Electroencephalogram decoding method based on power spectrum feature enhancement and brain-computer intelligent interface system

Abstract

The invention discloses an electroencephalogram decoding method based on power spectrum feature enhancement and a brain-computer intelligent interface system. The method comprises the following steps: acquiring an original electroencephalogram signal; intercepting a 48-52Hz part in the power spectrum, and performing fitting by using a noise fitting function to obtain a power frequency noise generation power spectrum; subtracting the noise component from the original power spectrum to obtain a denoised power spectrum, and fitting the intensity and frequency of the power spectrum according to a function by using the denoised power spectrum to obtain a power spectrum of a non-oscillation component; subtracting the power spectrum of the non-oscillation component and the variable constant from the denoised power spectrum to obtain a background component, and fitting the background component by using a multi-peak function to obtain an oscillation component; adding the power spectrum generated by the power frequency noise, the non-oscillation component and the oscillation component to obtain a reconstructed and enhanced signal power spectrum; and carrying out electroencephalogram decoding after stacking according to columns. Signal interference generated by irrelevant physiological activities is eliminated, and enhancement of spectral features and high-accuracy electroencephalogram decoding are achieved.

Description

Technical Field

[0001] The present invention relates to the field of brain-computer interface technology, and in particular to an electroencephalogram (EEG) decoding method based on power spectrum feature enhancement and a brain-computer intelligent interface system. Background Art

[0002] The brain-computer interface (BCI) is an emerging technology designed to enable direct communication between the brain and external devices. It has broad application prospects in disease diagnosis, medical rehabilitation, human-computer interaction, and other fields. Steady-state visual evoked potential (SSVEP)-based BCIs are non-invasive BCI technologies that use external visual stimulation of a specific frequency to induce the brain to generate oscillatory EEG signals of the corresponding frequency. These oscillatory EEG signals are then identified, extracted, classified, and decoded using relevant algorithms. This technology has the advantages of requiring minimal user training, high information transmission rates, and fast system response, and has been widely developed.

[0003] In such systems, various algorithms are used to decode EEG signals collected by electrodes and amplifiers under visual stimulation, thereby obtaining conscious instructions from the brain and enabling human-computer interaction. However, factors such as the EEG decoding algorithm can significantly affect the final decoding accuracy.

[0004] Decoding algorithms are one of the core technologies of brain-computer interfaces. In recent years, traditional machine learning algorithms, such as support vector machines (SVMs) and linear discriminant analysis (LDA), as well as rapidly developing deep learning algorithms, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), have gained widespread development. During the decoding process, the extraction and analysis of oscillatory electrical signals generated by the periodic oscillations of brain neuronal populations is particularly important.

[0005] Current research often uses methods such as Fourier transform to perform frequency domain analysis on EEG signals, and extracts different oscillatory EEG signals based on a fixed frequency range. Common ones include delta waves (1-4Hz), theta waves (4-8Hz), alpha waves (8-12Hz), beta waves (12-30Hz), etc.

[0006] However, factors such as individual differences, brain region differences, and cognitive state differences significantly affect the ability of such methods to extract key features such as the center frequency, frequency bandwidth, and signal power of oscillation signals under different tasks; at the same time, the non-oscillatory signals generated by irregular spontaneous activities of the brain further reduce the ability of such methods to extract key features of oscillation signals, ultimately affecting the operating efficiency and decoding accuracy of the brain-computer interface system. Summary of the Invention

[0007] In response to the shortcomings of the existing technology, the present invention aims to provide an EEG decoding method and brain-computer intelligent interface system based on power spectrum feature enhancement. The system can achieve synchronous acquisition of visual stimulation and EEG data, thereby improving the system's temporal resolution and response speed. At the same time, the system can effectively separate the noise and non-oscillatory components in the EEG power spectrum data through an EEG decoding method based on power spectrum feature enhancement, while efficiently extracting the oscillatory components, eliminating signal interference caused by irrelevant physiological activities, achieving spectral feature enhancement, and laying a good foundation for high-accuracy EEG decoding.

[0008] The technical solution adopted by the present invention to solve the technical problem is:

[0009] A method for decoding electroencephalogram (EEG) based on power spectrum feature enhancement, comprising the following steps:

[0010] 1) Obtain the original EEG signal, and extract the original power spectrum P of the single channel EEG signal after re-referencing and removing baseline drift. xx (f);

[0011] 2) Intercept the 48-52Hz part of the power spectrum psd and use the noise fitting function Perform fitting to obtain the power spectrum P generated by the power frequency noise noise (50Hz); then use the original power spectrum P xx (f) Subtract the noise component to obtain the denoised power spectrum Right now:

[0012]

[0013] 3) Use the denoised power spectrum The power spectrum intensity and frequency are The function is fitted, and the fitted curve represents the power spectrum of the non-oscillatory component

[0014] Where f is the frequency; α, σ, k c is the parameter to be fitted;

[0015] And use the power spectrum after denoising Power spectrum with non-oscillatory components subtracted With the variable constant P0, the background component is obtained Right now:

[0016]

[0017] 4) Using multimodal functions Background components Fitting to obtain the oscillatory components

[0018] Among them, A and μ are k0-dimensional vectors, which are the parameters to be fitted; Σ is a k0×k0-dimensional matrix, which represents the parameters to be fitted; k0 is the preset parameter;

[0019] 5) Use power frequency noise to generate power spectrum P noise (50Hz), non-oscillating components Oscillation components Add and obtain the reconstructed and enhanced EEG signal x(t) power spectrum

[0020] 6) Based on the Euclidean distance D, use the formula Evaluate the oscillation peak accuracy, f i is the frequency of the i-th sampling point, n is the number of sampling points;

[0021] If the value of ρ is not within the range of 0.6-0.8, adjust the values ​​of P0 in step 3) and k0 in step 4) and repeat steps 3)-5) until ρ is within the range of 0.6-0.8.

[0022] 7) Perform the above steps for each EEG channel to enhance the EEG signals of all channels, stack them by columns, and convert them into matrix X;

[0023] 8) For the matrix X, use the convolutional neural network model to perform EEG decoding.

[0024] Furthermore, in step 6), if ρ is too large, it means that the enhanced power spectrum feature is too close to the original power spectrum feature, and the extracted components contain more invalid components. It is necessary to increase P0 and reduce k0. If ρ is too small, it means that the enhanced power spectrum feature cannot contain the effective components in the original power spectrum feature. It is necessary to reduce P0 and increase k0.

[0025] Furthermore, P0 decreases or increases by 5% to 10% each time; k0 increases or decreases by 1 each time.

[0026] Furthermore, the process of obtaining the original EEG signal is:

[0027] The visual stimulation EEG acquisition software uses variable-length multi-color rectangular window encoding to control the display output image and generate the stimulus sequence. Its specific implementation is as follows:

[0028] (1) The width of the display is W and the height is H. Ten rectangular patterns are displayed on the screen. Each pattern has a width of w0. The ten rectangular patterns are evenly spaced along the width direction and are numbered 1-10 from left to right. There is a red solid circular pattern at the geometric center of the rectangular pattern with a diameter of 0.25w0. Before stimulation, the bottom edge of each rectangular pattern is located at the bottom of the screen. The initial height of the rectangular pattern is h0, and its color corresponds to the RGB color number (0, 0, 0). The pattern display brightness is l0.

[0029] (2) When the kth pattern is stimulated, the height h of the kth rectangular pattern k , RGB value of color, display brightness l k All of them change, and the changes are as follows:

[0030]

[0031] Among them, R k , G k 、B k Represents the RGB value of the color of the k-th rectangular pattern; f sk0 、 The stimulation frequency and initial phase parameters preset in the encoding mode; max is the maximum brightness of the display; t is the stimulus start time; max(0, x) is the positive function;

[0032] The electroencephalogram (EEG) signal is collected under the stimulation of the stimulation sequence to obtain the original electroencephalogram (EEG) signal.

[0033] Furthermore, in step 8), the structure of the convolutional neural network model is:

[0034]

[0035] a=min(e z , C)

[0036] y=W fc flatten(a)+b fc

[0037]

[0038] Among them, x ij ∈X,w ij ∈W, W is the neural network weight matrix, C is a constant, flatten is to expand the matrix in one dimension, b fc 、W fc are the bias matrix and weight matrix of the full connection; m represents the number of EEG channels, n represents the number of sampling points; X represents the input; z is the intermediate variable; a is the output of the activation function; b is the bias parameter of the neural network; y is the output of the full connection; Represents the model output.

[0039] Furthermore, after the oscillation components are obtained by fitting in step 4), they are resampled in the range of 0-60 Hz, and the number of sampling points in the resampling process is set to p, where p is less than n and is equidistant sampling; the resampled oscillation components of all EEG channels constitute an optimized power spectrum sequence matrix, and the matrix size is m×p; this optimized power spectrum sequence matrix is ​​input into the convolutional neural network model for EEG decoding.

[0040] The present invention also protects a brain-computer intelligent interface system, which executes the decoding method and includes an EEG acquisition module, a brain rhythm enhancement module and an EEG decoding module.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] (1) The method of the present invention does not rely on filters with predefined frequency bands (e.g., 8–12 Hz for the α band and 12–30 Hz for the β band) commonly used in traditional methods. Instead, it extracts the characteristic vectors of neural oscillation components, including the characteristic frequency μ, signal power A, and bandwidth Σ, through a data-driven approach using multiple fittings of different functions. This more accurately captures individual differences and dynamic changes. This improves the accuracy of the analysis, avoids the bias caused by predefined frequency bands, and enables the results to more realistically reflect the complexity and diversity of neural activity.

[0043] (2) The method of the present invention shows remarkable convenience and efficiency in the signal processing process. One of its core advantages is that it does not require complex filtering preprocessing. When processing signals, traditional methods often need to first use filtering preprocessing steps to remove the noise components generated by power frequency interference, which not only increases the processing steps, but may also introduce additional errors or distortions. The method of the present invention can automatically identify different brain activity areas under different conditions and remove the components in the signal corresponding to power frequency interference. It directly uses the original EEG signal for analysis and processing, eliminating the influence of external factors such as muscle activity, external electric field, power frequency interference, etc. on the data, simplifying the preprocessing process, reducing the preprocessing steps, and avoiding the loss or misjudgment of signal features that may be caused by improper filtering, thereby improving the overall efficiency and accuracy of signal processing. In this way, the method of the present invention provides a more direct, efficient and reliable solution for signal analysis, which is particularly suitable for scenarios where signals need to be processed quickly and accurately.

[0044] (3) The SSVEP stimulation paradigm in the present invention can enhance the amplitude, signal-to-noise ratio and other characteristics of the SSVEP signal by optimizing the stimulation method, making it easier to detect and identify, thereby improving the accuracy and reliability of the brain-computer interface system, while reducing visual fatigue and improving user comfort. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 Schematic diagram of the working process of the brain-computer intelligent interface system.

[0046] Figure 2 Schematic diagram of the data processing flow of the EEG rhythm enhancement module. DETAILED DESCRIPTION

[0047] The present invention is further explained below with reference to the embodiments and drawings, but they are not intended to limit the scope of protection of the present application.

[0048] Example 1

[0049] The present invention provides an EEG decoding method based on power spectrum feature enhancement, comprising the following steps:

[0050] 1) Obtain the original EEG signal, and extract the original power spectrum P of the single channel EEG signal after re-referencing and removing baseline drift. xx (f);

[0051] 2) Intercept the 48-52Hz part of the power spectrum psd (containing frequency and power spectrum intensity) and use the noise fitting function Perform fitting to obtain the power spectrum P generated by the power frequency noise noise (50Hz); then use the original power spectrum P xx (f) Subtract the noise component to obtain the denoised power spectrum Right now:

[0052]

[0053] 3) Use the denoised power spectrum The power spectrum intensity and frequency are calculated according to The function is fitted, and the fitted curve represents the power spectrum of the non-oscillatory component

[0054] Among them, f is the frequency, α, σ, k c is the parameter to be fitted;

[0055] And use the power spectrum after denoising Power spectrum with non-oscillatory components subtracted With the variable constant P0, the background component is obtained Right now:

[0056]

[0057] 4) Using multimodal functions Background components Fitting to obtain the oscillatory components

[0058] Where μ is a k0-dimensional vector representing the characteristic frequency; Σ is a k0×k0-dimensional matrix representing the bandwidth; k0 is a preset parameter representing the number of times the final fitting is performed, and the number of k0s is consistent with the number of peaks in the multimodal function; A is a k0-dimensional vector representing the signal power;

[0059] The eigenvector of the oscillation component can be determined by fitting, where μ represents the characteristic frequency of the oscillation component, A represents the signal power, and Σ represents the bandwidth.

[0060] 5) Use power frequency noise to generate power spectrum P noise (50Hz), non-oscillating components Oscillation components

[0061] Add and obtain the reconstructed and enhanced EEG signal x(t) power spectrum

[0062] 6) Based on the Euclidean distance D, use the formula Evaluate the oscillation peak accuracy, f i is the frequency of the i-th sampling point, n is the number of sampling points;

[0063] If the value of ρ is not within the range of 0.6-0.8, adjust the values ​​of P0 in step 3) and k0 in step 4) and repeat steps 3)-5) until ρ is within the range of 0.6-0.8.

[0064] 8) Perform the above steps for each EEG channel to enhance the EEG signals of all channels, stack them by columns, and convert them into matrix X;

[0065] 8) For the matrix X, use the convolutional neural network model to perform EEG decoding.

[0066] Example 2

[0067] The brain-computer intelligent interface system of this embodiment includes:

[0068] 1. EEG acquisition module:

[0069] The high-time-precision EEG acquisition function includes the following: a head-mounted 32-channel electrode cap and signal amplifier that complies with the 10-20 international standard, a 240 Hz refresh rate monitor, a computer, and visual stimulation EEG acquisition software.

[0070] The visual stimulation EEG acquisition software uses variable-length multi-color rectangular window encoding to control the display output image and generate the stimulus sequence. Its specific implementation is as follows:

[0071] (1) The width of the display is W and the height is H. Ten rectangular patterns are displayed on the screen. Each pattern has a width of w0. The ten rectangular patterns are evenly spaced along the width direction and are numbered 1-10 from left to right (the stimulation frequency and initial phase of different modes are different and can be set according to actual conditions). There is a red solid circular pattern at the geometric center of the rectangular pattern with a diameter of 0.25w0. Before stimulation, the bottom edge of each pattern is at the bottom of the screen, its initial height is h0, its color corresponds to the RGB color number (0, 0, 0), and the pattern display brightness is l0.

[0072] (2) When the kth pattern is stimulated, the height h of the kth rectangular pattern k , RGB values ​​of the color (respectively R k , G k , B k Display brightness l k All of them change, and the changes are as follows:

[0073]

[0074] Among them, f sk0 、 The stimulation frequency and initial phase parameters preset in the encoding mode; max is the maximum brightness of the display, t is the stimulus start time, max(0, x) is the positive function, that is,

[0075]

[0076] Where x is the variable of the positive function.

[0077] 2. Brain Rhythm Enhancement Module:

[0078] (1) The collected original EEG signal is preprocessed by re-referencing and removing baseline drift. Re-referencing can eliminate common mode noise, reduce the influence of reference electrodes on the signal, and improve signal consistency. Removing the baseline can eliminate the baseline drift caused by equipment, environment and other factors in the signal. The original power spectrum P of the single channel EEG signal x(t) is obtained using the Welch method (power spectrum density estimation method). xx (f). Where P xx (f) is a 31×n dimensional matrix, where 31 is the number of EEG signal channels and n is the number of power spectrum sampling points calculated using the Welch method.

[0079] (2) Intercept the 48-52Hz part of the power spectrum PSD (the power spectrum contains frequency and power spectrum intensity data) and use the noise fitting function Perform fitting to obtain the power spectrum P generated by the power frequency noise noise (50Hz); then use the original power spectrum Pxx (f) Subtract the noise component to obtain the denoised power spectrum Right now:

[0080]

[0081] (3) Using the denoised power spectrum The power intensity and frequency are The function is fitted, and the fitted curve represents the power spectrum of the non-oscillatory component

[0082] Among them, f is the frequency, α, σ, and k are the parameters to be fitted.

[0083] And use the power spectrum after denoising Power spectrum with non-oscillatory components subtracted With the variable constant P0, the background component is obtained Right now:

[0084]

[0085] Among them, P0 is a variable constant used to adjust the similarity between the enhanced spectral features and the original spectral features.

[0086] (4) Using multimodal functions Background components Fitting to obtain the oscillatory components

[0087] Where μ is a k0-dimensional vector, representing the center value of each peak in the multi-peak function, that is, the maximum point of the function, corresponding to the center frequency of each component; Σ is a k0×k0-dimensional matrix; A is a k0-dimensional vector, representing the function value of each extreme point of the multi-peak function, corresponding to the maximum power of each component; all three are parameters to be fitted; k0 is a preset parameter, representing the number of times the final fitting is performed, and the number of k0 is consistent with the number of peaks.

[0088] (5) Use power frequency noise to generate power spectrum P noise (50Hz), non-oscillating components Oscillation components Add and obtain the reconstructed and enhanced EEG signal x(t) power spectrum

[0089] (6) Based on the Euclidean distance D, use the formula Evaluate the oscillation peak accuracy, f i is the frequency of the i-th sampling point;

[0090] The value range of ρ is (0, 1]. The closer the value is to 1, the higher the similarity between the enhanced power spectrum feature and the original power spectrum. In actual use, the value of ρ should be between 0.6 and 0.8. If the value of ρ does not meet the requirements, it is necessary to adjust P0 in step (3) and k0 in step (4), and complete steps (3) to (5) again.

[0091] If ρ is too large, it means that the enhanced feature is too close to the original power spectrum, and the extracted components contain more invalid components. It is necessary to increase P0 and reduce k0. If ρ is too small, it means that the enhanced power spectrum feature cannot contain the effective components in the element spectrum feature. It is necessary to reduce P0 and increase k0.

[0092] P0 decreases or increases by 5 to 10% each time, and k0 increases or decreases by 1 each time.

[0093] where f i P xx Frequency sampling points calculated in (f).

[0094] (7) Perform the above steps for each EEG channel to complete the processing of the EEG signals of all channels, stack them by columns, and convert them into matrix X.

[0095] 3. EEG decoding module:

[0096] For the multi-channel optimized power spectrum sequence matrix X, a convolutional neural network model is used, in which the activation function a=min(e z , C) decode it.

[0097] The convolutional neural network model structure is as follows:

[0098]

[0099] a=min(e z , C)

[0100] y=W fc flatten(a)+b fc

[0101]

[0102] Among them, x ij ∈X,w ij ∈W, W is the neural network weight matrix, C is a constant, flatten is to expand the matrix in one dimension, b fc 、W fc are the bias matrix and weight matrix of the full connection; m = 31 represents the number of EEG channels, n represents the number of sampling points; X represents the input; z is the intermediate variable; a is the output of the activation function; b is the bias parameter of the neural network; y is the output of the full connection; Represents the model output.

[0103] The EEG signals were collected and divided into datasets. Ten types of patterns were used for testing, and the f set in each type of pattern was sk0 As the EEG data label, the convolutional neural network model training is completed on the dataset to realize the decoding of EEG signals.

[0104] The present invention adopts activation function a=min(e z By introducing nonlinearity and manually setting the constant C, we can effectively prevent gradient explosion while controlling model complexity, improving model generalization and optimizing convergence speed. Experiments show that the decoding accuracy reaches 95% when testing and decoding using the visual stimulation method in step 1.

[0105] Example 3

[0106] This embodiment, based on Example 1, resamples the oscillatory components obtained by fitting in step 4) within the 0-60 Hz range. The number of sampling points during the resampling process is set to p, where p is less than n and the sampling is equidistant. The resampled oscillatory components of all EEG channels form an optimized power spectrum sequence matrix of size m × p. This optimized power spectrum sequence matrix is ​​input into a convolutional neural network model for EEG decoding. In this case, steps 7) and 8) of Example 1 are omitted, while steps 5) and 6) are still performed to ensure that appropriate P0 and k0 are obtained.

[0107] This processing method can also significantly improve the decoding accuracy and has a good effect.

[0108] Any matters not described in the present invention are applicable to the prior art.

Claims

1. An EEG decoding method based on power spectrum feature enhancement, characterized in that: The decoding method comprises the following steps: 1) Obtain the original EEG signal, and extract the original power spectrum P of the single channel EEG signal after re-referencing and removing baseline drift. xx (f); 2) Intercept the 48-52Hz part of the power spectrum psd and use the noise fitting function Perform fitting to obtain the power spectrum P generated by the power frequency noise noise (50Hz); then use the original power spectrum P xx (f) Subtract the noise component to obtain the denoised power spectrum Right now: 3) Use the denoised power spectrum The power spectrum intensity and frequency are calculated according to The function is fitted, and the fitted curve represents the power spectrum of the non-oscillatory component Where f is the frequency; α, σ, k c is the parameter to be fitted; And use the power spectrum after denoising Power spectrum with non-oscillatory components subtracted With the variable constant P0, the background component is obtained Right now: 4) Using multimodal functions Background components Fitting to obtain the oscillatory components Among them, A and μ are k0-dimensional vectors, which are the parameters to be fitted; Σ is a k0×k0-dimensional matrix, which represents the parameters to be fitted; k0 is the preset parameter; 5) Use power frequency noise to generate power spectrum P noise (50Hz), non-oscillating components Oscillation components Add and obtain the reconstructed and enhanced EEG signal x(t) power spectrum 6) Based on the Euclidean distance D, use the formula Evaluate the oscillation peak accuracy, f i is the frequency of the i-th sampling point, n is the number of sampling points; If the value of ρ is not within the range of 0.6-0.8, adjust the values ​​of P0 in step 3) and k0 in step 4) and repeat steps 3)-5) until ρ is within the range of 0.6-0.

8. 7) Perform the above steps for each EEG channel to enhance the EEG signals of all channels, stack them by columns, and convert them into matrix X; 8) For the matrix X, use the convolutional neural network model to perform EEG decoding.

2. The decoding method according to claim 1, wherein: In step 6), if ρ is too large, it means that the enhanced power spectrum feature is too close to the original power spectrum feature, and the extracted components contain more invalid components. It is necessary to increase P0 and reduce k0. If ρ is too small, it means that the enhanced power spectrum feature cannot contain the valid components in the original power spectrum feature. It is necessary to reduce P0 and increase k0.

3. The decoding method according to claim 2, wherein: P0 decreases or increases by 5% to 10% each time; k0 increases or decreases by 1 each time.

4. The decoding method according to claim 1, wherein: The process of acquiring raw EEG signals is as follows: the visual stimulation EEG acquisition software uses variable-length multi-color rectangular window encoding to control the display output image and generate the stimulus sequence. The specific implementation is as follows: (1) The width of the display is W and the height is H. Ten rectangular patterns are displayed on the screen. Each pattern has a width of w0. The ten rectangular patterns are evenly spaced along the width direction and are numbered 1-10 from left to right. There is a red solid circular pattern at the geometric center of the rectangular pattern with a diameter of 0.25w0. Before stimulation, the bottom edge of each rectangular pattern is located at the bottom of the screen. The initial height of the rectangular pattern is h0, and its color corresponds to the RGB color number (0,0,0). The pattern display brightness is l0. (2) When the kth pattern is stimulated, the height h of the kth rectangular pattern k , RGB value of color, display brightness l k All of them change, and the changes are as follows: Among them, R k , G k 、B k Represents the RGB value of the color of the k-th rectangular pattern; f sk0 、 The stimulation frequency and initial phase parameters preset in the encoding mode; max is the maximum brightness of the display; t is the stimulus start time; max(0,x) is the positive function; The electroencephalogram (EEG) signal is collected under the stimulation of the stimulation sequence to obtain the original electroencephalogram (EEG) signal.

5. The decoding method according to claim 1, wherein: In step 8), the structure of the convolutional neural network model is: a=min(e z ,C) y=W fc ·flatten(a)+b fc Among them, x ij ∈X,w ij ∈W, W is the neural network weight matrix, C is a constant, flatten is to expand the matrix in one dimension, b fc 、W fc are the bias matrix and weight matrix of the full connection; m represents the number of EEG channels, n represents the number of sampling points; X represents the input; z is the intermediate variable; a is the output of the activation function; b is the bias parameter of the neural network; y is the output of the full connection; Represents the model output.

6. The decoding method according to claim 1, wherein: After fitting and obtaining the oscillation components in step 4), they are resampled in the range of 0-60 Hz. The number of sampling points in the resampling process is set to p, where p is less than n and the sampling is equidistant. The resampled oscillation components of all EEG channels constitute an optimized power spectrum sequence matrix with a matrix size of m×p. This optimized power spectrum sequence matrix is ​​input into the convolutional neural network model for EEG decoding.

7. A brain-computer intelligent interface system, characterized in that: The system executes the decoding method described in any one of claims 1-6, including an EEG acquisition module, a brain rhythm enhancement module and an EEG decoding module.

Images