Brain control system based on non-invasive brain-computer interface and its implementation method

Abstract

The present invention provides a brain control system based on a non-invasive brain-computer interface and an implementation method thereof, which includes two systems. The second system is an extension and improvement of the first system, and includes a signal acquisition module, which provides brain imagination induction and electrical signal acquisition of the cerebral cortex; a signal preprocessing module, which processes the acquired electrical signals; a feature projection matrix establishment module, which first uses the signal to establish a feature projection matrix and perform feature extraction, and then uses an improved power spectrum density method to extract features, and merges the eigenvalues ​​extracted by the two methods to obtain the final eigenvalue. A classifier modeling module, which uses the eigenvalues ​​for classifier modeling; a signal conversion module, which converts real-time EEG signals into control instructions. The present invention is aimed at brain-computer interface control and adds a new scheme for brain-computer interface control, which can convert the brain's motor imagination signals into control instructions to control and manipulate the outside world, and greatly improves the accuracy of EEG signal discrimination.

Description

Technical Field

[0001] The present invention relates to the technical field of brain-computer interface, and specifically to a brain control system based on a non-invasive brain-computer interface and an implementation method thereof. Background Art

[0002] Different from traditional control devices, brain-computer interface control is a new type of technology that does not rely on the triggering of the body's neural network and muscles, but directly reads the signals from the cerebral cortex, allowing the brain to directly control external devices. This technology involves various interdisciplinary subjects, such as neurology, signal processing, signal conversion, etc., and is a hot topic of current research.

[0003] At present, this technology can serve patients with nerve or limb injuries well, allowing them to issue instructions to devices directly through their brains to achieve rehabilitation treatment, or enable them to manipulate various devices to live more independently. In the future, this technology can be applied to virtual reality technology, allowing users to interact more directly with the virtual world, making the user experience more realistic, and can also be used to manipulate various devices such as robots, computers, etc., and manipulate all computer-controllable devices.

[0004] At present, domestic brain-computer interface devices are mostly dependent on foreign countries. They are expensive, bulky, and the technology is not mature enough, so it is difficult to widely use brain control systems. Currently, this technology is most widely used in universities, research institutes and hospitals, but its prospects are bright. Summary of the invention

[0005] In order to overcome the shortcomings of domestic brain-computer interface control technology, the present invention proposes a new brain-computer interface control system for the problem of brain-computer interface control, which realizes the process of converting EEG signals into control signals. A new method for brain-computer interface control is also proposed, which greatly enhances the accuracy of brain-computer control. The present invention is a system and method for converting cerebral cortex signals into control signals for machines.

[0006] To achieve the above object, the present invention provides a brain control system based on a non-invasive brain-computer interface, comprising two systems, the first system comprising:

[0007] Signal acquisition module, providing brain imagination induction and cerebral cortex electrical signal acquisition;

[0008] The signal preprocessing module performs cutting and processing on the collected electrical signals, calculates the power spectrum density of the signals using the Fisher criterion theory, and selects the channel with high discrimination;

[0009] The feature extraction module uses the improved power spectrum density method to extract features from the processed signal and obtain the eigenvalue;

[0010] A classifier modeling module uses feature values ​​for classifier modeling;

[0011] The signal conversion module obtains control instructions after preprocessing, feature extraction and classification of real-time EEG signals.

[0012] Furthermore, in the signal acquisition module, the imagination induction provided to the brain is in the form of visual induction, providing visual manipulation instructions to induce the organism to actively imagine; the electrical signal acquisition of the cerebral cortex adopts a non-invasive brain-computer interface, using 32 electrodes to collect cerebral cortex signals, and then collecting earlobe signals as reference electrodes.

[0013] Furthermore, in the signal preprocessing module, the collected electrical signals are clipped to eliminate useless signals. The signals are then filtered using a 5th-order Butterworth filter to obtain 8-30Hz signals. The filtered signals are measured and sorted for power spectral density separability of each channel signal between 8-30Hz using the Fisher criterion, and the 15 electrode signals with the strongest separability are selected.

[0014] Furthermore, the feature extraction module performs PSD feature extraction on the selected 15 electrode signals, takes the average power spectral density of the frequency bands where the Mu and Beta waves of the 15 channels are located as the feature value, and constructs a feature vector.

[0015] Furthermore, the classifier modeling module uses a linear discriminant analysis algorithm to establish a classifier data model using the feature vector obtained by feature extraction.

[0016] The system is based on a brain control implementation method of a non-invasive brain-computer interface and includes two parts: the first part is the training part, and the second part is the practical operation part;

[0017] The training part includes the following:

[0018] First, the signal acquisition module collects the signal, and the signal preprocessing module cuts and classifies the collected signal, and then uses a 5th-order Butterworth filter to filter the signal to obtain an 8-30Hz signal. The filtered signal is sorted by the power spectral density average value of each channel signal between 8-30Hz, using the idea that the larger the difference between the average values ​​between classes, the better, and the smaller the variance within the class, the better, and the separability is sorted, and the 15 electrode signals with the strongest power spectral density separability are selected.

[0019] The feature extraction module performs PSD feature extraction on the selected 15 electrode signals, takes the average power spectral density of the frequency bands where the Mu and Beta waves of these 15 channels are located as the eigenvalue, and constructs the eigenvector.

[0020] The feature vector obtained by feature extraction is used to establish a classifier data model using a linear discriminant analysis algorithm;

[0021] The practical operation part includes the following contents:

[0022] The read signal is preprocessed, and the electrode signal selected in the training part is used. Then, the average power spectral density of the frequency bands where the Mu and Beta waves of the 15 channels are located is taken to form a feature vector. The feature vector is judged by the imagination category using the classifier data model obtained in the training part and converted into a control instruction. The obtained control instruction is transmitted to various control targets to achieve the desired conversion of brain imagination into control commands in the present invention.

[0023] The second system is an extension and improvement of the first system. The second system includes:

[0024] Signal acquisition module, providing brain imagination induction and cerebral cortex electrical signal acquisition;

[0025] The signal preprocessing module performs cutting and processing on the collected electrical signals, calculates the power spectrum density of the signals using the Fisher criterion theory, and selects the channel with high discrimination;

[0026] The feature projection matrix building module first uses the processed signal to build the feature projection matrix, and extracts the features of the signal through the projection matrix. Then, the signal is extracted using the power spectrum density method, and the eigenvalues ​​extracted by the two methods are combined to obtain the final eigenvalue.

[0027] Classifier modeling module, which uses the final feature value of the previous step for classifier modeling;

[0028] The signal conversion module obtains control instructions after preprocessing, feature extraction and classification of real-time EEG signals.

[0029] Furthermore, in the signal preprocessing module, the collected electrical signals are clipped to eliminate useless signals. The signal is then filtered using a 5th-order Butterworth filter to obtain a signal of 8-30Hz. The power spectral density of the signal frequency between 8-30Hz is then discriminated by the Fisher criterion function to measure and sort the separability of each electrode signal, and finally the 15 electrode signals with the strongest separability are selected. The selection of lead signals can remove redundant leads with weak separability, reduce the amount of calculation, and reduce the impact on real-time performance due to the high computational complexity of the EMD module. Selecting leads with high power spectral density discrimination is of great help to the eigenvalues ​​extracted by the PSD method, and is significantly effective in improving the accuracy. After selecting leads for a subject, only the signals of these leads can be taken in subsequent experiments, making the preparation of the experiment easier.

[0030] Furthermore, in the feature projection matrix establishment module, the empirical mode decomposition method is first used to decompose the data according to its own time scale characteristics, so that the data is decomposed into components that form the composite signal; then the wavelength optimal spatial filtering process is performed on these components to make the classification characteristics more obvious; then the common space mode feature extraction method is used to calculate the most suitable projection matrix for these data, and the data is feature extracted through the projection matrix to obtain the eigenvector of the CSP method, and the projection matrix is ​​saved. The projection matrix will be used for the signal conversion module to extract the signal. Then the power spectral density method is used to extract the signal obtained by the signal preprocessing module, and the eigenvalues ​​extracted by the two methods are combined to obtain the final eigenvector.

[0031] Furthermore, the classifier modeling module uses a linear discriminant analysis algorithm to establish a classifier data model for the final feature vector.

[0032] The system is based on a brain control implementation method of a non-invasive brain-computer interface and includes two parts: the first part is the training part, and the second part is the practical operation part;

[0033] The training part includes the following:

[0034] First, the signal acquisition module collects the signal, and the signal preprocessing module cuts and classifies the collected signal, and then uses a 5th-order Butterworth filter to filter the signal to obtain an 8-30Hz signal. Then, the power spectrum density of the signal at 8-30Hz is judged by the Fisher criterion function to measure and sort the separability of each electrode signal, and finally select the 15 electrode signals with the strongest separability;

[0035] In the feature projection matrix establishment module, the empirical mode decomposition method is first used to decompose the data according to its own time scale characteristics, so that the data is decomposed into components that form the composite signal; then these components are subjected to wavelength optimal spatial filtering to make the classification characteristics more obvious; then the common space mode feature extraction method is used to calculate the most suitable projection matrix for these data, and the data is feature extracted through the projection matrix to obtain the eigenvector of the CSP method, and the projection matrix is ​​saved. Then the power spectral density method is used to extract features from the signal obtained by the signal preprocessing module, and the eigenvalues ​​extracted by the two methods are combined to obtain the final eigenvector;

[0036] The final feature vector is used to establish a classifier data model using the linear discriminant analysis algorithm;

[0037] The practical operation part includes the following contents:

[0038] The read signal is preprocessed, and the electrode signal selected in the training part is used. Then, the average power spectral density of the frequency bands of the Mu and Beta waves of these 15 channels is taken to form the eigenvector of PSD. Then, the selected electrode signal data is subjected to empirical mode decomposition and optimal wavelength filtering. Then, the projection matrix obtained in the training part is used for feature extraction to obtain the eigenvector of CSP. Finally, the eigenvalues ​​extracted by the two methods are combined to obtain the final eigenvector.

[0039] The final feature vector is judged by the classifier data model obtained in the training part according to the imagination category and converted into control instructions, and the obtained control instructions are transmitted to various control targets to achieve the desired conversion of brain imagination into control commands of the present invention.

[0040] The beneficial effects of the present invention are:

[0041] Aiming at brain-computer interface control, the present invention proposes a new brain-computer interface control system and method, which adds a new scheme for the control of brain-computer interface. The brain's motor imagination signals can be converted into control instructions to control and manipulate the outside world, and the accuracy of EEG signal discrimination is greatly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a structural block diagram of the first system and method of the present invention;

[0043] Figure 2 is a structural block diagram of a second system and method of the present invention;

[0044] Figure 3 This is a lead diagram for the brain-computer interface;

[0045] Figure 4 Flowchart for obtaining feature projection matrix;

[0046] Figure 5 For experimental cases. DETAILED DESCRIPTION

[0047] The present invention is further technically described below in conjunction with the accompanying drawings, but the specific implementation method is only used to explain the present invention and does not limit the protection scope of the present invention.

[0048] Embodiment 1: One aspect of this embodiment provides a brain control system based on a non-invasive brain-computer interface, such as Figure 1 As shown, including:

[0049] Signal acquisition module, providing brain imagination induction and cerebral cortex electrical signal acquisition;

[0050] The signal preprocessing module cuts the collected electrical signals and eliminates useless signals. Then, a 5th-order Butterworth filter is used to filter the signals to obtain 8-30Hz signals. The power spectrum density of the signals is calculated using the Fisher criterion theory to select channels with high discrimination.

[0051] The feature extraction module uses the improved power spectrum density method to extract features from the processed signal and obtain the eigenvalue;

[0052] A classifier modeling module uses feature values ​​for classifier modeling;

[0053] The signal conversion module obtains control instructions after preprocessing, feature extraction and classification of real-time EEG signals.

[0054] Another aspect of this embodiment is to provide a method for implementing brain control based on a non-invasive brain-computer interface, such as Figure 1 As shown, it is divided into two parts. The first part is the training part, which includes the signal preprocessing module, the improved PSD feature extraction module and the classifier modeling module. The second part is the practical operation part, which includes the signal preprocessing module, the improved PSD feature extraction, the classification discrimination and the instruction output module.

[0055] In the visual trigger module, the present invention adopts the motion imagery visual trigger, a direction arrow pops up on the computer screen, when it is to the left, imagine the left hand moving, when it is to the right, imagine the right hand moving. At the same time, the g.tec device starts to collect data.

[0056] When the brain-computer interface starts to collect data, the present invention uses a 32-lead interface, such as Figure 3 As shown, another earlobe signal is read as a reference signal. When the picture is displayed, the subject starts to imagine, and the brain-computer interface also starts to read the electrical signals of the cerebral cortex; when the picture disappears, the subject stops imagining and starts to rest, and the brain-computer interface also stops reading the EEG signals and inputs the read electrical signals into the preprocessing module.

[0057] There are two types of operations in the preprocessing module. The first is when training, the read signal is clipped, classified, and filtered. The 5th-order Butterworth filter is used to filter the signal to obtain an 8-30Hz signal. The power spectrum density of the signal is calculated using the Fisher criterion theory to select channels with high discrimination. The second is when practicing, only clipping and filtering are required, and no classification is required. The channel selection is the channel selected during training.

[0058] The idea of ​​Fisher criterion (the larger the difference between the mean values ​​between classes, the better, and the smaller the variance within classes, the better) is used to measure and sort the power spectral density separability of each channel signal between 8-30Hz, select the 15 electrode signals with the strongest separability, and then use the improved PSD feature extraction module to extract PSD features.

[0059] The method of sorting the separability of each electrode signal using the Fisher criterion is as follows:

[0060] First, calculate the average power spectral density of each channel in the 8-30Hz frequency band for each trial. For the signal Xij(t) of the jth channel in the ith trial, calculate its power spectral density and obtain the power spectral density values ​​P of m different frequency bands. ij (u), among these m power spectrum density values, take the power spectrum density value P in the frequency band range of 8-30Hz ij (j)(where f∈{1, 2, 3, ..., n}), for P ij (f) Taking the average value, we get Should It is the average value of the power spectral density of the jth channel signal in the i-th trial, representing the energy characteristics of the channel in the range of 8-30 Hz.

[0061]

[0062] Then, the Fisher criterion value of each lead channel is obtained. First, the energy characteristics of each channel of all trials C1 in the first category are averaged to obtain m1(j), and the energy characteristics of each channel of all trials C2 in the second category are averaged to obtain m2(j).

[0063]

[0064]

[0065] Then, the variance of each channel energy feature of all trials C1 in the first category is calculated to obtain V1(j), and the variance of each channel energy feature of all trials C2 in the second category is calculated to obtain V2(j)

[0066]

[0067]

[0068] Finally, the Fisher criterion value is calculated for each channel

[0069]

[0070] Then, the criterion value F(j) of each channel is sorted, and the first 15 channel port signals with the largest criterion value are selected.

[0071] The method of PSD feature extraction is then as follows:

[0072] For the signal Xij(t) of the jth channel in the i-th trial, the power spectrum density can be obtained to obtain the power spectrum density values ​​P of m different frequency bands. ij (u), in the P band of Beta wave (i.e. 8-12Hz band) ij (u) Take the average value to get Beta ij , in the P band where the Mu wave (i.e. 8-12Hz band) is located ij (u) Take the average value to get Mu ij .

[0073] Because the brain waves related to motor imagery are mainly Beta waves and Mu waves, the average power spectral density of these two waves can well represent the characteristics of the imagination signal. In addition, when selecting the channel port, the channel port with the highest separability of the power spectral density of the frequency band where the two waveforms are located is selected, so that the average power spectral density of these two waves can better distinguish the two types of signals, thereby achieving a good classification effect.

[0074] Then repeat the above process for all 15 channels of the ith trial, and get 30 average values ​​of power spectrum density. These 30 values ​​can well represent the characteristics of the trial, and can also have a good degree of distinction in the classification of two types of signals. Therefore, these 30 values ​​form the feature vector of the ith trial, and so on, the feature vectors of all trials can be obtained. The feature extraction step of the signal is completed. And the obtained feature values ​​are added to the classifier modeling module.

[0075] The next step is to calculate the classifier model parameters, as follows:

[0076] First, the feature vectors of category one and category two are averaged to obtain average vectors m1 and m2, and then the average of all feature vectors is obtained to obtain m3.

[0077] Then, subtract m1 from the feature vectors of all trials in class 1 to obtain vector S[i]. Subtract m2 from the feature vectors of all trials in class 2 to obtain vector S[i].

[0078] Perform the following processing on all S[i]:

[0079] T[i]=S[i] T S[i]

[0080] The sum of all T[i] obtained is used to obtain Sw.

[0081] Subtract m3 from the eigenvectors of all trials to get P[j], and perform the following processing on all P[j]:

[0082] H[j]=P[j] T P[j]

[0083] Sum all the H[i] obtained to get St. And perform the following operations on Sw and St:

[0084] G=Sw -1 *St

[0085] Next, the singular value decomposition of G is performed to obtain the eigenvector matrix V. The first column of V is taken and transposed to obtain the vector WT. Finally, WT, m1, and m2 are saved as the model parameters of the classifier. At this time, the classifier model parameters are obtained.

[0086] The first part is over. Now let's talk about the second part. We will select the same channels as in the training for the preprocessed data and use the PSD feature extraction method consistent with that in the training to obtain the feature vector x. Then we will classify and discriminate the feature vector. The discrimination method is as follows:

[0087] First, find mu1, mu2

[0088] mu1=WT*m1 T

[0089] mu2=WT*m2 T

[0090] Next, find y

[0091] y=WT*x

[0092] Then find the discriminant numbers d1, d2

[0093] d1=|y-mu1|

[0094] d2=|y-mu2|

[0095] If d1 is smaller than d2, then the trial is of category 2, otherwise it is of category 1.

[0096] Finally, the category judgment is converted into a control instruction and input into the device to be controlled. Figure 5 As shown in the figure, the rehabilitation robotic arm is controlled by EEG signals.

[0097] Embodiment 2: One aspect of this embodiment provides a brain control system based on a non-invasive brain-computer interface, such as Figure 2 As shown, including:

[0098] Signal acquisition module, providing brain imagination induction and cerebral cortex electrical signal acquisition;

[0099] The signal preprocessing module cuts the collected electrical signals and then filters them with a 5th-order Butterworth filter to obtain an 8-30Hz signal. The power spectrum density of the signal is then judged by the Fisher criterion function to measure and sort the separability of each electrode signal, and finally the 15 electrodes with the strongest separability are selected.

[0100] The feature projection matrix building module first uses the processed signal to build the feature projection matrix, and extracts the features of the signal through the projection matrix. Then, the signal is extracted using the power spectrum density method, and the eigenvalues ​​extracted by the two methods are combined to obtain the final eigenvalue.

[0101] Classifier modeling module, which uses the final feature value of the previous step for classifier modeling;

[0102] The signal conversion module obtains control instructions after preprocessing, feature extraction and classification of real-time EEG signals.

[0103] In the feature projection matrix establishment module, an improved CSP algorithm is used, and the method is as follows: first, the empirical mode decomposition method is used to decompose the data according to its own time scale characteristics, so that the data is decomposed into components that form the composite signal; then, these components are subjected to wavelength optimal spatial filtering to make the classification characteristics more obvious; then, the common space mode feature extraction method is used to calculate the most suitable projection matrix for these data, and the data is feature extracted through the projection matrix to obtain the feature vector of the improved CSP method, and the projection matrix is ​​saved. The projection matrix will be used in the signal conversion module to extract features from the signal. Then, the power spectral density method is used to extract features from the signal obtained by the signal preprocessing module, and the feature values ​​extracted by the two methods are combined to obtain the final feature vector;

[0104] The classifier modeling module uses the linear discriminant analysis algorithm to build a classifier data model for the final feature vector.

[0105] Another aspect of this embodiment is to provide a method for implementing brain control based on a non-invasive brain-computer interface, such as Figure 2 As shown, it is divided into two parts. The first part is the training part, which includes the signal preprocessing module, the feature projection matrix establishment module and the classifier modeling module. The second part is the practical operation part, which includes the signal preprocessing module, feature extraction, classification discrimination and instruction output module.

[0106] In the visual trigger module, the present invention adopts the motion imagery visual trigger, a direction arrow pops up on the computer screen, when it is to the left, imagine the left hand moving, when it is to the right, imagine the right hand moving. At the same time, the g.tec device starts to collect data.

[0107] When the brain-computer interface starts to collect data, the present invention uses a 32-lead interface, such as Figure 3 As shown, another earlobe signal is read as a reference signal. When the picture is displayed, the subject starts to imagine, and the brain-computer interface also starts to read the electrical signals of the cerebral cortex; when the picture disappears, the subject stops imagining and starts to rest, and the brain-computer interface also stops reading the EEG signals and inputs the read electrical signals into the preprocessing module.

[0108] There are two types of operations in the preprocessing module. The first is to crop, classify, and filter the read signal during training. A 5th-order Butterworth filter is used to filter the signal to obtain an 8-30Hz signal. The power spectrum density of the signal is then judged by the Fisher criterion function to measure and sort the separability of each lead signal. Finally, the 15 most separable lead signals are selected, and the data is input into the next module after all training is completed. The second is during practical operation. At this time, only cropping and filtering are required, and no classification is required. The optimal lead calculated during training is selected and these lead signals are input into the next module. The method of sorting the separability of each electrode signal using the Fisher criterion is the same as system 1.

[0109] The projection matrix is ​​established by using an improved CSP algorithm, which consists of empirical mode decomposition (EMD), wavelength optimal spatial filtering and common space pattern feature extraction (CSP), and the feature matrix is ​​obtained from it. The flow chart is as follows: Figure 4 shown.

[0110] The steps of Empirical Mode Decomposition (EMD) are as follows:

[0111] First, all the maximum values ​​and all the minimum values ​​are extracted from the preprocessed signal u(t), and all the maximum values ​​are fitted into the upper envelope y1(t), and all the minimum values ​​are fitted into the lower envelope y2(t).

[0112] Then find the average curve n1(t) of the upper and lower envelopes.

[0113]

[0114] Subtract the average signal n1(t) from the signal u(t) to obtain the signal h1(t). Determine whether h1(t) meets the conditions of the IMF component. If not, replace u(t) with h1(t) and repeat the above process until hn(t) meets the conditions of the IMF and assign hn(t) to IMF1(t).

[0115] Then, the signal r(t) is obtained by subtracting IMF1(t) from the signal u(t), and r(t) is used to replace the signal u(t) obtained after preprocessing, and the above process is repeated to obtain multiple IMF signals until r(t) is a monotonic function or a constant. However, here, the present invention only seeks two. Because the amount of data is large, it is quite time-consuming to seek these signals, so only the first two IMF signals are obtained and named x to replace the original u(t) signal to continue the next steps.

[0116] Then, these signals are subjected to wavelength-optimal spatial filtering, and the value of each channel port at the next moment is subtracted from the value at the previous moment. The method is as follows:

[0117] x=x(2:n,:)-x(1:n-1,:)

[0118] The next step is to extract the common spatial pattern features. First, each trial data X[i] is subjected to covariance normalization to obtain R[i], and the formula is as follows:

[0119]

[0120] Next, we calculate the average value of R according to the category to which the trial belongs. Taking two categories of data as an example (if it is multi-category data, it can be divided into multiple binary classifications), the average values ​​of the two categories are Ra and Rb respectively. Next, we add the two average values ​​to get the mixed space covariance matrix Rc, and perform eigenvalue decomposition on Rc to get the eigenvector and eigenvalue.

[0121] Rc=PλP T

[0122] Where P is the eigenvector, λ is the eigenvalue, and then the whitening matrix W is calculated as follows

[0123] W=λ -0.5 P T

[0124] Whiten Ra to get Sa

[0125] Sa=W*Ra*W T

[0126] Perform eigendecomposition on Sa to obtain the eigenvector V

[0127] The final projection matrix Projm is obtained as follows:

[0128] Projm=W T V

[0129] Next, the training data is converted into feature vectors using the projection matrix and the projection matrix is ​​saved. Then, the power spectral density method is used to extract features from the signal obtained by the preprocessing module. The specific method has been described in detail in Example 1. Finally, the eigenvalues ​​extracted by the two methods are combined to obtain the final feature vector.

[0130] The next step is to calculate the classifier model parameters. The mathematical model of the classifier is established through the final feature vectors of category one and category two, and WT, m1, m2 are obtained and saved as the model parameters of the classifier. The specific method is described in detail in system one.

[0131] The first part is over. Now let's talk about the second part. The preprocessed data is first processed according to the empirical mode decomposition and wavelength optimal spatial filtering in part one, and then projected through the projection matrix obtained during training to obtain the eigenvector of CSP. Then the power spectral density method is used to extract the features of the preprocessed signal to obtain the eigenvector of PSD. Then the two eigenvectors are combined to obtain the eigenvector x, and then the eigenvector is classified and judged. The judgment method is as follows:

[0132] First, find mu1, mu2

[0133] mu1=WT*m1 T

[0134] mu2=WT*m2 T

[0135] Next, find y

[0136] y=WT*x

[0137] Then find the discriminant numbers d1, d2

[0138] d1=|y-mu1|

[0139] d2=|y-mu2|

[0140] If d1 is smaller than d2, then the trial is of category 2, otherwise it is of category 1.

[0141] Finally, the category judgment is converted into a control instruction and input into the device to be controlled. Figure 5 As shown in the figure, the rehabilitation robotic arm is controlled by EEG signals.

[0142] The above detailed explanation is only an application of the present invention and cannot be used to limit the scope of rights of the present invention. Any modifications made within the spirit and principles of the present invention are within the protection scope of the present invention.

Claims

1. A brain control system based on a non-invasive brain-computer interface, characterized in that: It includes signal acquisition module, signal preprocessing module, feature projection matrix building module, classifier modeling module and signal conversion module; The signal acquisition module provides brain imagination induction and electrical signal acquisition of the cerebral cortex; The signal preprocessing module cuts the collected electrical signals and eliminates useless signals. It then uses a 5th-order Butterworth filter to filter the signals to obtain 8-30Hz signals. It then uses the Fisher criterion function to discriminate the power spectrum density of the signal frequency between 8-30Hz, measures the separability of each electrode signal and sorts it, and finally selects the 15 electrode signals with the strongest separability. The feature projection matrix building module first uses the empirical mode decomposition method to decompose the signal according to its own time scale characteristics to form component signals; then the component signals are subjected to wavelength optimal spatial filtering to make the classification characteristics more obvious; Then, the common space mode feature extraction method is used to calculate the most suitable projection matrix, and the feature extraction is performed through the projection matrix to obtain the feature vector I of the common space mode feature extraction method, and the projection matrix is ​​saved; then, the power spectral density method is used to extract the PSD features of the 15 electrode signals obtained by the signal preprocessing module, and the average power spectral density of the frequency bands of the Mu and Beta waves of the channels where the above 15 electrodes are located is taken as the eigenvalue, and the eigenvector II is formed, and the above eigenvector I and eigenvector II are merged to obtain the final eigenvector; The classifier modeling module uses the final feature vector obtained by the feature projection matrix building module for classifier modeling; The signal conversion module obtains control instructions after preprocessing, feature extraction and classification of real-time EEG signals.

2. The brain control system based on non-invasive brain-computer interface according to claim 1, characterized in that: In the signal acquisition module, the imagination induction provided to the brain is in the form of visual induction, providing visual manipulation instructions to induce the organism to actively imagine; the electrical signal acquisition of the cerebral cortex adopts a non-invasive brain-computer interface, using 32 electrodes to collect cerebral cortex signals, and then collecting earlobe signals as reference electrodes.

3. The brain control system based on non-invasive brain-computer interface according to claim 1, characterized in that: The classifier modeling module uses a linear discriminant analysis algorithm to establish a classifier data model for the final feature vector.

4. A method for implementing brain control based on a non-invasive brain-computer interface, characterized in that: It consists of two parts: the first part is the training part, and the second part is the practical operation part; The training part includes the following: First, the signal acquisition module collects the signal, and the signal preprocessing module cuts and classifies the collected signal, and then uses the 5th-order Butterworth filter to filter the signal to obtain the 8-30Hz signal; then the power spectrum density of the signal at 8-30Hz is judged by the Fisher criterion function to measure and sort the separability of each electrode signal, and finally select the 15 electrode signals with the strongest separability; In the feature projection matrix building module, the empirical mode decomposition method is first used to decompose the signal according to its own time scale characteristics to form component signals; then the component signals are subjected to wavelength optimal spatial filtering to make the classification characteristics more obvious; Then, the common space mode feature extraction method is used to calculate the most suitable projection matrix, and feature extraction is performed through the projection matrix to obtain the feature vector I of the common space mode feature extraction method, and the projection matrix is ​​saved; then, the power spectral density method is used to extract the PSD features of the 15 electrode signals obtained by the signal preprocessing module, and the average power spectral density of the frequency bands of the Mu and Beta waves of the channels where the above 15 electrodes are located is taken as the eigenvalue, and the feature vector II is formed, and the above eigenvector I and eigenvector II are merged to obtain the final eigenvector; The final feature vector is used to establish a classifier data model using the linear discriminant analysis algorithm; The practical operation part includes the following contents: The read signal is preprocessed, and the 15 electrodes selected in the training part are used. Then, the average power spectral density of the frequency bands of the Mu and Beta waves of the channels where the 15 electrodes are located is taken to form the eigenvector III of the PSD. Then, the signal data obtained from the selected 15 electrodes are subjected to empirical mode decomposition and optimal wavelength filtering. The projection matrix obtained in the training part is used for feature extraction to obtain the eigenvector IV of the CSP. Finally, the above eigenvector III and eigenvector IV are merged to obtain the final eigenvector. The final eigenvector is judged by the classifier data model obtained in the training part for imagination category and converted into control instructions.

Images