Data-driven electroencephalogram channel expansion method and system

Abstract

The invention discloses a data-driven electroencephalogram channel expansion method and system, and the method comprises the steps: obtaining a low-channel electroencephalogram signal, dividing the low-channel electroencephalogram signal along a time dimension, integrating a plurality of partitioned signals obtained through division along a channel dimension, forming a partitioned signal matrix, carrying out the feature extraction and fusion processing of each partitioned signal in the partitioned signal matrix, and carrying out the feature extraction and fusion processing of each partitioned signal in the partitioned signal matrix; performing channel expansion on the fusion feature matrix to generate an expanded feature matrix, splicing the expanded feature matrix and the fusion feature matrix, and mapping the feature matrix obtained after splicing back to a time domain signal through feedforward operation to obtain a high-channel electroencephalogram signal. According to the method, the state space model is applied to the channel expansion task of the EEG signal, compared with an existing method, the method has higher time sequence modeling capacity and feature extraction capacity, and the spatial-temporal features of the EEG signal can be better captured.

Description

Technical Field

[0001] The present invention relates to the field of brain-computer interface technology, and in particular to a data-driven electroencephalogram channel expansion method and system. Background Art

[0002] Electroencephalography (EEG) is a non-invasive electrophysiological monitoring technology used to measure the electrical activity of the brain. However, due to the low electrode density and volume conduction effect, EEG often suffers from limited spatial resolution: high-channel EEG systems (generally equipped with 100 to 1000 electrodes) can improve the channel, but these systems are expensive, complex to configure, and may cause user discomfort during long-term monitoring; in contrast, low-channel EEG systems (generally equipped with 8 or 16 electrodes) are more cost-effective, user-friendly and practical. Traditional solutions to improve EEG spatial resolution mainly rely on hardware, such as increasing the number of electrodes, which not only greatly increases the hardware cost, but also causes problems such as discomfort in wearing the equipment and prolonged preparation time. Therefore, improving the spatial resolution of EEG from a software perspective has important application value. Existing EEG channel expansion methods mainly include feature engineering methods and deep learning-based methods: feature engineering methods mainly rely on mathematical models such as interpolation and sparse representation, and estimate missing EEG channels by interpolating adjacent channels or sparse coding based on predefined dictionaries. The main disadvantage of this type of method is that it cannot effectively capture the complex nonlinear relationship between EEG signal channels, resulting in unsatisfactory reconstruction quality; although deep learning-based methods have improved performance, they often require a large amount of training data and computing resources, and existing methods generally have problems such as insufficient time series modeling capabilities and insufficient feature extraction. Summary of the invention

[0003] In response to the above problems, the present invention application proposes a data-driven EEG channel expansion method and system, which can accurately reconstruct multi-channel high-precision EEG signals based on low-channel EEG signals collected by a small number of electrodes, achieving the goal of obtaining high-channel EEG signals while keeping the hardware simple.

[0004] On the one hand, the present invention provides a data-driven electroencephalogram channel expansion method, which comprises the following steps:

[0005] S1. Obtain a low-channel EEG signal, divide the low-channel EEG signal along the time dimension to obtain a number of block signals, and integrate the several block signals along the channel dimension to form a block signal matrix;

[0006] S2, using the state space model to perform feature extraction and fusion processing on the block signal matrix to obtain a fusion feature matrix;

[0007] S3, performing statistical analysis on the fusion feature matrix, performing channel expansion on the statistical features obtained after the statistical analysis, and generating an expanded feature matrix;

[0008] S4. Concatenate the extended feature matrix and the fused feature matrix, and map the concatenated feature matrix back to the time domain signal through a feedforward operation to obtain a high-channel EEG signal.

[0009] Preferably, the block signal matrix in S1 is specifically expressed as:

[0010] X patch (nc,t)=X l (c,nT w +t), 0≤t<T w

[0011] Where n∈[0,N-1], c∈[0,C l -1],t∈[0,T w -1]

[0012] Where, X patch represents the block signal matrix, X l Represents low-channel EEG signals, C l represents the number of channels of the low-channel EEG signal, n represents the index of the block, c represents the channel index, t represents the position index in the time window, T w Represents the length of each time window, T represents the total time length of the low-channel EEG signal, and N represents the number of blocks.

[0013] Preferably, the state space model includes a feature embedding module, a feature enhancement module, a selective scanning module and a feature fusion module connected in sequence. The specific process of S2 is as follows:

[0014] S21, the feature embedding module maps each block signal in the block signal matrix to the feature space through a learnable embedding weight matrix to obtain an embedded feature matrix;

[0015] S22, the feature enhancement module performs average pooling of the channel dimension on the embedded feature matrix and learns the attention weights, applies the attention weights to the embedded feature matrix, realizes the selective enhancement of the embedded features, and obtains the enhanced feature matrix;

[0016] S23, the selective scanning module performs selective scanning operations on the enhanced feature matrix in a plurality of preset different directions to obtain scanning feature matrices in a plurality of different directions;

[0017] S24, the feature fusion module receives several scanning feature matrices in different directions and performs weighted fusion and layer normalization processing to obtain a fused feature matrix.

[0018] Preferably, the embedding feature matrix in S21 is specifically expressed as:

[0019] X embed =W embed X patch +b embed

[0020] Among them, X embed represents the embedding feature matrix, W embed represents the embedding weight matrix, b embed represents the bias vector, L represents the embedding feature dimension.

[0021] Preferably, the enhanced feature matrix in S22 is specifically expressed as:

[0022] X enhanced =X embed ⊙A ct

[0023] Among them, A ct =σ(W2·ReLU(W1·Avg(X embed )+b1)+b2)

[0024] In the formula, X enhanced represents the enhanced feature matrix, ⊙ represents the element-by-element multiplication operation, A ct represents the attention weight, W1 and W2 both represent weight matrices, b1 and b2 represent bias vectors, σ represents the sigmoid activation function, and Avg represents the l Averaging operation over dimensions.

[0025] Preferably, the several different directions preset in S23 include: horizontally from left to right, horizontally from right to left, vertically from top to bottom, and vertically from bottom to top.

[0026] Preferably, the scanning feature matrices in several different directions in S23 are specifically expressed as:

[0027] F d =S d (X enhanced )=W d Scan d (X enhanced )+b d , d∈{htb,hbt,vtb,vbt}

[0028] In the formula, F d represents the scanning feature matrix in direction d, S d (·) represents the scanning operation in direction d, X enhanced represents the enhanced feature matrix, Scan d (·) represents the reordering operation in direction d, W d and b d They represent transformation weight and bias respectively, htb represents horizontal from left to right, hbt represents horizontal from right to left, vtb represents vertical from top to bottom, and vbt represents vertical from bottom to top.

[0029] Preferably, the extended features in S3 are specifically expressed as:

[0030] F new =Expand(Mean(F merged ), C h -C l )

[0031] In the formula, represents the extended feature matrix, F merged represents the fusion feature matrix, Mean represents the average operation in the channel dimension, Expand represents the feature copy expansion operation, C h Indicates the number of channels of high-channel EEG signals, C l Indicates the number of channels of low-channel EEG signals, C h >C l .

[0032] Preferably, the high-channel EEG signal in S4 is specifically expressed as:

[0033] X h =W out ·F concat +b out

[0034] Among them, F concat =Concat(F merged , F new )

[0035] In the formula, represents the concatenated feature matrix, W out and b out denote the output weight and output bias respectively, X h represents the high-channel EEG signal,

[0036] On the other hand, the present invention provides an electroencephalogram channel expansion system, which uses the above-mentioned data-driven electroencephalogram channel expansion method to expand the electroencephalogram channel. The system includes a computer system and an electroencephalogram signal acquisition module connected to the computer system. The computer system is provided with a signal preprocessing module, a state space module, a channel expansion module and a signal reconstruction module connected in sequence. The state space module is also connected to the signal reconstruction module, wherein:

[0037] The EEG signal acquisition module is used to obtain low-channel EEG signals and transmit the low-channel EEG signals to the computer system;

[0038] The signal preprocessing module is used to process the low-channel EEG signal in blocks to obtain a number of block signals, and arrange the block signals into a block signal matrix;

[0039] The state space model is used to receive the block signal matrix and perform feature extraction and feature fusion to obtain a fused feature matrix;

[0040] The channel expansion module is used to receive the fused feature matrix and perform channel expansion to obtain an extended feature matrix;

[0041] The signal reconstruction module is used to receive the fused feature matrix and the extended feature matrix, and perform splicing and signal reconstruction on them to obtain a high-channel EEG signal.

[0042] The above-mentioned data-driven EEG channel expansion method and system obtains low-channel EEG signals and divides them along the time dimension, integrates several block signals obtained by division along the channel dimension to form a block signal matrix, extracts and fuses each block signal in the block signal matrix to obtain a fused feature matrix, performs channel expansion on the fused feature matrix to generate an extended feature matrix, splices the extended feature matrix and the fused feature matrix, maps the spliced ​​feature matrix back to the time domain signal through a feedforward operation to obtain a high-channel EEG signal. This method can accurately reconstruct multi-channel high-precision EEG signals based on low-channel EEG signals collected by a small number of electrodes, achieving the goal of obtaining high-quality EEG signals while keeping the hardware simple, and has stronger time series modeling and feature extraction capabilities than existing methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a flow chart of a data-driven EEG channel expansion method in one embodiment of the present invention. DETAILED DESCRIPTION

[0044] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings.

[0045] See also Figure 1, a data-driven electroencephalogram channel expansion method, the method comprising:

[0046] S1. Obtain a low-channel EEG signal, divide the low-channel EEG signal along a time dimension to obtain a plurality of block signals, and integrate the plurality of block signals along a time dimension to form a block signal matrix.

[0047] Specifically, since EEG signals have continuity and timing characteristics, it is necessary to obtain low-channel EEG signals containing information in two dimensions, channel and time. The low-channel EEG signal Divide the signal into multiple time windows along the time dimension for processing. Specifically, divide the low-channel EEG signal into several equal-sized block signals along the time dimension, and each block signal contains a time series of fixed length. This block processing method helps to capture local time features and facilitates subsequent parallel computing. The selection of block size requires a balance between computational efficiency and the adequacy of feature extraction. Too small block signals may lose important timing information, and too large block signals will increase the computational burden. Therefore, choosing a window length that is suitable for the main rhythm cycle of the EEG signal (such as 0.5 seconds to 1 second) can achieve better results.

[0048] It should be noted that when the last time window length is less than T w When , all time windows are guaranteed to have the same length by zero padding to ensure the consistency of subsequent processing. The block-based signal can be rearranged along the channel dimension to form a block signal matrix. This facilitates subsequent feature extraction operations. The size of a block signal is C l ×T w , the size of the block signal matrix composed of N blocks is (N·C l )×T w .

[0049] Furthermore, the block signal matrix in S1 is specifically expressed as:

[0050] X patch (nc, t) = X l (c,nT w +t), 0≤t<T w

[0051] Where n∈[0,N-1], c∈[0,C l -1],t∈[0,T w -1]

[0052] In the formula, X patch represents the block signal matrix, X l Represents low-channel EEG signals, C l represents the number of channels of the low-channel EEG signal, n represents the index of the block, c represents the channel index, t represents the position index in the time window, T w Represents the length of each time window, T represents the total time length of the low-channel EEG signal, and N represents the number of blocks.

[0053] S2. Use the state space model to perform feature extraction and fusion processing on the block signal matrix to obtain a fused feature matrix.

[0054] Furthermore, the state space model includes a feature embedding module, a feature enhancement module, a selective scanning module and a feature fusion module which are connected in sequence. The specific process of S2 is as follows:

[0055] S21. The feature embedding module maps each block signal in the block signal matrix to the feature space through a learnable embedding weight matrix to obtain an embedded feature matrix.

[0056] In order to better express and extract the intrinsic features of each block signal in the block signal matrix, feature embedding is performed on each block signal through a learnable embedding weight matrix, and each block signal is mapped from a time domain signal to a feature space. This process is similar to word embedding in natural language processing.

[0057] The embedding feature matrix is ​​specifically expressed as:

[0058] X embed =W embed X patch +b embed

[0059] Where, X embed represents the embedding feature matrix, W embed represents the embedding weight matrix, b embed represents the bias vector, L represents the dimension of the embedded features.

[0060] S22, the feature enhancement module performs average pooling of the channel dimension on the embedded feature matrix and learns the attention weights, applies the attention weights to the embedded feature matrix, realizes selective enhancement of the embedded features, and obtains the enhanced feature matrix.

[0061] Specifically, the feature enhancement module first performs average pooling of the channel dimension on the embedded feature matrix, and then learns the attention weight through feedforward operation. The calculation formula of the attention weight is as follows:

[0062] A ct=σ(W2·ReLU(W1·Avg(X embed )+b1)+b2)

[0063] Among them, A ct represents the attention weight, W1 and W2 represent weight matrices, b1 and b2 represent bias vectors, σ represents the sigmoid activation function, and Avg represents the l The average operation on the dimension, ReLU represents the rectified linear unit activation function, d hidden Represents the feature dimension of the feature enhancement module output.

[0064] This design enables the channel expansion method to focus on important components in the embedded feature matrix, such as specific EEG rhythms or event-related potentials. The use of the sigmoid activation function ensures that the weight value is between 0 and 1, which is conducive to maintaining the stability of the embedded feature matrix signal.

[0065] The attention weights are applied to the embedded feature matrix to achieve selective enhancement of the embedded features and obtain the enhanced feature matrix. This process can be achieved through element-by-element multiplication to ensure that the important features in the embedded feature matrix are enhanced while maintaining the embedded feature matrix X embed structural characteristics.

[0066] The attention weights are applied to the embedded feature matrix to obtain the enhanced feature matrix, which can be specifically expressed as:

[0067] X enhanced =X embed ⊙A ct

[0068] Among them, X enhanced represents the enhanced feature matrix, ⊙ represents an element-by-element multiplication operation.

[0069] The core purpose of feature enhancement is to enhance the key features embedded in the feature matrix signal and suppress irrelevant features. This feature enhancement mechanism can improve the sensitivity of the channel enhancement method to key EEG features, adaptively adjust the feature weights of different time points and different channels, and facilitate subsequent spatial resolution reconstruction.

[0070] S23, the selective scanning module performs selective scanning operations on the enhanced feature matrix in several preset different directions to obtain scanning feature matrices in several different directions.

[0071] Specifically, several different directions are preset, and the several different directions can be set arbitrarily. In this embodiment, four different scanning directions are preset, including: horizontal from left to right (htb), horizontal from right to left (hbt), vertical from top to bottom (vtb) and vertical from bottom to top (vbt).

[0072] Selective scanning can fully capture the spatial dependencies of EEG signals by scanning the enhanced feature matrix in the above four different directions. Each direction of scanning is equipped with an independent transformation matrix, which can learn the feature expression specific to each direction. This multi-directional selective scanning strategy can improve the state space model's ability to model spatial features.

[0073] The scanning feature matrices in several different directions in S23 are specifically expressed as:

[0074] F d =S d (X enhanced )=W d Scan d (X enhanced )+b d , d∈{htb,hbt,vtb,vbt}

[0075] In the formula, F d represents the scanning feature matrix in direction d, S d (·) represents the scanning operation in direction d, X enhanced represents the enhanced feature matrix, Scan d (·) represents the reordering operation in direction d, W d and b d They represent transformation weight and bias respectively, htb represents horizontal from left to right, hbt represents horizontal from right to left, vtb represents vertical from top to bottom, and vbt represents vertical from bottom to top.

[0076] S24, the feature fusion module receives several scanning feature matrices in different directions and performs weighted fusion and layer normalization processing to obtain a fused feature matrix.

[0077] Specifically, the scan feature matrices in four directions are integrated to form a complete feature representation. Through weighted fusion and layer normalization, it can be ensured that the scan feature matrices in different directions can contribute to the final feature representation in a balanced manner. This process not only integrates the spatial information in multiple directions, but also ensures the stability of the features through normalization operations.

[0078] Furthermore, the fusion feature matrix can be specifically expressed as:

[0079]

[0080] Among them, F merged Represents the fused feature matrix, and LayerNorm represents layer normalization.

[0081] In this embodiment, the feature extraction and fusion adopts the state space model SS2D module. The feature extraction and fusion may also adopt other neural networks, such as CNN (Convolutional Neural Networks) and Transformer modules.

[0082] S3. Perform statistical analysis on the fused feature matrix, perform channel expansion on the statistical features obtained after the statistical analysis, and generate an expanded feature matrix.

[0083] Specifically, channel expansion is a key step to achieve spatial resolution improvement. First, the fusion feature F merged Perform statistical analysis, specifically, perform average calculation along the channel dimension to obtain the statistical feature Mean (F merged ), according to the expansion requirements, the statistical feature Mean(F merged ) is repeatedly expanded to form the initial features of the new channel.

[0084] For example, the current EEG only has channel 1 and channel 2, the time dimension is 10 points, and the corresponding matrix size is 2*10. Now to double the spatial resolution, it actually means to convert 2*10 into 4*10, and expand 2 channels into 4 channels. The initial features of the 2 new channels here are generated based on the average features of the 2 old channels.

[0085] This expansion method based on the characteristics of existing channels ensures the spatial consistency and physiological rationality between the newly generated channels and the original channels.

[0086] Furthermore, the extended feature matrix in S3 is specifically expressed as:

[0087] F new =Expand(Mean(F merged ), C h -C l )

[0088] In the formula, represents the extended feature matrix, Mean represents the average operation on the channel dimension, Expand represents the feature copy expansion operation, C h Indicates the number of channels of high-channel EEG signals, C l Indicates the number of channels of low-channel EEG signals, C h >C l .

[0089] S4. Concatenate the extended feature matrix and the fused feature matrix, and map the concatenated feature matrix back to the time domain signal through a feedforward operation to obtain a high-channel EEG signal.

[0090] Specifically, the fused feature matrix and the expanded feature matrix are concatenated, and then the concatenated feature matrix is ​​mapped back to the time domain signal through a feedforward operation to reconstruct the high-channel EEG signal. This process needs to ensure the temporal continuity and spatial consistency of the reconstructed signal while maintaining the physiological characteristics of the signal.

[0091] Furthermore, the high-channel EEG signal in S4 is specifically expressed as:

[0092] X h =W out ·F concat +b out

[0093] Among them, F concat =Concat(F merged , F new )

[0094] In the formula, represents the concatenated feature matrix, W out and b out denote the output weight and output bias respectively, X h represents the high-channel EEG signal,

[0095] In one embodiment, an electroencephalogram channel expansion system uses the above-mentioned data-driven electroencephalogram channel expansion method to perform channel expansion processing on the electroencephalogram. The system includes a computer system and an electroencephalogram signal acquisition module connected to the computer system. The computer system is provided with a signal preprocessing module, a state space module, a channel expansion module and a signal reconstruction module connected in sequence. The state space module is also connected to the signal reconstruction module, wherein:

[0096] The EEG signal acquisition module is used to obtain low-channel EEG signals and transmit the low-channel EEG signals to the computer system;

[0097] The signal preprocessing module is used to process the low-channel EEG signal in blocks to obtain a number of block signals, and arrange the block signals into a block signal matrix;

[0098] The state space model is used to receive the block signal matrix and perform feature extraction and feature fusion to obtain a fused feature matrix;

[0099] The channel expansion module is used to receive the fused feature matrix and perform channel expansion to obtain an extended feature matrix;

[0100] The signal reconstruction module is used to receive the fused feature matrix and the extended feature matrix, and perform splicing and signal reconstruction on them to obtain a high-channel EEG signal.

[0101] For the specific definition of an EEG channel expansion system, please refer to the definition of a data-driven EEG channel expansion method above, which will not be repeated here.

[0102] Furthermore, a data-driven EEG channel expansion method in the present invention is specifically described using a specific embodiment.

[0103] Assume that a low-channel EEG signal is input. The low-channel EEG signal contains 1 channel, the sampling rate is 256Hz, and the signal length is 1 second. The low-channel EEG signal is represented as The goal is to reconstruct it into a 2-channel high channel signal The specific processing process is as follows:

[0104] Step 1: Process the low-channel EEG signal in blocks to obtain block signals:

[0105] Choose the time window length T w = t2, the low-channel EEG signal is divided along the time dimension to obtain 8 block signals, and the 8 block signals are spliced ​​along the channel dimension to form a block signal matrix (1 channel × 8 blocks);

[0106] Step 2: Select the state space model and set the embedding dimension L = t2. The feature embedding module (i.e., embedding layer) of the state space model maps the block signal matrix to the feature space through a learnable embedding weight matrix to obtain an embedded feature matrix Feature augmentation module for state-space models learns attention weights Apply the attention weights to the embedded feature matrix to obtain the enhanced feature matrix The selective scanning module of the state space model scans in four preset directions to generate scanning feature matrices in four directions. The feature fusion module of the state space model receives the scanned feature matrices in four directions and performs weighted fusion and layer normalization to obtain the fused feature matrix

[0107] Step 3: Perform statistical analysis on the fused feature matrix, perform channel expansion on the statistical features obtained after the statistical analysis, and generate an expanded feature matrix with 1 new channel

[0108] Step 4: Expand the feature matrix And the fusion feature matrix Splicing, the spliced ​​feature matrix is ​​concatenated through a learnable output layer Mapping back to the time domain signal, we get a high channel signal with 2 channels

[0109] Assume that the original signal fragment is:

[0110] [0.12, 0.15, 0.18, 0.22, ..., 0.35] # 256 sampling points

[0111] After being processed by the above method, the following two channel signals are generated at the corresponding positions:

[0112] Channel 1: [0.12, 0.15, 0.18, 0.22, ..., 0.35] #keep the original signal

[0113] Channel 2: [0.10, 0.16, 0.15, 0.23, ..., 0.34] # The original signal value of the newly generated channel

[0114] The newly generated channels maintain the overall trend of the signal while adding subtle changes through the learned spatial relationship, ensuring the spatial continuity and physiological rationality of the signal. The entire processing process achieves an exponential increase in spatial resolution while maintaining the time domain characteristics.

[0115] In addition, the above-mentioned data-driven EEG channel expansion method was applied to actual application scenarios, such as motor imagery recognition tasks. After expanding the number of channels by 2 times, the recognition accuracy of two-category motor imagery increased by about 6.5%, and the recognition accuracy of four-category motor imagery increased by about 8.0%. This proves that the data-driven EEG channel expansion method proposed in the present invention performs well in actual brain-computer interface applications.

[0116] The above-mentioned data-driven EEG channel expansion method and system obtains low-channel EEG signals and divides them along the time dimension, integrates several block signals obtained by division along the channel dimension to form a block signal matrix, performs feature extraction and fusion processing on each block signal in the block signal matrix to obtain a fused feature matrix, performs channel expansion on the fused feature matrix to generate an extended feature matrix, splices the extended feature matrix and the fused feature matrix, and maps the spliced ​​feature matrix back to the time domain signal through a feedforward operation to obtain a high-channel EEG signal. This method applies the state space model to the channel expansion task of EEG signals, and can accurately reconstruct multi-channel high-precision EEG signals based on low-channel EEG signals collected by a small number of electrodes, achieving the goal of obtaining high-quality EEG signals while keeping the hardware simple. Compared with existing methods, it has stronger time series modeling capabilities and feature extraction capabilities, and can better capture the spatiotemporal characteristics of EEG signals by introducing attention mechanisms and multi-directional scanning mechanisms.

[0117] The above is a detailed introduction to a data-driven EEG channel expansion method and system provided by the present invention. This article uses specific examples to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the core idea of ​​the present invention. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present invention, the present invention can also be improved and modified, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A data-driven electroencephalogram channel expansion method, characterized in that: The method comprises the following steps: S1. Obtain a low-channel EEG signal, divide the low-channel EEG signal along the time dimension to obtain a number of block signals, and integrate the several block signals along the channel dimension to form a block signal matrix; S2, using the state space model to perform feature extraction and fusion processing on the block signal matrix to obtain a fusion feature matrix; S3, performing statistical analysis on the fusion feature matrix, performing channel expansion on the statistical features obtained after the statistical analysis, and generating an expanded feature matrix; S4. Concatenate the extended feature matrix and the fused feature matrix, and map the concatenated feature matrix back to the time domain signal through a feedforward operation to obtain a high-channel EEG signal.

2. The data-driven EEG channel expansion method according to claim 1, characterized in that: The block signal matrix in S1 is specifically expressed as: X patch (nc,t)=X l (c,nT w +t),0≤t<T w Where n∈[0,N-1], c∈[0,C l -1],t∈[0,T w -1] Where, X patch表 shows the block signal matrix, Xl represents the low-channel EEG signal, C l represents the number of channels of the low-channel EEG signal, n represents the index of the block, c represents the channel index, t represents the position index in the time window, T w Represents the length of each time window, T represents the total time length of the low-channel EEG signal, and N represents the number of blocks.

3. The data-driven EEG channel expansion method according to claim 2, characterized in that: The state space model includes a feature embedding module, a feature enhancement module, a selective scanning module, and a feature fusion module connected in sequence. The specific process of S2 is as follows: S21, the feature embedding module maps each block signal in the block signal matrix to the feature space through a learnable embedding weight matrix to obtain an embedded feature matrix; S22, the feature enhancement module performs average pooling of the channel dimension on the embedded feature matrix and learns the attention weights, applies the attention weights to the embedded feature matrix, realizes the selective enhancement of the embedded features, and obtains the enhanced feature matrix; S23, the selective scanning module performs selective scanning operations on the enhanced feature matrix in a plurality of preset different directions to obtain scanning feature matrices in a plurality of different directions; S24, the feature fusion module receives several scanning feature matrices in different directions and performs weighted fusion and layer normalization processing to obtain a fused feature matrix.

4. The data-driven EEG channel expansion method according to claim 3, characterized in that: The embedding feature matrix in S21 is specifically expressed as: X embed =W embed X patch +b embed Among them, X embed represents the embedding feature matrix, W embed represents the embedding weight matrix, b embed represents the bias vector, L represents the embedding feature dimension.

5. The data-driven EEG channel expansion method according to claim 4, characterized in that: The enhanced feature matrix in S22 is specifically expressed as: X enhanced =X embed ⊙A ct Among them, A ct =σ(W2·ReLU(W1·Avg(X embed )+b1)+b2) Where, X enhanced represents the enhanced feature matrix, ⊙ represents the element-by-element multiplication operation, A ct represents the attention weight, W1 and W2 both represent weight matrices, b1 and b2 represent bias vectors, σ represents the sigmoid activation function, and Avg represents the l Averaging operation over dimensions.

6. The data-driven EEG channel expansion method according to claim 5, characterized in that: The several different directions preset in S23 include: horizontally from left to right, horizontally from right to left, vertically from top to bottom, and vertically from bottom to top.

7. The data-driven EEG channel expansion method according to claim 6, characterized in that: The scanning feature matrices in several different directions in S23 are specifically expressed as: F d =S d (X ehanced )=W d ·Scan d (X enhanced )+b d ,d∈{htb,hbt,vtb,vbt} In the formula, F d represents the scanning feature matrix in direction d, S d (·) represents the scanning operation in direction d, X enhanced represents the enhanced feature matrix, Scan d (·) represents the reordering operation in direction d, W d and b d They represent transformation weight and bias respectively, htb represents horizontal from left to right, hbt represents horizontal from right to left, vtb represents vertical from top to bottom, and vbt represents vertical from bottom to top.

8. The data-driven EEG channel expansion method according to claim 7, characterized in that: The extended features in S3 are specifically expressed as: F new =Expand(Mean(F merged ),C h -C l ) In the formula, represents the extended feature matrix, F merged represents the fusion feature matrix, Mean represents the average operation in the channel dimension, Expand represents the feature copy expansion operation, C h Indicates the number of channels of high-channel EEG signals, C l Indicates the number of channels of low-channel EEG signals, C h >C l .

9. The data-driven EEG channel expansion method according to claim 8, characterized in that: The high-channel EEG signal in S4 is specifically expressed as: X h =W out ·F concat +b out Among them, F concat =Concat(F merged , F new ) In the formula, represents the concatenated feature matrix, W out and b out denote the output weight and output bias respectively, X h represents the high-channel EEG signal, 10. An electroencephalogram channel expansion system, using the data-driven electroencephalogram channel expansion method according to any one of claims 1 to 9 to expand the electroencephalogram channel, characterized in that: The system includes a computer system and an EEG signal acquisition module connected to the computer system. The computer system is provided with a signal preprocessing module, a state space module, a channel expansion module and a signal reconstruction module connected in sequence. The state space module is also connected to the signal reconstruction module, wherein: The EEG signal acquisition module is used to obtain low-channel EEG signals and transmit the low-channel EEG signals to the computer system; The signal preprocessing module is used to process the low-channel EEG signal in blocks to obtain a number of block signals, and arrange the block signals into a block signal matrix; The state space model is used to receive the block signal matrix and perform feature extraction and feature fusion to obtain a fused feature matrix; The channel expansion module is used to receive the fused feature matrix and perform channel expansion to obtain an extended feature matrix; The signal reconstruction module is used to receive the fused feature matrix and the extended feature matrix, and perform splicing and signal reconstruction on them to obtain a high-channel EEG signal.

Images