing layer, a gateway and a cloud end, and firstly, electroencephalogram data of a user are collected through the sensing layer; then, transmitting the acquired electroencephalogram data into a gateway, and downloading a trained classifier model from a cloud end through the gateway; and then, carrying out online classification on the acquired electroencephalogram data based on the trained classifier model, and uploading the EEG segments to a cloud server after being calibrated by a doctor for an incremental training model. The method can be directly applied to the original EEG without preprocessing and feature extraction, and has high classification result precision and real-time performance.

Description

An online classification method of EEG data based on lightweight convolutional neural network

technical field

The invention relates to the technical field of computers, in particular to an online classification method of EEG data based on a lightweight convolutional neural network.

Background technique

Fast and accurate online classification of electroencephalogram (EEG) is a prerequisite for brain-computer interface, neurofeedback application level-1 brain disease monitoring system. Accurate assessment and timely assessment of brain status can greatly reduce the risk of patients with brain disorders. Therefore, online EEG classification has always been a research hotspot for neuroscience researchers and clinicians.

In the prior art, the collected EEG data is usually classified after denoising and feature extraction.

In the process of implementing the present invention, the inventor of the present application found that the method of the prior art has at least the following technical problems:

Because EEG data is susceptible to noise interference and the evolution of brain states is non-stationary, the classification performance strongly depends on data preprocessing (removing noise and human interference) and feature extraction (extracting EEG key information). How to classify high-dimensional weak EEG signals with strong noise interference is a huge challenge.

It can be seen from this that the methods in the prior art have technical problems of complex implementation and slow classification speed.

SUMMARY OF THE INVENTION

In view of this, the present invention provides an online classification method for EEG data based on a lightweight convolutional neural network, to solve or at least partially solve the technology in the prior art with complex implementation and slow classification speed. question.

In order to solve the above technical problems, the present invention provides an online classification method of EEG data based on a lightweight convolutional neural network, which is applied to a cloud service platform. The cloud service platform includes a sensor layer, a gateway and a cloud. The online classification method Methods include:

Step S1: collecting the user's EEG data through the sensing layer;

Step S2: The collected EEG data is transmitted to the gateway, and the trained classifier model is downloaded from the cloud through the gateway. After designing the lightweight convolutional neural network structure, the trained classifier model adopts the back-propagation-based method. The mini-batch stochastic gradient descent method is trained;

Step S3: online classification of the collected EEG data based on the trained classifier model.

In one embodiment, the structure of designing a lightweight convolutional neural network in step S2 includes one dropout layer, two convolutional layers with the same convolutional domain, one pooling layer and three fully connected layers, wherein three The number of neurons in each fully connected layer is reduced in a preset manner.

In one embodiment, the activation function used by the Dropout layer and the first convolutional layer is Relu,

relu(x)=max(0,x)

Among them, x is a variable;

The second convolutional layer, pooling layer, and three fully connected layers all use the Sigmoid activation function. The last activation function of Lightweight, Sigmoid, identifies the classification results of specific EEG segments. The form of the Sigmoid activation function is as follows:

where x is a variable.

In one embodiment, the method further includes canceling the activation function of the second convolutional layer.

In one embodiment, in step S2, a back-propagation-based mini-batch stochastic gradient descent method is used for training, including:

The sample space is divided into training set, validation set and test set in advance according to a preset ratio, among which, the training set and validation set are stored in the cloud, and the test set is stored in the gateway;

The neural network structure is trained through the training set, wherein the weight update rules used in the training process are:

where i is the number of iterations, ν is the momentum term variable, ε is the learning rate, Represents the average value of the target derivative of the _i -th batch Di when the weight is ω _i , ω _i represents the weight in the i-th iteration process, and ω _i+1 represents the i+1-th iteration. The weight in the process.

In one embodiment, the method further includes: the cloud is used to incrementally train the classifier model offline by collecting the doctor-calibrated EEG data from the gateway.

In one embodiment, step S3 specifically includes:

Step S3.1: After segmenting the collected EEG data, organize it into a three-dimensional format of channel cascade;

Step S3.2: Process the data in the three-dimensional format of the channel cascade through the Dropout layer;

Step S3.3: Perform a convolution operation on the data processed in step S3.2 through two convolution layers with the same convolution domain, wherein the convolution kernels of the two convolution layers are both 3×3, and the convolution The operation is expressed as:

表示第l层中第i个和第j个特征图之间的卷积核；b_i ^(l)代表第l层中的第i个特征图的偏置项；In the formula, a _i ^(l) and a _i ^(l-1) represent the ith output channel of the lth layer and the ith output channel of the l-1th layer, respectively;

represents the convolution kernel between the i-th and j-th feature maps in the l-th layer; b _i ^(l) represents the bias term of the i-th feature map in the l-th layer;

Step S3.4: perform a pooling operation on the convolved data through a pooling layer;

Step S3.5: After the pooled data is expanded into a 1*14112 vector, it is processed through three fully connected layers, wherein the output sizes of the three fully connected layers are 250, 60 and 1, respectively. The join operation is expressed as follows:

表示第l-1层的第k个神经元连接到第l层的第j个神经元的权重；b_j ^(l)表示第l层的第j个神经元的偏置，在全连接层使用的激活函数为sigmoid，最后一层激活函数的输出为对用户脑电数据片段的分类结果。a _j ^(l) and a _j ^(l-1) represent the output of the ith neuron in the lth layer and the l-1th layer, respectively;

Represents the weight of the kth neuron in the l-1th layer connected to the jth neuron in the lth layer; b _j ^(l) represents the bias of the jth neuron in the lth layer, used in the fully connected layer The activation function of sigmoid is sigmoid, and the output of the last layer of activation function is the classification result of the user's EEG data segment.

The above-mentioned one or more technical solutions in the embodiments of the present application have at least one or more of the following technical effects:

Thanks to the method provided by the present invention, after collecting the user's EEG data through the sensing layer, the collected EEG data is transmitted to the gateway, and the trained classifier model is downloaded from the cloud through the gateway; The collected EEG data is classified online. The trained classifier model in the cloud is based on the design of a lightweight convolutional neural network structure and then trained using the back-propagation-based mini-batch stochastic gradient descent method. Since the Lightweight Convolutional Neural Network (LightweightCNN) uses as few layers as possible and can obtain high classification accuracy, and the trained classifier model is stored in the cloud, the gateway can directly download the corresponding classifier model from the cloud to achieve online Therefore, it does not need to go through complex preprocessing and feature processing, which improves the classification efficiency and realizes real-time performance.

Further, by designing high convolutional layers to process high-dimensional EEG signals, minimizing the number of CNN hidden layers, and setting an hourglass-type fully connected layer, the closer to the output layer, the number of neurons in the fully connected layer is nearly linear. Drop, like the edge of an hourglass, to quickly reduce the connection parameters of the neural network and improve training performance.

Further, to solve the problem of unbalanced classification due to unbalanced samples in the neural network, it is solved by adjusting the number of activation functions.

Further, the cloud can continuously optimize the classifier by collecting doctor-calibrated EEG data from the gateway for incremental offline training of the classifier model.

Description of drawings

In order to illustrate the embodiments of the present invention or the technical solutions in the prior art more clearly, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are For some embodiments of the present invention, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without creative efforts.

1 is a schematic flowchart of an online classification method for EEG data based on a lightweight convolutional neural network provided by the present invention;

Fig. 2 is the working flow chart of the brain health care cloud platform of online EEG classification in the embodiment of the present invention;

FIG. 3 is a schematic structural diagram of a lightweight convolutional neural network in an embodiment of the present invention.

Detailed ways

The purpose of the present invention is to realize the complex and slow classification method for EEG data classification method caused by the prior art classification method which strongly relies on data preprocessing (removing noise and human interference) and feature extraction (extracting EEG key information). In order to solve the technical problem, an online classification method of EEG based on lightweight convolutional neural network is proposed, which aims to directly classify the original EEG signal quickly and accurately without preprocessing and feature extraction.

For achieving the above object, main design of the present invention is as follows:

Design a lightweight convolutional neural network, process high-dimensional EEG signals through high convolutional layers, minimize the number of CNN hidden layers, and design an "hourglass" fully connected layer block to quickly reduce the number of output neurons. . Aiming at the problem of unbalanced classification of the neural network due to unbalanced samples, the present invention solves the problem by adjusting the number of activation functions. The cloud platform includes three parts: the sensing layer, the cloud and the gateway. The sensing layer is used to directly collect the user's EEG data, the cloud is responsible for training the classifier, and the gateway downloads the trained classifier from the cloud and enters the EEG segment directly. Online classification and real-time display of classification results on smart devices. The EEG clips are calibrated by doctors and uploaded to the cloud server for incremental training of the model. The invention can be directly applied to the original EEG without preprocessing and feature extraction, and the classification result has high precision and real-time performance.

In order to make the purposes, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments These are some embodiments of the present invention, but not all 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.

This embodiment provides an online classification method for EEG data based on a lightweight convolutional neural network, see FIG. 1 , and the method includes:

Step S1: Collect the user's EEG data through the sensing layer.

Specifically, the sensing layer can be equipped with a collection device to collect data, such as EEG data, such as sensors.

Step S2: The collected EEG data is transmitted to the gateway, and the trained classifier model is downloaded from the cloud through the gateway. After designing the lightweight convolutional neural network structure, the trained classifier model adopts the back-propagation-based method. Trained with mini-batch stochastic gradient descent.

Specifically, the gateway in the embodiment of the present invention is used to interact with the user. After the EEG data collected by the sensor layer is transmitted to the gateway, the gateway also downloads the trained classifier model from the cloud. Lightweight Convolutional Neural Network (LightweightCNN) is a neural network with a simplified structure, which uses as few layers as possible to obtain high classification accuracy at the same time.

Step S3: online classification of the collected EEG data based on the trained classifier model.

Specifically, online classification is carried out at the gateway. After online classification, the results can be visualized through computers, tablets, mobile phones, etc., and users can be reminded in real time. The gateway in this application is an intelligent gateway, and the intelligent gateway must be able to satisfy advanced functions such as real-time response, processing data cache, and load balancing, which also poses a challenge to its design.

In the specific implementation process, the infrastructure of the intelligent gateway mainly includes three parts, the operating system, the flowing EEG database and the network protocol to provide support for the fog computing layer. The fog computing layer is an important part of the intelligent gateway, which contains the user-facing API of the brain health service. These services mainly include user management (opening different permissions for different user groups), online EEG classification (downloading the latest classifier from the cloud for real-time status classification), visualization, real-time reminders, etc.

In one embodiment, the structure of designing a lightweight convolutional neural network in step S2 includes one dropout layer, two convolutional layers with the same convolutional domain, one pooling layer and three fully connected layers, wherein three The number of neurons in each fully connected layer is reduced in a preset manner.

Specifically, please refer to Figure 3, which is the structure diagram of the lightweight convolutional neural network, Conv1 represents the first convolutional layer, Conv2 represents the second convolutional layer, Maxpooling represents the pooling layer, FC1, FC2, FC3 respectively Indicates the fully connected layer, the lightweight convolutional neural network in this application, the entire neural network has only 7 layers, far less than the existing deep CNN model.

In one embodiment, the activation function used by the Dropout layer and the first convolutional layer is Relu,

relu(x)=max(0,x)(1)

Among them, x is a variable;

The second convolutional layer, pooling layer, and three fully connected layers all use the Sigmoid activation function. The last activation function of Lightweight, Sigmoid, identifies the classification results of specific EEG segments. The form of the Sigmoid activation function is as follows:

where x is a variable.

Specifically, high convolutional layers handle high-dimensional EEG segments by placing depthwise convolutional filters on one convolutional layer. For each time window, the data sequence (X) of each electrode is matrixed into a matrix (A×B), and then the entire EEG data (the number of channels is Z) is organized into a three-dimensional data through channel cascade Blocks (A×B×Z), put into convolutional layers. LightweightCNN starts with a dropout layer, then connects two convolutional layers with the same convolutional domain (3×3) and a Maxpooling layer (1×1), and finally three fully connected layers.

In one embodiment, the method further includes canceling the activation function of the second convolutional layer.

Specifically, in EEG classification, training samples cannot always be guaranteed to be balanced, which affects the performance of classification. In neural networks, nonlinear fitting ability is largely dependent on multiple activation functions. The inventor of the present application has found through a large number of experiments and research that the best performance of unbalanced learning can be achieved by canceling the activation function of the second convolutional layer. Generally speaking, there are two metrics to verify the ability of a classifier to handle imbalanced samples, F-score and GMEAN.

The formula of F-score is:

Among them, precision represents precision, recall represents recall, and β represents a parameter that determines the weighted importance of recall and precision. The formula of GMEAN is:

Sensitivity means sensitivity and Specificity means specificity.

By canceling the activation function of the second convolutional layer, the best unbalanced learning performance can be obtained through the calculation of the above two indicators.

In the embodiment of the present invention, the size of the convolution kernel of the convolution layer is set to 3×3, the purpose is to perform convolution with the nearest neighbor of each element, so as to better maintain the structural information of the EEG data.

The time complexity of a neural network is proportional to the product of the number of hidden neurons N and the number of layers L. The time complexity of all convolutional layers can be expressed as:

分别表示卷积核和feature map(特征图)的空间大小。1 indicates the number of convolution layers, d indicates the number of layers, n _l indicates the number of convolution kernels in the first layer, n _l-1 indicates the number of input channels in the first layer, and

Represent the spatial size of the convolution kernel and feature map (feature map), respectively.

It should be noted that the EEG data contains multiple channels, each channel collects one data at a time point (each data is called an element), and a time series of data is collected within a period of time.

In one embodiment, in step S2, a back-propagation-based mini-batch stochastic gradient descent method is used for training, including:

The sample space is divided into training set, validation set and test set in advance according to a preset ratio, among which, the training set and validation set are stored in the cloud, and the test set is stored in the gateway;

The neural network structure is trained through the training set, wherein the weight update rules used in the training process are:

表示在权值为ω_i时，第i批D_i的目标导数平均值，ω_i表示第i次迭代过程中的权值，ω_i+1表示第i+1次迭代过程中的权值。where i is the number of iterations, ν is the momentum term variable, ε is the learning rate,

Represents the average value of the target derivative of the _i -th batch Di when the weight is ω _i , ω _i represents the weight in the i-th iteration process, and ω _i+1 represents the i+1-th iteration. The weight in the process.

Specifically, during the training process, the feedforward network obtains the output of the current layer by connecting the topology structure and the activation function, and the feedback network uses the BP algorithm according to the topology structure to adjust the model connection weight based on the residual between the output and the target based on stochastic gradient descent. Optimize; the chain rule for the learning of connection weights is to output to the current connection start point for chain update, and the offset is only updated by the chain rule between adjacent layers, regardless of the cross-layer Impact.

In addition, the optimized classification model is obtained after training. After the test data is put into the model, the learned parameters can be used to learn features, and finally the EEG data (such as epilepsy seizure period, etc.) can be predicted and classified according to each feature.

In one embodiment, the method further includes: the cloud is used to incrementally train the classifier model offline by collecting the doctor-calibrated EEG data from the gateway.

After the model is trained, the smart gateway can download the latest classifier from the cloud and test it on the test set without modifying the parameters. The cloud server incrementally trains new corrected EEG data from the gateway to continuously optimize new models. Through incremental offline training of the classifier model, the classifier can be continuously optimized to obtain more accurate classification prediction results.

In one embodiment, step S3 specifically includes:

Step S3.1: After segmenting the collected EEG data, organize it into a three-dimensional format of channel cascade;

Step S3.2: Process the data in the three-dimensional format of the channel cascade through the Dropout layer;

Step S3.3: Perform a convolution operation on the data processed in step S3.2 through two convolution layers with the same convolution domain, wherein the convolution kernels of the two convolution layers are both 3×3, and the convolution The operation is expressed as:

表示第l层中第i个和第j个特征图之间的卷积核；b_i ^(l)代表第l层中的第i个特征图的偏置项；In the formula, a _i ^(l) and a _i ^(l-1) represent the ith output channel of the lth layer and the ith output channel of the l-1th layer, respectively;

represents the convolution kernel between the i-th and j-th feature maps in the l-th layer; b _i ^(l) represents the bias term of the i-th feature map in the l-th layer;

Step S3.4: perform a pooling operation on the convolved data through a pooling layer;

Step S3.5: After the pooled data is expanded into a 1*14112 vector, it is processed through three fully connected layers, wherein the output sizes of the three fully connected layers are 250, 60 and 1, respectively. The join operation is expressed as follows:

表示第l-1层的第k个神经元连接到第l层的第j个神经元的权重；b_j ^(l)表示第l层的第j个神经元的偏置，在全连接层使用的激活函数为sigmoid，最后一层激活函数的输出为对用户脑电数据片段的分类结果。a _j ^(l) and a _j ^(l-1) represent the output of the ith neuron in the lth layer and the l-1th layer, respectively;

Represents the weight of the kth neuron in the l-1th layer connected to the jth neuron in the lth layer; b _j ^(l) represents the bias of the jth neuron in the lth layer, used in the fully connected layer The activation function of sigmoid is sigmoid, and the output of the last layer of activation function is the classification result of the user's EEG data segment.

After deep neural network is applied to EEG classification, it has achieved certain results compared with traditional machine learning such as SVM in identifying epileptic seizures and Parkinson's disease. But currently this highly data- and computationally-intensive method can only be used for offline EEG classification. The technical solution adopted in the present invention is a lightweight convolutional neural network applied to a cloud platform, which can implement a response to the classification result of the non-stationary EEG signal. In the specific implementation, the collected EEG signals are put into the intelligent gateway, and with the support of various infrastructures, the mature lightweight CNN classifier trained in the cloud is used to classify the EEG, and the classification results are visualized. Notify users in real time.

In a specific implementation process, step S3.1 may organize the data in the 20*1024 format into a 20*32*32 format, where 20 represents 20 channels of EEG data.

In step S3.3, the output data of the first convolutional layer is 20*30*30, the data is put into the second convolutional layer, and the output data is 18*28*28.

In step S3.4, flatten can be used to expand the 18*28*28 data into a 1*14112 vector. Then use the "hourglass" type of three fully connected layers to reduce the number of neurons. The three fully connected layers of the "hourglass" type refer to: the closer to the output layer, the number of neurons in the fully connected layer decreases in a near-linear manner, like the edge of an hourglass, to rapidly reduce the connection parameters of the neural network, thereby improving training. performance.

The online classification method provided by the present invention will be introduced and described below through specific examples.

Referring to FIG. 2 , the embodiment applies the present invention to classify the brain state of the original EEG of patients with depression. The data set used is MPHC EEG Data, a public data set for depression, with 20 channels and a sampling rate of 256 Hz. With the data window set to 1024 (4 seconds), all sample spaces are 18442 segments, including 9789 segments for depression and 8653 segments for healthy. The cloud server used is Google Colaboratory, the classifier is based on the deep learning framework Keras, and the data collected by the sensing layer is stored in the ongoing database. Convert 20×1024 data into 20×32×32 3D format. After the input data passes through the dropout layer, two convolution layers, and a Maxpooling (pooling) layer, the 3D data blocks in the middle process will be tiled into A vector, the vector is passed to the last three fully connected layers with output sizes of 250, 60 and 1, which finally identify whether the state of the EEG segment is depression or health.

Lightweight CNN is trained with a mini-batch size of 80 using SGD (Stochastic Gradient Descent). The learning rate is set to 0.01, and the dropout is set to 0.1. After the classifier is designed, its performance is evaluated. In this embodiment, the data is divided into training set, validation set and test set, which account for 64%, 16% and 20% respectively. In the training phase, a 5-fold cross-validation algorithm is used, with 10 iterations. All sample fragments in the training set are divided into 5 copies, and in each iteration, 4 copies are used for training and the remaining one is used for validation. Take the average of 10 iteration outputs (test set) as the result. The accuracy of the classification of depression was 98.59% ± 0.28%, the sensitivity was 98.59% ± 0.28%, and the specificity was 99.51% ± 0.19%.

Operational efficiency is measured by training and online classification. The model training is completed within one hour. After the model training is completed, it is downloaded to the gateway to classify the original EEG segment (about 4 seconds in length) that is being input, and it takes less than 0.01 seconds. Improved classification efficiency.

During specific implementation, those skilled in the art can use computer software technology to realize the automatic operation of the above process.

It should be understood that the parts not described in detail in this specification belong to the prior art. The above description of the preferred embodiment is relatively detailed, and therefore should not be considered as a limitation on the scope of protection of the patent of the present invention. Those of ordinary skill in the art, under the inspiration of the present invention, do not deviate from the scope of protection of the claims of the present invention. Below, alternatives or modifications can also be made, which all fall within the protection scope of the present invention, and the claimed protection scope of the present invention shall be subject to the appended claims.

Claims (7)

1. an online classification method based on the EEG data of lightweight convolutional neural network, is characterized in that, is applied to cloud service platform, cloud service platform comprises sensor layer, gateway and cloud, and described online classification method comprises: Step S1: collecting the user's EEG data through the sensing layer; Step S2: The collected EEG data is transmitted to the gateway, and the trained classifier model is downloaded from the cloud through the gateway. After designing the lightweight convolutional neural network structure, the trained classifier model adopts the back-propagation-based method. The mini-batch stochastic gradient descent method is trained; Step S3: online classification of the collected EEG data based on the trained classifier model. 2. The method of claim 1, wherein the structure of designing a lightweight convolutional neural network in step S2 comprises a dropout layer, two convolutional layers with the same convolutional domain, a pooling layer, and three convolutional layers. A fully connected layer, wherein the number of neurons in the three fully connected layers is reduced in a preset manner. 3. method as claimed in claim 2, is characterized in that, the activation function that Dropout layer and first convolution layer use is Relu, relu(x)=max(0,x) Among them, x is a variable; The second convolutional layer, pooling layer, and three fully connected layers all use the Sigmoid activation function. The last activation function of Lightweight, Sigmoid, identifies the classification results of specific EEG segments. The form of the Sigmoid activation function is as follows: where x is a variable. 4. The method of claim 3, further comprising: canceling the activation function of the second convolutional layer. 5. The method of claim 1, wherein in step S2, a back-propagation-based mini-batch stochastic gradient descent method is used for training, comprising: The sample space is divided into training set, validation set and test set in advance according to a preset ratio, wherein, the training set and validation set are stored in the cloud, and the test set is stored in the gateway; The neural network structure is trained through the training set, wherein the weight update rules used in the training process are:

ω _i+1 ←ω _i +ν _i+1 where i is the number of iterations, ν is the momentum term variable, ε is the learning rate,

Represents the average value of the target derivative of the _i -th batch Di when the weight is ω _i , ω _i represents the weight in the i-th iteration process, and ω _i+1 represents the i+1-th iteration. The weight in the process. 6 . The method of claim 5 , further comprising: the cloud is used to incrementally train the classifier model offline by collecting doctor-calibrated EEG data from the gateway. 7 . 7. The method of claim 2, wherein step S3 specifically comprises: Step S3.1: After segmenting the collected EEG data, organize it into a three-dimensional format of channel cascade; Step S3.2: Process the data in the three-dimensional format of the channel cascade through the Dropout layer; Step S3.3: Perform a convolution operation on the data processed in step S3.2 through two convolution layers with the same convolution domain, wherein the convolution kernels of the two convolution layers are both 3×3, and the convolution The operation is expressed as:

In the formula, a _i ^(l) and a _i ^(l-1) represent the ith output channel of the lth layer and the ith output channel of the l-1th layer, respectively;

represents the convolution kernel between the i-th and j-th feature maps in the l-th layer; b _i ^(l) represents the bias term of the i-th feature map in the l-th layer; Step S3.4: perform a pooling operation on the convolved data through a pooling layer; Step S3.5: After the pooled data is expanded into a vector of 1*14112, it is processed through three fully connected layers, wherein the output sizes of the three fully connected layers are 250, 60 and 1, respectively. The join operation is expressed as follows:

a _j ^(l) and a _j ^(l-1) represent the output of the ith neuron in the lth layer and the l-1th layer, respectively;

Represents the weight of the kth neuron in the l-1th layer connected to the jth neuron in the lth layer; b _j ^(l) represents the bias of the jth neuron in the lth layer, used in the fully connected layer The activation function of sigmoid is sigmoid, and the output of the last layer of activation function is the classification result of the user's EEG data segment.

Images