SSVEP (Steady-State Visual Evoked Potential) extraction and classification method, system and device based on electroencephalogram acquisition equipment and medium

By extracting and fusing the SSVEP signals of the EEG acquisition device in wearable EEG devices, using multi-scale time domain and spatial domain feature fusion technology, the problem of insufficient detection and classification accuracy of SSVEP signals under limited channel layout is solved, and the implementation of the implementation of the More efficient signal identification and classification are achieved, and the applicability of the equipment is improved.

CN119089264APending Publication Date: 2024-12-06QUFU NORMAL UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411132608.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-19
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

In the prior art, in wearable EEG devices using limited channel layout, it is difficult to effectively enhance the detection and classification accuracy of SSVEP signals, limiting the applicability of the device.

Method used

By acquiring the steady-state visual evoked potential data of the EEG acquisition device, slice processing is performed to separate the characteristic channel and the non-feature channel data, and extracting the recessive visual stimulation feature map using adaptive pooling layer, convolutional layer and activation function, combined with The parameter weight matrix can be learned for convex combination processing. After recombining the feature map, the spatial domain features are extracted through multi-scale time domain fusion and deep convolution, and finally classified through the fully connected layer.

Benefits of technology

It effectively overcomes the limitations of limited channel layout, efficiently utilizes all electrode channel information, improves the accurate identification and classification of SSVEP targets, and promotes the applicability of portable wearable EEG devices.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119089264A_ABST
    Figure CN119089264A_ABST
Patent Text Reader

Abstract

The invention belongs to the technical field of data acquisition and processing, and particularly relates to an SSVEP (Steady-State Visual Evoked Potential) extraction and classification method, system and device based on electroencephalogram acquisition equipment and a medium. According to electrode distribution of the electroencephalogram acquisition equipment, feature channel data and non-feature channel data are divided, and after recessive visual feature extraction is carried out on the non-feature channel data, the feature channel data is extracted; the method comprises the following steps of: performing multi-scale time-domain feature extraction and fusion on feature channel data, performing reorganization on the feature channel data, performing deep convolution to extract spatial domain features, fusing all features through separable convolution, and outputting a classification result through flattening and a full connection layer, thereby effectively overcoming the actual limitation of wearable electroencephalogram equipment with limited channel layout; all electrode channel information is efficiently utilized, accurate recognition and classification of the SSVEP target of the wearable device are improved, and the applicability of the portable wearable electroencephalogram device is promoted.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of data acquisition and processing, and specifically relates to a SSVEP extraction and classification method, system, device and medium based on an electroencephalogram (EEG) acquisition device. Background Art

[0002] Brain computer interfaces (BCI) technology achieves information transmission and control between the brain and external devices by capturing brain signals (such as EEG signals, neuronal discharge activities, etc.) and converting them into electrical signals or control instructions. Among various BCI paradigms, steady-state visual evoked potential (SSVEP), as a natural response signal of the brain to visual stimulation of a constant frequency (usually greater than 4Hz), stands out in BCI systems due to its robustness and easy detection.

[0003] In practical applications, EEG signal acquisition devices are often constrained by various factors. Traditional EEG systems usually consist of many electrodes and extensive wiring, which makes their setup and use complicated and cumbersome. Instead of deploying a sufficient number of electrodes in any single area, portable wearable EEG acquisition devices distribute a limited number of electrodes across different scalp areas to facilitate the practical use of brain-computer interfaces in daily environments.

[0004] At present, the main methods for enhancing the detection and classification of SSVEP signals include: canonical correlation analysis (CCA), task-related component analysis (TRCA), multivariate synchronization index (MSI) and filter bank canonical correlation analysis (FBCCA) in machine learning, as well as methods for adding deep learning networks such as CNN, EEGNet, and SSVEPNet. However, the above methods usually require a large number of EEG channels to achieve high accuracy, which limits their applicability in wearable devices with fewer channels. Even theoretical studies rarely utilize all electrode channels on objective devices.

[0005] In summary, in real-life scenarios where wearable EEG devices with limited channel layouts are used, further research is needed to enhance the detection of SSVEP signals and improve classification accuracy, thereby promoting the applicability of wearable EEG devices. Summary of the invention

[0006] In order to solve the problems raised by the background technology, the present invention provides a SSVEP extraction and classification method, system, device and medium based on an electroencephalogram (EEG) acquisition device.

[0007] The technical solution of the present invention is as follows:

[0008] The present invention provides a SSVEP extraction and classification method based on an electroencephalogram (EEG) acquisition device, comprising the following steps:

[0009] S1: Obtain the steady-state visual evoked potential data of the EEG acquisition device, and obtain the characteristic channel data and non-characteristic channel data through slicing processing;

[0010] S2: The non-feature channel data is processed by the adaptive pooling layer, the first linear layer, the first activation function, the second linear layer, and the second activation function in sequence to obtain a first implicit visual stimulus feature map;

[0011] The non-feature channel data is processed by the convolution layer and the pooling layer in sequence to obtain the second implicit visual stimulus feature map;

[0012] According to the learnable parameter weight matrix, the first implicit visual stimulus feature map and the second implicit visual stimulus feature map are subjected to convex combination correlation processing to obtain an implicit visual stimulus feature map;

[0013] S3: The implicit visual stimulus feature map is recombined with the explicit visual stimulus feature map in the feature channel data to obtain a recombined feature map;

[0014] S4: The reconstructed feature map is processed through three parallel convolutional layers, specifically:

[0015] After the reorganized feature map is processed by the first convolutional layer, a first feature map is obtained;

[0016] After the reorganized feature map is processed by the second convolutional layer, a second feature map is obtained;

[0017] After the reorganized feature map is processed by the third convolutional layer, a third feature map is obtained;

[0018] After the first feature map, the second feature map and the third feature map are concatenated, they are processed by point convolution, batch normalization and the third activation function in sequence to obtain a multi-scale time-domain fusion feature map;

[0019] S5: The multi-scale time-domain fusion feature map is processed by deep convolution and separable convolution in turn to obtain the spatial-domain fusion feature map;

[0020] S6: The spatial domain fusion feature map is processed by flattening and fully connected layers in turn, and then classified to obtain the classification result.

[0021] Step S2, using the attention function to obtain the implicit visual stimulus feature map, the specific operation is as follows:

[0022] SE(x)=(σ(σ(W2·ReL(W1·AvP(x)+b1)+b2)·σ(W weight )+AvP(Conv(x))·(1-σ(W weight ))))x

[0023] Where SE(x) is the attention function, x is the non-feature channel data, σ is the second activation function, ReL is the first activation function, W1 and W2 are the learnable parameter matrices of the first linear layer and the second linear layer, b1 and b2 are the biases of the first linear layer and the second linear layer, AvP is the pooling function, Conv is the convolution function, and W weight is the learnable parameter weight matrix.

[0024] In step S4, the first feature map, the second feature map and the third feature map are concatenated by a Concat layer, specifically:

[0025] According to the formula: x4 = Concat (x1, x2, x3), concatenate;

[0026] In the formula, x1, x2, x3 are the first feature map, the second feature map, and the third feature map respectively, and x4 is the concatenated feature map.

[0027] In step S4, point convolution, batch normalization, and the third activation function are sequentially processed to obtain a multi-scale time-domain fusion feature map, specifically:

[0028] According to the formula: x5 = Dropout (ELU (BN (Conv2D (1,1) (x4)))), obtain the multi-scale time domain fusion feature map;

[0029] In the formula, x5 is the multi-scale time-domain fusion feature map, x4 is the concatenated feature map, and Conv2D (1,1) is the point convolution function, BN is the batch normalization function, ELU is the third activation function, and Dropout is used to train the neural network.

[0030] In step S4, the first convolution layer, the second convolution layer, and the third convolution layer are two-dimensional convolution layers with convolution kernels of different sizes.

[0031] In step S2, the first activation function is the ReLU activation function, and the second activation function is the Sigmoid activation function; in step S4, the third activation function is the ELU activation function.

[0032] The present invention also provides a SSVEP extraction and classification system based on an electroencephalogram acquisition device, comprising:

[0033] Data partitioning module: used to obtain the steady-state visual evoked potential data of the EEG acquisition device, and obtain the characteristic channel data and non-characteristic channel data through slicing processing;

[0034] Implicit visual stimulus feature extraction module: After the non-feature channel data is processed by the adaptive pooling layer, the first linear layer, the first activation function, the second linear layer, and the second activation function in sequence, the first implicit visual stimulus feature map is obtained;

[0035] The non-feature channel data is processed by the convolution layer and the pooling layer in sequence to obtain the second implicit visual stimulus feature map;

[0036] According to the learnable parameter weight matrix, the first implicit visual stimulus feature map and the second implicit visual stimulus feature map are subjected to convex combination correlation processing to obtain an implicit visual stimulus feature map;

[0037] Feature recombination module: used to recombine the implicit visual stimulus feature map with the explicit visual stimulus feature map in the feature channel data to obtain a recombined feature map;

[0038] Multi-scale temporal feature fusion module: The reconstructed feature map is processed through three parallel convolutional layers, specifically:

[0039] After the reorganized feature map is processed by the first convolutional layer, a first feature map is obtained;

[0040] After the reorganized feature map is processed by the second convolutional layer, a second feature map is obtained;

[0041] After the reorganized feature map is processed by the third convolutional layer, a third feature map is obtained;

[0042] After the first feature map, the second feature map and the third feature map are concatenated, they are processed by point convolution, batch normalization and the third activation function in sequence to obtain a multi-scale time-domain fusion feature map;

[0043] Spatial domain feature fusion module: The multi-scale time domain fusion feature map is processed by deep convolution and separable convolution in turn to obtain the spatial domain fusion feature map;

[0044] Classification module: The spatial domain fusion feature map is processed by flattening and fully connected layers in sequence, and then classified to obtain the classification result.

[0045] The present invention also provides an SSVEP extraction and classification device based on an electroencephalogram (EEG) acquisition device, comprising a processor and a memory, wherein the processor implements the SSVEP extraction and classification method based on the EEG acquisition device when executing a computer program stored in the memory.

[0046] The present invention also provides a SSVEP extraction and classification medium based on an electroencephalogram (EEG) acquisition device, which is used to store a computer program, wherein when the computer program is executed by a processor, the SSVEP extraction and classification method based on the EEG acquisition device is implemented.

[0047] Beneficial Effects

[0048] The present invention divides the electrode distribution of the EEG acquisition device into feature channel data and non-feature channel data, extracts implicit visual features from the non-feature channel data, and then recombines them with the feature channel data. The recombined features are extracted and fused through multi-scale time domain features, and then the spatial domain features are extracted through deep convolution. All features are fused through separable convolution, and the classification results are output through flattening and fully connected layers. This method effectively overcomes the practical limitations of wearable EEG devices with limited channel layouts, efficiently utilizes all electrode channel information, improves the accurate recognition and classification of SSVEP targets of wearable devices, and promotes the applicability of portable wearable EEG devices.

[0049] The data division performed in the present invention not only ensures that the electrode data with significant visual features can be processed efficiently, but also fully utilizes all channel data, thereby ensuring the integrity of the SSVEP features.

[0050] By performing parallel processing on non-feature channel data, the present invention can not only extract the SSVEP features implicit in the non-feature channel data, but also increase the attention paid to certain obvious non-feature channels, thereby effectively utilizing the data of the non-feature channels under the restriction of a limited number of channels.

[0051] The present invention uses multiple convolution kernels in parallel to reorganize features to extract time domain features on different time scales, generate features on multiple time scales, and fuse multiple groups of features by a point convolution layer, and fuse multiple groups of features into one group of features, which is convenient for subsequent feature extraction. Feature extraction is performed with the dimension of one group of features, which optimizes the calculation process and can effectively improve the computing efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 Schematic diagram of a flow chart of a SSVEP extraction and classification method based on an EEG acquisition device in an embodiment. DETAILED DESCRIPTION

[0053] The following examples are intended to illustrate the present invention rather than to further limit the present invention.

[0054] The present invention provides a SSVEP extraction and classification method based on EEG acquisition equipment. The method is applicable to wearable brain-computer interface data acquisition equipment. Such equipment cannot be equipped with all the electrode channels of the international 10-20 or 10-10 standards due to the pursuit of portability and comfort. The method aims to make full use of limited electrode channel data to extract features of the SSVEP paradigm, thereby realizing the classification function.

[0055] like Figure 1 As shown, the following steps are included:

[0056] S1: Obtain the steady-state visual evoked potential (SSVEP) data from the EEG acquisition device, and obtain the characteristic channel data and non-characteristic channel data through slicing processing.

[0057] For portable wearable brain-computer interface data acquisition devices, under the premise of a limited number of electrode configurations, the characteristics of the SSVEP paradigm mainly appear in the occipital lobe area. The characteristic channels (occipital lobe electrodes) and non-characteristic channels are selected according to the actual electrode channel configuration.

[0058] After obtaining the visual stimulation data at different frequencies, the data is used as the original data. Assume that the original data is a multidimensional time series matrix. According to the electrode distribution of the wearable EEG device, the electrode channels in the occipital region with obvious visual features and other non-feature electrode channels are sliced ​​and divided to obtain feature channel data and non-feature channel data. These two parts of data remain unchanged with the original data in other dimensions, and in the channel dimension, the sum of the channel dimensions of the two parts of data is the size of the original data dimension.

[0059] The data division in step S1 of the present invention not only ensures that the electrode data with significant visual features can be processed efficiently, but also fully utilizes all channel data, thereby ensuring the integrity of the SSVEP features.

[0060] S2: The non-feature channel data is processed by the adaptive pooling layer, the first linear layer, the first activation function, the second linear layer, and the second activation function in sequence to obtain a first implicit visual stimulus feature map;

[0061] The non-feature channel data is processed by the convolution layer and the pooling layer in sequence to obtain the second implicit visual stimulus feature map;

[0062] According to the learnable parameter weight matrix, the first implicit visual stimulation feature map and the second implicit visual stimulation feature map are processed by convex combination correlation to obtain an implicit visual stimulation feature map.

[0063] In addition to feature channel data, feature extraction of non-feature channel data helps to improve the comprehensiveness of feature extraction and the accuracy of classification.

[0064] Preferably, in step S2, the first activation function is a ReLU activation function, and the second activation function is a Sigmoid activation function.

[0065] The present invention can extract the SSVEP features implicit in the non-feature channel data and increase the attention paid to certain obvious non-feature channels by performing parallel processing on the non-feature channel data in step S2, thereby effectively utilizing the data of the non-feature channels under the restriction of a limited number of channels.

[0066] Preferably, in step S2, the implicit visual stimulus feature map is obtained using an attention function, and the specific operation is as follows:

[0067] SE(x)=(σ(σ(W2·ReL(W1·AvP(x)+b1)+b2)·σ(W weight )+AvP(conv(x))·(1-σ(W weight ))))x

[0068] Where SE(x) is the attention function, x is the non-feature channel data, σ is the Sigmoid activation function, ReL is the ReLU activation function, W1 and W2 are the learnable parameter matrices of the first linear layer and the second linear layer, b1 and b2 are the biases of the first linear layer and the second linear layer, AvP is the pooling function, Conv is the convolution function, and W weight is the learnable parameter weight matrix.

[0069] The attention function SE(x) enhances the features of important channels and suppresses the features of unimportant channels by weighting the features through the channel attention mechanism, thereby improving the representation ability of the features and the classification accuracy.

[0070] Specifically, the step-by-step operation is as follows:

[0071] y1=AvP(x)

[0072] z1=W1·y1+b1

[0073] a1=ReLU(z1)

[0074] z2=W2·a1+b2

[0075] a2=σ(z2)

[0076] y2=Conv(x)

[0077] w=σ(W weight )

[0078] g=a2·w+y2·(1-w)

[0079] SE(x)=σ(g)·x

[0080] The specific process is as follows: non-feature channel data x is processed in parallel. First, non-feature channel data x is first subjected to a one-dimensional adaptive pooling to obtain y1. Then y1 is compressed by a first linear layer to obtain z1. z1 is activated by a ReLU activation function to obtain a1. a1 is expanded to its pre-compression size by a second linear layer to obtain z2. z2 is processed by a Sigmoid activation function to limit the feature map between (0, 1) to obtain the first implicit visual stimulus feature map a2.

[0081] Second, the non-feature channel data x is processed by the convolution layer and the pooling layer in sequence to obtain the second implicit visual stimulus feature map y2;

[0082] Introducing a learnable parameter weight matrix W weight , W weight The size of W is the same as the number of non-feature channels in dimension. weight After activation by the Sigmoid activation function, w is obtained. The first implicit visual stimulus feature map a2 and the second implicit visual stimulus feature map y2 are combined with w to perform convex combination correlation to obtain the attention mechanism g. After g is activated by the Sigmoid activation function, it is multiplied with the non-feature channel data x to obtain the attention function SE(x). Among them, the convolution step size parameter is set to 'same', so this process does not change the feature output size.

[0083] S3: The implicit visual stimulus feature map is recombined with the explicit visual stimulus feature map in the feature channel data to obtain a recombined feature map. The dimension size of the recombined feature is the same as the original data size.

[0084] S4: The reconstructed feature map is processed through three parallel convolutional layers, specifically:

[0085] After the reorganized feature map is processed by the first convolutional layer, a first feature map is obtained;

[0086] After the reorganized feature map is processed by the second convolutional layer, a second feature map is obtained;

[0087] After the reorganized feature map is processed by the third convolutional layer, a third feature map is obtained;

[0088] After the first feature map, the second feature map and the third feature map are spliced, they are processed by point convolution, batch normalization and the third activation function in sequence to obtain a multi-scale time-domain fusion feature map.

[0089] When extracting multi-scale time domain features, three parallel convolution kernels are included to extract features in large, medium, and small scale ranges. Among them, the large-scale convolution kernel is used to extract the global features of the data, the medium-scale convolution kernel is used to extract features within the medium time domain range, and the smallest-scale convolution kernel is used to extract detailed features in the time domain space. The three groups of convolution features are fused through point convolution to output a group of features, which optimizes the calculation process and can effectively improve the calculation efficiency.

[0090] Furthermore, the third activation function in step S4 is an ELU activation function.

[0091] When extracting multi-scale time domain features, preferably, in step S4, the first convolution layer, the second convolution layer, and the third convolution layer are two-dimensional convolution layers with convolution kernels of different sizes.

[0092] Specifically, first, the reorganized feature map is input into three networks with different sizes of convolution kernels. After convolution, batch normalization is performed to obtain three sets of n-dimensional feature maps. Point convolution will fuse the three sets of feature maps in the feature map dimension, and finally output the feature map dimension as n. Among them, the output of three parallel multi-scale convolutions is as follows:

[0093] x1=BN(Conv2d (1,4) (x0))

[0094] x2=BN(Conv2d (1,64) (x0))

[0095] x3=BN(Conv2d (1,128) (x0))

[0096] In the formula, x1, x2, and x3 are the output feature maps of the three convolutional layers, namely the first feature map, the second feature map, and the third feature map; x0 is the reorganized feature map; Conv2D (1,n) represents a two-dimensional convolution layer with a convolution kernel size of (1, n), and BN represents the batch normalization function.

[0097] Preferably, in step S4, the first feature map, the second feature map and the third feature map are spliced, preferably, they are spliced ​​through a Concat layer, and after splicing, they are sequentially processed by point convolution, batch normalization, and ELU activation function to obtain a multi-scale time domain fusion feature map, specifically:

[0098] According to the formula: x4 = Concat (x1, x2, x3), concatenate;

[0099] In the formula, x1, x2, x3 are the first feature map, the second feature map, and the third feature map respectively, and x4 is the concatenated feature map;

[0100] According to the formula: x5 = Dropout (ELU (BN (Conv2D (1,1) (x4)))), obtain the multi-scale time domain fusion feature map;

[0101] Where x5 is the multi-scale time-domain fusion feature map, Conv2D (1,1) is the point convolution function, BN is the batch normalization function, ELU is the ELU activation function, and Dropout is used to train the neural network.

[0102] The present invention uses multiple convolution kernels in parallel to reorganize features to extract time domain features on different time scales, generate features on multiple time scales, and fuse multiple groups of features by a point convolution layer, and fuse multiple groups of features into one group of features, so as to facilitate feature extraction in the dimension of one group of features during subsequent feature extraction.

[0103] S5: The multi-scale time-domain fusion feature map is processed by deep convolution and separable convolution in sequence to obtain the spatial-domain fusion feature map.

[0104] The multi-scale time-domain fusion feature map is used to extract data spatial features through a deep convolution layer, and then all feature map information is fused through a separable convolution layer to improve the feature representation capability and ensure the comprehensiveness and accuracy of the spatial domain fusion feature map.

[0105] S6: The spatial domain fusion feature map is processed by flattening and fully connected layers in sequence, and then classified, such as using the Softmax function to obtain the classification result.

[0106] The spatial domain fusion feature map is flattened by flattening, mapped to the output space of the classification task through a fully connected layer, and the Softmax function is used to generate the probability distribution of each category (visual stimulus signals at different frequencies) to achieve accurate classification of the SSVEP target.

[0107] The present invention divides the electrode distribution of the EEG acquisition device into feature channel data and non-feature channel data, extracts implicit visual features from the non-feature channel data, and then recombines them with the feature channel data. The recombined features are subjected to multi-scale time domain feature extraction and fusion, and then the spatial domain features are extracted through deep convolution. All features are fused through separable convolution, and the classification results are output through flattening and fully connected layers. This method effectively overcomes the practical limitations of wearable EEG devices with limited channel layouts, efficiently utilizes all electrode channel information, and improves the accurate recognition and classification of SSVEP targets of wearable devices.

[0108] The present invention also provides a SSVEP extraction and classification system based on an electroencephalogram acquisition device, comprising:

[0109] Data partitioning module: used to obtain the steady-state visual evoked potential data of the EEG acquisition device, and obtain the characteristic channel data and non-characteristic channel data through slicing processing;

[0110] Implicit visual stimulus feature extraction module: After the non-feature channel data is processed by the adaptive pooling layer, the first linear layer, the first activation function, the second linear layer, and the second activation function in sequence, the first implicit visual stimulus feature map is obtained;

[0111] The non-feature channel data is processed by the convolution layer and the pooling layer in sequence to obtain the second implicit visual stimulus feature map;

[0112] According to the learnable parameter weight matrix, the first implicit visual stimulus feature map and the second implicit visual stimulus feature map are subjected to convex combination correlation processing to obtain an implicit visual stimulus feature map;

[0113] Feature recombination module: used to recombine the implicit visual stimulus feature map with the explicit visual stimulus feature map in the feature channel data to obtain a recombined feature map;

[0114] Multi-scale temporal feature fusion module: The reconstructed feature map is processed through three parallel convolutional layers, specifically:

[0115] After the reorganized feature map is processed by the first convolutional layer, a first feature map is obtained;

[0116] After the reorganized feature map is processed by the second convolutional layer, a second feature map is obtained;

[0117] After the reorganized feature map is processed by the third convolutional layer, a third feature map is obtained;

[0118] After the first feature map, the second feature map and the third feature map are concatenated, they are processed by point convolution, batch normalization and the third activation function in sequence to obtain a multi-scale time-domain fusion feature map;

[0119] Spatial domain feature fusion module: The multi-scale time domain fusion feature map is processed by deep convolution and separable convolution in turn to obtain the spatial domain fusion feature map;

[0120] Classification module: The spatial domain fusion feature map is processed by flattening and fully connected layers in sequence, and then classified to obtain the classification result.

[0121] The present invention also provides an SSVEP extraction and classification device based on an electroencephalogram (EEG) acquisition device, comprising a processor and a memory, wherein the processor implements the SSVEP extraction and classification method based on the EEG acquisition device when executing a computer program stored in the memory.

[0122] The present invention also provides a SSVEP extraction and classification medium based on an electroencephalogram (EEG) acquisition device, which is used to store a computer program, wherein the computer program, when executed by a processor, implements the SSVEP extraction and classification method based on an EEG acquisition device.

Claims

1. A SSVEP extraction and classification method based on EEG acquisition equipment, characterized in that: The following steps are involved: S1: Obtain the steady-state visual evoked potential data of the EEG acquisition device, and obtain the characteristic channel data and non-characteristic channel data through slicing processing; S2: The non-feature channel data is processed by the adaptive pooling layer, the first linear layer, the first activation function, the second linear layer, and the second activation function in sequence to obtain a first implicit visual stimulus feature map; The non-feature channel data is processed by the convolution layer and the pooling layer in sequence to obtain the second implicit visual stimulus feature map; According to the learnable parameter weight matrix, the first implicit visual stimulus feature map and the second implicit visual stimulus feature map are subjected to convex combination correlation processing to obtain an implicit visual stimulus feature map; S3: The implicit visual stimulus feature map is recombined with the explicit visual stimulus feature map in the feature channel data to obtain a recombined feature map; S4: The reconstructed feature map is processed through three parallel convolutional layers, specifically: After the reorganized feature map is processed by the first convolutional layer, a first feature map is obtained; After the reorganized feature map is processed by the second convolutional layer, a second feature map is obtained; After the reorganized feature map is processed by the third convolutional layer, a third feature map is obtained; After the first feature map, the second feature map and the third feature map are concatenated, they are processed by point convolution, batch normalization and the third activation function in sequence to obtain a multi-scale time-domain fusion feature map; S5: The multi-scale time-domain fusion feature map is processed by deep convolution and separable convolution in turn to obtain the spatial-domain fusion feature map; S6: The spatial domain fusion feature map is processed by flattening and fully connected layers in turn, and then classified to obtain the classification result.

2. The SSVEP extraction and classification method based on the EEG acquisition device according to claim 1 is characterized in that: Step S2, using the attention function to obtain the implicit visual stimulus feature map, the specific operation is as follows: SE(x)=(σ(σ(W2·ReL(W1·AvP(x)+b1)+b2)·σ(W weight )+AvP(Conv(x))·(1-σ(W weight ))))x Where SE(x) is the attention function, x is the non-feature channel data, σ is the second activation function, ReL is the first activation function, W1 and W2 are the learnable parameter matrices of the first linear layer and the second linear layer, b1 and b2 are the biases of the first linear layer and the second linear layer, AvP is the pooling function, Conv is the convolution function, and W weight is the learnable parameter weight matrix.

3. The SSVEP extraction and classification method based on the EEG acquisition device according to claim 1 is characterized in that: In step S4, the first feature map, the second feature map and the third feature map are concatenated by a Concat layer, specifically: According to the formula: x4 = Concat (x1, x2, x3), concatenate; In the formula, x1, x2, x3 are the first feature map, the second feature map, and the third feature map respectively, and x4 is the concatenated feature map.

4. The SSVEP extraction and classification method based on the EEG acquisition device according to claim 1 is characterized in that: In step S4, point convolution, batch normalization, and the third activation function are sequentially processed to obtain a multi-scale time-domain fusion feature map, specifically: According to the formula: x5 = Dropout (ELU (BN (Conv2D (1,1) (x4)))), obtain the multi-scale time domain fusion feature map; In the formula, x5 is the multi-scale time-domain fusion feature map, x4 is the concatenated feature map, and Conv2D (1,1) is the point convolution function, BN is the batch normalization function, ELU is the third activation function, and Dropout is used to train the neural network.

5. The SSVEP extraction and classification method based on the EEG acquisition device according to claim 1 is characterized in that: In step S4, the first convolution layer, the second convolution layer, and the third convolution layer are two-dimensional convolution layers with convolution kernels of different sizes.

6. The SSVEP extraction and classification method based on the EEG acquisition device according to claim 1 is characterized in that: In step S2, the first activation function is the ReLU activation function, and the second activation function is the Sigmoid activation function; in step S4, the third activation function is the ELU activation function.

7. A SSVEP extraction and classification system based on EEG acquisition equipment, characterized in that: include: Data partitioning module: used to obtain the steady-state visual evoked potential data of the EEG acquisition device, and obtain the characteristic channel data and non-characteristic channel data through slicing processing; Implicit visual stimulus feature extraction module: After the non-feature channel data is processed by the adaptive pooling layer, the first linear layer, the first activation function, the second linear layer, and the second activation function in sequence, the first implicit visual stimulus feature map is obtained; The non-feature channel data is processed by the convolution layer and the pooling layer in sequence to obtain the second implicit visual stimulus feature map; According to the learnable parameter weight matrix, the first implicit visual stimulus feature map and the second implicit visual stimulus feature map are subjected to convex combination correlation processing to obtain an implicit visual stimulus feature map; Feature recombination module: used to recombine the implicit visual stimulus feature map with the explicit visual stimulus feature map in the feature channel data to obtain a recombined feature map; Multi-scale temporal feature fusion module: The reconstructed feature map is processed through three parallel convolutional layers, specifically: After the reorganized feature map is processed by the first convolutional layer, a first feature map is obtained; After the reorganized feature map is processed by the second convolutional layer, a second feature map is obtained; After the reorganized feature map is processed by the third convolutional layer, a third feature map is obtained; After the first feature map, the second feature map and the third feature map are concatenated, they are processed by point convolution, batch normalization and the third activation function in sequence to obtain a multi-scale time-domain fusion feature map; Spatial domain feature fusion module: The multi-scale time domain fusion feature map is processed by deep convolution and separable convolution in turn to obtain the spatial domain fusion feature map; Classification module: The spatial domain fusion feature map is processed by flattening and fully connected layers in sequence, and then classified to obtain the classification result.

8. A SSVEP extraction and classification device based on an electroencephalogram acquisition device, characterized in that: The method comprises a processor and a memory, wherein the processor implements the SSVEP extraction and classification method based on the EEG acquisition device as described in any one of claims 1 to 6 when executing the computer program stored in the memory.

9. A SSVEP extraction and classification medium based on an electroencephalogram acquisition device, characterized in that: Used to store a computer program, wherein when the computer program is executed by a processor, the SSVEP extraction and classification method based on an electroencephalogram acquisition device as described in any one of claims 1 to 6 is implemented.