CN101558997A - EEG signal recognition method based on second-order blind recognition - Google Patents

Abstract

The invention pertains to the technical field of biomedical engineering and information, and realizes individual identity authentication and identification through acquiring and analyzing the electroencephalograph (EEG) signals in human brain. Experimenters are trained through different stimulation patterns, thus leading the experimenters to adapt to the different stimulation patterns and generate different EEG signals; the EEG signals generated due to the stimulation are analyzed through second-order blind identification; individual characteristic signals are extracted for classification and identification so as to determine the target characteristics; and finally the train is finished. When in identification, only the target characteristics of the acquired EEG signals are extracted for classification and compared with the formwork of each experimenter, the identification result can be determined. Research results show that the highest identification rate is 88 percent, the average identification rate is about 83 percent, and the method can realize individual identification and identity authentication.

Description

EEG signal recognition method based on second-order blind recognition

technical field

The technical invention belongs to the fields of biomedical engineering and information technology.

technical background:

Identification and verification are important prerequisites for ensuring national public safety and information security. In national security, public security, justice, e-commerce, e-government, security inspection, security monitoring and other application fields, accurate identification and identification are required. The traditional verification method for identification items (such as keys, certificates, bank cards); the other is a verification method based on identification knowledge (such as user name, password, etc.). However, marked items are easily lost or counterfeited, and marked knowledge is easily forgotten or deciphered. Biometric identification technology has brought the possibility of realizing this wish. People may forget or lose their cards or passwords, but they will never forget or lose their biometrics, such as faces, fingerprints, irises, palm prints, brain waves, etc. Therefore, the personal identification system based on biometric identification technology has better security, reliability and effectiveness, and is attracting more and more attention, and has begun to enter various fields of our social life to meet the challenges of the new era. Since the late 1980s and early 1990s, with the increasing importance of information security, biometric technology research has become a research hotspot. Biometrics refers to the use of the inherent physiological characteristics of the human body (such as fingerprints, faces, iris, brain waves, pulse, etc.) ) or behavioral characteristics (such as handwriting, voice, gait, etc.) for personal identity authentication. Biometric identification technology has the advantages of not being forgotten, not easy to be forged or stolen, portable and available anytime, anywhere, etc. It is more secure, confidential and convenient than traditional identity authentication methods. Biometrics that can be used to identify and authenticate identities should have the characteristics of universality, uniqueness, stability, and collectability. At present, several biometric identification technologies that are relatively mature and have the most application prospects include fingerprints, faces, face thermograms, irises, retinas, hand shapes, voiceprints, and signatures. Among them, iris recognition and fingerprint recognition are recognized as the two most reliable biometric technologies.

Any physiological or behavioral characteristic of a person can be used as a biological characteristic for identification in principle as long as it meets the following conditions: (1) universal, everyone has it; (2) unique, everyone is different; (3) ) stability, which is constant in a certain period of time; (4) collectability, which can be conveniently measured quantitatively. Of course, it may not be feasible to just meet the above conditions. The actual system should also consider: (1) performance, that is, the recognition accuracy, speed, robustness and resources needed to meet the requirements; (2) acceptability, people Acceptance of this kind of biometric identification; (3) Deceptiveness, whether it is easy to deceive the system through subjective fraud.

Currently commonly used biometric technologies have one or another problem, for example, face recognition is powerless for twins; voiceprint recognition is easy to imitate; fingerprint recognition will be affected by finger injuries, and it is also easy to steal. Electroencephalogram (EEG) signals are not only a very useful clinical diagnostic tool, but also a good biometric tool for identity authentication. First of all, it is universal, everyone has brain waves; secondly, there are different EEG signals between people due to the differences in brain characteristics, thinking methods, memory, etc. of each person; thirdly, EEG also has a certain degree of stability , within a certain period of time, the EEG signal can maintain relative stability, and finally, the EEG signal is easy to collect. The biometric system based on EEG signal can achieve a certain accuracy and speed, and it will not cause any harm to the human body, which is also acceptable to people. Because the EEG signal comes from the thinking activity of the brain, it is difficult to forge, and the robustness of the system is very high. The research and analysis of the EEG signals of the human brain shows that different individuals will produce different nerve impulse responses in different brain regions. , which can make the EEG signal have individual specificity. Based on the above analysis, the biometric identification system based on EEG is a new identification system with application prospects.

Contents of the invention

The present invention uses the classification of motor imagery EEG signals to realize the identification of the subject, only uses the electrode signals related to motor imagery for data analysis, and adopts the signal processing and characteristics that are very effective for motor imagery EEG signals after research. Extraction method - second-order blind identification and Fisher distance to extract features, and use neural network to classify features, so as to realize identity recognition. The method is suitable for various groups of people including physical disabilities and visual defects, and has good applicability.

The present invention comprises the following steps:

Step 1. The subject puts on the electrode cap. The original EEG signal is collected through a 64-channel EEG amplifier conforming to the 10/20 method calibrated by the International Electroencephalography Society. The sampling rate is 250Hz, and the left mastoid is taken as the The reference electrode, the passband of the band-pass filter is 1-50Hz, select 6 electrodes to collect EEG signals (that is, C3, C4, P3, P4, O1 and O2 in the 10/20 international standard calibrated by the International Electroencephalography Society A total of 6 electrode positions), to collect the EEG signals of the subjects in different motor imagery processes.

Step 2. According to the set stimulus program on the computer screen (prompting the subject to start imagining the movement), the subject made four different types of motor imaginations (imagining left hand movement, right hand movement, leg movement and movement) according to the experimental requirements. tongue movement). The subjects were trained and familiar with the experimental procedure.

Because each person has different adaptability to different motor imagery, we let the subjects try four different types of motor imagery during the learning process. Through learning and testing, we can decide which type of motor imagery is most suitable. For example, after the subject passed the study and test, we found that the recognition rate of imagining tongue movements was higher than that of the other three types of motor imagery, we thought that this subject might be more suitable for imagining tongue movements. In future use, you only need to imagine the tongue movement, and you don't need to imagine other movements.

Step 3. Preprocessing the collected EEG signals. Firstly, the obtained EEG signals are screened to exclude some obviously abnormal EEG signals, and then the rest of the EEG signals are preprocessed by oculogram removal, artifact removal, baseline correction, and linear correction. EEG signals will be further signal processed.

Step 4, EEG signal processing. The preprocessed EEG signal still contains a lot of irrelevant evoked potentials and background noise, so the spatial resolution and signal-to-noise ratio are very low. In order to identify useful signals from the background noise, further processing is necessary. The present invention uses a Blind source separation algorithm - second-order blind identification to further process EEG signals, the specific algorithm is as follows:

Let the n column vectors of x(t) correspond to the continuous-time EEG signals of n electrodes, then x _i (t) corresponds to the EEG signals of the i-th electrode. Each x _i (t) can be regarded as a linear instantaneous mixture of n sources s _i (t), and the mixing matrix is A, then

x(t)=A _s (t) (1)

SOBI only uses the EEG signal x(t) measured by the sensor to obtain an approximate A ^-1 decomposition matrix W, so that

$\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; wxya</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

is the recovered continuous-time source signal.

The SOBI algorithm has two steps: first, the EEG signal is zero-meanized, as shown in the following formula:

y(t)=B(x(t)-<x(t)>) (3)

Angle brackets <·> indicate time averaging, so the mean of y is zero. The value of matrix B is such that the correlation matrix <y(t)y(t) ^T > of y is the identity matrix, whose value is given by

$\mathrm{<span\; class="notranslate">\; B</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; diag</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; u</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

Where λ _i is the eigenvalue of the correlation matrix <(x(t)-<x(t)>)(x(t)-<x(t)>) ^T >, and each column of U is its corresponding eigenvector .

The second step is to construct a set of diagonal matrices: select a set of time delays τ, and calculate the symmetric correlation matrix of the signal y(t) and its time-delayed signal y(t+τ):

R _τ = sym(<y(t)y(t+τ) ^T >) (5)

in

sym(M)=(M+M ^T )/2 (6)

This is a function that turns an asymmetric matrix into a related symmetric matrix. The process of symmetrization loses some information, but it provides an effective solution.

After calculating Rτ, then diagonalize Rτ: through the rotation matrix V, use the iterative method, so that

∑ _τ ∑ _i≠j (V ^T R _τ V) _ij ² (7)

obtains a minimum value, the estimate of the separation matrix

W = V ^T B (8)

Step 5. Feature extraction

In order to perform classification, feature extraction is first performed. Feature extraction is to extract data points that can distinguish sample types from all data, that is, feature points. The present invention uses Fisher distance to determine features. In classification research, Fisher distance is often used to represent the difference between types. The size of Fisher distance is proportional to the degree of discrimination between types. If the degree of discrimination between types is large, that is, the difference is obvious, the Fisher distance is larger. Otherwise, Fisher distance is smaller.

The formula for calculating the Fisher distance between two classes is as follows:

$\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; =</span>\frac{{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 9</span>}<span\; class="notranslate">\; )</span>$

Among them, F represents the Fisher distance, μ and σ are the mean and variance, respectively, and the subscripts 1 and 2 represent two different classes respectively.

For three or more types of situations, the Fisher distance formula can be extended as follows:

$\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; =</span>\frac{{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; 3</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; +</span>{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 10</span>}<span\; class="notranslate">\; )</span>$

For each data point, the size of the Fisher distance represents the contribution of the data point as a feature to the classification. The point with a larger Fisher distance has a greater contribution to the classification.

The number of feature points is closely related to the final recognition rate, algorithm complexity and recognition rate. Too many or too few feature points will affect the recognition rate and reduce the recognition rate. In addition, the more feature points, the more complex the algorithm and the faster the recognition. On the contrary, the fewer feature points, the simpler the algorithm and the faster the recognition. Therefore, the number of feature points has a great influence on the final performance. After repeated tests, it is found that the feature extraction uses the Fisher distance to determine the feature, extracts 8-12 feature points for each electrode data, and a total of 48--72 feature points; it can achieve a higher recognition rate, and the recognition speed is also faster. Not slow, the algorithm complexity is average.

Step 6. Use the BP neural network for classification learning and testing. The BP neural network has 60 units in the input layer, 10 units in the hidden layer, and 1 unit in the output layer. Use the above 60 features as the input layer of the BP neural network, and input the features extracted from the EEG signal of each subject’s motor imagery through step 5 to the input layer, and each subject has 20 data during the learning process (five for each of the four types of motor imagery), we determined the parameters of the neural network through the learning process. The test process also has 20 data, through the test process, we can determine the type of motor imagery suitable for the subject (the recognition rate is the highest).

Step 7. Input unknown EEG data into the neural network for identification and authentication. After the subject has passed the above steps 1-6, the BP neural network structure and the type of motor imagery suitable for him (her) are determined, and identification and authentication can be performed at this time. The subject puts on the electrode cap, starts motor imagery according to step 2 (just the most suitable type of motor imagery determined in step 6), collects EEG signals, and after preprocessing, extracts 60 EEG signals according to the algorithm introduced in step 5. Feature quantity, input the extracted feature quantity into the neural network determined in step 6. If it is identification, the neural network outputs the code of the subject; if it is authentication, the neural network output changes whether the subject (0 or 1).

The identification described in the present invention refers to selecting which subject this EEG signal belongs to from several subjects; and the authentication process is to determine whether this EEG signal belongs to a certain subject, The former is a multiple-choice question, while the latter is a true-false question.

The present invention uses EEG signals to extract information features of EEG signals. The method used is to imagine various movement patterns through the brain, and perform feature extraction and classification on them, so as to realize the identification of individual identities through EEG signals. The process of performing identification or authentication. EEG signals are used as identification to provide a new type of password system, which can not only solve the problem that some disabled people cannot complete daily identification, but also can be used in occasions that have higher requirements for identification.

The innovations of this method are:

1. EEG signals are used as input signals for identification, which is different from previous fingerprints and irises.

2. The EEG signals based on motor imagery were collected, that is, the EEG signals generated by the subjects when they imagined four kinds of sports, of course, it is also applicable to other EEG signals (such as visual evoked potentials, event evoked potentials, etc.).

3. Second-order blind recognition and Fisher distance are used to extract information from EEG signals.

4. At the same time, the identification and authentication functions are realized. Recognition refers to judging whose EEG signal is from the EEG signals of several individuals, and authentication refers to judging whether a certain EEG signal is the EEG signal of the target person.

5. According to the characteristics of different subjects, automatically select the type of motor imagery suitable for the subjects.

Description of drawings

The schematic diagram of electrode selection in Fig. 1,

Figure 2 is a flow chart of the feature extraction of the identity recognition system based on EEG signals,

Figure 3 is a flow chart of identity recognition based on EEG signals.

Detailed ways

The method of the present invention is used to realize the identification of the individual identity in the EEG signal identification system, according to accompanying drawings 1, 2, and 3. This can be achieved by following steps:

Step 1. The subject puts on the electrode cap. The original EEG signal is collected through a 64-channel EEG amplifier conforming to the 10/20 method calibrated by the International Electroencephalography Society. The sampling rate is 250Hz, and the left mastoid is taken as the The reference electrode, the passband of the band-pass filter is 1-50 Hz, and 6 electrodes are selected to collect EEG signals. The specific electrode positions are shown in Figure 1. That is, the 6 electrode positions of C3, C4, P3, P4, O1 and O2 in the 10/20 international standard calibrated by the International Electroencephalography Society collect the EEG signals of subjects with different motor imagery processes.

Step 2. According to the set stimulus program on the computer screen (prompting the subject to start imagining the movement), the subject made four different types of motor imaginations (imagining left hand movement, right hand movement, leg movement and movement) according to the experimental requirements. tongue movement). The subjects were trained and familiar with the experimental procedure.

Step 3. Preprocessing the collected EEG signals. Firstly, the obtained EEG signals are screened to exclude some obviously abnormal EEG signals, and then the rest of the EEG signals are preprocessed by oculogram removal, artifact removal, baseline correction, and linear correction. EEG signals will be further signal processed.

Step 4, EEG signal processing. The preprocessed EEG signal still contains a lot of irrelevant evoked potentials and background noise, so the spatial resolution and signal-to-noise ratio are very low. In order to identify useful signals from the background noise, further processing is necessary. The present invention uses a Blind source separation algorithm - second-order blind identification to further process EEG signals, the specific algorithm is as follows:

Let the n column vectors of x(t) correspond to the continuous-time EEG signals of n electrodes, then x _i (t) corresponds to the EEG signals of the i-th electrode. Each x _i (t) can be regarded as a linear instantaneous mixture of n sources s _i (t), and the mixing matrix is A, then

x(t)=As(t) (1)

SOBI only uses the EEG signal x(t) measured by the sensor to obtain an approximate A ^-1 decomposition matrix W, so that

$\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; wxya</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

is the recovered continuous-time source signal.

The SOBI algorithm has two steps: first, the EEG signal is zero-meanized, as shown in the following formula:

y(t)=B(x(t)-<x(t)>) (3)

Angle brackets <·> indicate time averaging, so the mean of y is zero. The value of matrix B is such that the correlation matrix <y(t)y(t) ^T > of y is the identity matrix, whose value is given by

$\mathrm{<span\; class="notranslate">\; B</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; diag</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; u</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

Where λ _i is the eigenvalue of the correlation matrix <(x(t)-<x(t)>)(x(t)-<x(t)>) ^T >, and each column of U is its corresponding eigenvector .

The second step is to construct a set of diagonal matrices: select a set of time delays τ, and calculate the symmetric correlation matrix of the signal y(t) and its time-delayed signal y(t+τ):

R _τ = sym(<y(t)y(t+τ) ^T >) (5)

in

sym(M)=(M+M ^T )/2 (6)

This is a function that turns an asymmetric matrix into a related symmetric matrix. The process of symmetrization loses some information, but it provides an effective solution.

After calculating Rτ, then diagonalize Rτ: through the rotation matrix V, use the iterative method, so that

∑ _τ ∑ _i≠j (V ^T R _τ V) _ij ² (7)

obtains a minimum value, the estimate of the separation matrix

W = V ^T B (8)

Step 5. Feature extraction

In order to perform classification, feature extraction is first performed. Feature extraction is to extract data points that can distinguish sample types from all data, that is, feature points. The present invention uses Fisher distance to determine features. In classification research, Fisher distance is often used to represent the difference between types. The size of Fisher distance is proportional to the degree of discrimination between types. If the degree of discrimination between types is large, that is, the difference is obvious, the Fisher distance is larger. Otherwise, Fisher distance is smaller.

The formula for calculating the Fisher distance between two classes is as follows:

$\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; =</span>\frac{{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 9</span>}<span\; class="notranslate">\; )</span>$

Among them, F represents the Fisher distance, μ and σ are the mean and variance, respectively, and the subscripts 1 and 2 represent two different classes respectively.

For three or more types of situations, the Fisher distance formula can be extended as follows:

$\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; =</span>\frac{{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; 3</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; +</span>{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 10</span>}<span\; class="notranslate">\; )</span>$

For each data point, the size of the Fisher distance represents the contribution of the data point as a feature to the classification. The point with a larger Fisher distance has a greater contribution to the classification.

The number of feature points is closely related to the final recognition rate, algorithm complexity and recognition rate. Too many or too few feature points will affect the recognition rate and reduce the recognition rate. In addition, the more feature points, the more complex the algorithm and the faster the recognition. On the contrary, the fewer feature points, the simpler the algorithm and the faster the recognition. Therefore, the number of feature points has a great influence on the final performance. After repeated tests, it is found that the feature extraction uses the Fisher distance to determine the feature, and extracts 10 feature points from the data of each electrode, a total of 60 feature points; it can achieve a high recognition rate, and the recognition speed is not slow, and the algorithm is complex Moderate.

Step 6. Extract the EEG signal features of each subject according to the process shown in Figure 2, learn and identify through the BP neural network, determine the neural network structure and motor imagery type, and the training process ends. The BP neural network has 60 units in the input layer, 10 units in the hidden layer, and 1 unit in the output layer. Use the above 60 features as the input layer of the BP neural network, and input the features extracted from the EEG signal of each subject’s motor imagery through step 5 to the input layer, and each subject has 20 data during the learning process (five for each of the four types of motor imagery), we determined the parameters of the neural network through the learning process. The test process also has 20 data, through the test process, we can determine the type of motor imagery suitable for the subject (the recognition rate is the highest).

Step 7. Identify and authenticate the identity of each subject according to the process shown in FIG. 3 . Input unknown EEG data into the neural network for identification and authentication. After the subject has passed the above steps 1-6, the BP neural network structure and the type of motor imagery suitable for him (her) are determined, and identification and authentication can be performed at this time. The subject puts on the electrode cap, starts motor imagery according to step 2 (just the most suitable type of motor imagery determined in step 6), collects EEG signals, and after preprocessing, extracts 60 EEG signals according to the algorithm introduced in step 5. Feature quantity, input the extracted feature quantity into the neural network determined in step 6. If it is identification, the neural network outputs the code of the subject; if it is authentication, the neural network output changes whether the subject (0 or 1).

At present, some foreign studies mainly use visual stimulation or myoelectricity as a feature source. This study uses motor imagery EEG signals as identification, which can be suitable for various groups of people such as disabilities, and has good adaptability. In terms of method, second-order blind identification is used for signal processing, and Fisher distance is used to extract features. It can be seen from the results that the recognition rate of imaginary tongue movement EEG signal is the highest, reaching 88.1%, and the average recognition rate of four kinds of imaginary movement EEG signals is 82.8%, which is about 5 percentage points higher than other foreign methods.

Claims (2)

1. The EEG signal recognition method based on second-order blind recognition is characterized in that: the present invention comprises the following steps: Step (1), the subject puts on the electrode cap, and the original EEG signal is collected through a 64-channel EEG amplifier that conforms to the 10/20 method calibrated by the International Electroencephalography Society. The sampling rate is 250Hz, and the left mastoid is used as a reference. Electrodes, the passband of the band-pass filter is 1-50Hz, select 6 electrodes to collect EEG signals, the 6 electrodes are C3, C4, P3, P4, O1 and O2 6 electrode positions to collect the EEG signals of subjects with different motor imagery processes; Step (2), according to the set stimulus program on the computer screen, prompt the subject to start imagining movement, the subject makes four different types of movement imagination according to the experimental requirements, imagine left hand movement, right hand movement, leg movement and tongue movement, the subjects were trained and familiar with the experimental process; Step (3), preprocessing the collected EEG signals, first screening the acquired EEG signals, excluding a part of the obviously abnormal EEG signals, and then removing the oculoelectricity and false removal of the remaining EEG signals Trace, baseline correction, linear correction preprocessing; Step (4), EEG signal processing, the preprocessed EEG signal still contains a lot of irrelevant evoked potentials and background noise, and the second-order blind identification in the blind source separation algorithm is used to further process the EEG signal; Step (5), feature extraction, adopt Fisher distance to determine feature, extract 8-12 feature points to the data of each electrode, totally 48--72 feature points; Step (6), using BP neural network to carry out classification learning and testing. We can identify the type of motor imagery that is appropriate for the subject; Step (7), inputting the unknown EEG data into the neural network for identification and authentication. After the subject has passed the above steps 1-6, the BP neural network structure and the suitable motor imagery type are determined. At this time, identification and authentication can be carried out. The subject puts on the electrode cap and starts motor imagery according to step 2. The most suitable type of motor imagery determined in step 6 is needed to collect EEG signals. After preprocessing, extract feature quantities according to the algorithm introduced in step 5, and input the extracted feature quantities into the neural network determined in step 6. If it is identification, the neural network outputs the code of the subject; if it is authentication, the neural network output is changed to whether the subject is. 2. The EEG signal recognition method based on second-order blind recognition as claimed in claim 1, characterized in that: Step (4), the specific algorithm of EEG signal processing is as follows: Let the n column vectors of x(t) correspond to the continuous time EEG signals of n electrodes, then x _i (t) corresponds to the EEG signals of the i-th electrode . Each x _i (t) can be regarded as a linear instantaneous mixture of n sources s _i (t), and the mixing matrix is A, then x(t)=As(t) (1) SOBI only uses the EEG signal x(t) measured by the sensor to obtain an approximate A ^-1 decomposition matrix W, so that $\widehat{\mathrm{the\; s}}\left(t\right)=\mathrm{wxya}\left(t\right)---\left(2\right)$ For the recovered continuous-time source signal, The SOBI algorithm has two steps: first, the EEG signal is zero-meanized, as shown in the following formula: y(t)＝B(x(t)-<x(t)>) (3) Angle brackets <·> indicate time averaging, so the mean of y is zero. The value of matrix B is such that the correlation matrix <y(t)y(t) ^T > of y is the identity matrix, and its value is given by the following formula, $\mathrm{<span\; class="notranslate">\; B</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; diag</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; u</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$ Where λ _i is the eigenvalue of the correlation matrix <(x(t)-<x(t)>)(x(t)-<x(t)>) ^T >, and each column of U is its corresponding eigenvector , The second step is to construct a set of diagonal matrices: select a set of time delays τ, and calculate the symmetric correlation matrix of the signal y(t) and its time-delayed signal y(t+τ): R _τ = sym(<y(t)y(t+τ) ^T >) (5) where sym(M)=(M+M ^T )/2 (6) This is a function that turns an asymmetric matrix into a related symmetric matrix. The process of symmetrization loses some information, but it provides an effective solution, After calculating Rτ, then diagonalize Rτ: through the rotation matrix V, use the iterative method, so that ∑ _τ ∑ _i≠j (V ^T R _τ V) _ij ² (7) obtains a minimum, then the estimate of the separation matrix W=V ^T B (8).

Images