CN115105093A - EEG signal classification and identification method based on power spectral density predetermined frequency band - Google Patents

Abstract

The invention relates to an EEG signal classification and identification method based on a predetermined frequency band based on power spectral density, comprising: obtaining an original signal, filtering and preprocessing the original signal with a bandpass filter; , obtain the frequency corresponding to the maximum power value on the spectrum; take the frequency corresponding to the maximum value as the center, and determine the frequency band according to the set length; perform secondary bandpass filtering on the preprocessed signal under the frequency band to obtain the target frequency Use the common space pattern algorithm to perform feature extraction on the target frequency components in the air domain to obtain the target features; use the neural network model to classify the target features, and output the EEG signal recognition results. Compared with the prior art, the present invention has the advantages of improving the decoding accuracy of the motion imagery EEG signal and the like.

Description

Classification and identification method of EEG signal based on power spectral density pre-determined frequency band

technical field

The invention relates to the field of EEG signal processing, in particular to an EEG signal classification and identification method based on a predetermined frequency band based on power spectral density.

Background technique

In recent years, Brain Computer Interface (BCI) technology has been extensively studied, and this technology plays an active role in promoting the rehabilitation process of post-stroke patients. Brain-computer interface technology is a new type of human-computer interaction technology that does not rely on conventional brain information output pathways (including peripheral nervous system and muscle tissue). It allows users to directly interact with the external environment through the brain or control various types of External control devices, such as wheelchairs, robots, etc. Motor imagery EEG (brain wave) is an important paradigm in brain-computer interface technology, which has the advantage of not requiring additional external stimulation, but requires a longer training process for the subjects. The motor imagery EEG signal is a kind of unstable signal, there are great differences between different subjects, and the performance of the same subject at different times will also have certain differences. This is because the EEG signal is also easily interfered by external irrelevant signals. , such as the power frequency interference of the external environment, the electromyographic signal generated by the body movement, and so on. Therefore, improving the decoding accuracy of motor imagery EEG signals is still a research difficulty in the field of brain-computer interface.

Patent CN202010022435.1 discloses an online processing method of EEG signals of motor imagery, using band-pass filter, co-space mode and support vector machine, which can realize the classification and recognition of EEG signals of motor imagery. However, this scheme is not sufficient for the extraction of components related to motor imagery in the EEG signal, which affects the decoding accuracy of the final EEG signal.

Patent CN201810044806.9 discloses a method for identifying motor imagery EEG signals based on energy features, extracting energy features in EEG signals, and classifying them using a support vector machine based on radial basis kernel function. In the process, only the energy feature of the signal is used, and the spatial feature between each channel is not considered, which affects the decoding accuracy of the final EEG signal.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a kind of EEG signal classification and identification method based on power spectral density pre-determined frequency band in order to overcome the above-mentioned defects in the prior art. Improvements have been made to finally improve the decoding accuracy of motor imagery EEG signals.

The object of the present invention can be realized through the following technical solutions:

A method for classifying and identifying EEG signals based on a predetermined frequency band based on power spectral density, comprising:

S1. Obtain the original signal, and use a band-pass filter to filter and preprocess the original signal;

S2. Perform power spectral density estimation on the preprocessed signal to obtain the frequency corresponding to the maximum power value on the spectrum;

S3. Take the frequency corresponding to the maximum value as the center, and determine the frequency band according to the set length;

S4. Perform secondary band-pass filtering on the preprocessed signal under the frequency band to obtain the target frequency component;

S5, use the common space mode algorithm to perform feature extraction operation on the target frequency components in the air domain to obtain the target features;

S6. Use the neural network model to classify the target features, and output the EEG signal recognition result.

Further, in step S1, a Butterworth bandpass digital filter of 1-30 Hz is used to filter the original signal to remove the baseline drift of the original signal and the interference component of the power frequency signal.

Further, in step S2, the power spectral density is calculated for each channel of the preprocessed signal, and the frequency corresponding to the maximum power value is obtained.

Further, in step S3, it is determined that the set length of the frequency band is 3-6 Hz.

Further, in step S4, the secondary band-pass filtering is to use a Butterworth band-pass digital filter in the range of the frequency band to perform a filtering operation on the preprocessed signal.

Further, in step S5, the co-spatial mode algorithm includes: diagonalizing the covariance matrix of the target frequency components, generating a set of optimal spatial filters, and then finding a projection matrix to project each target frequency component to the same common Space, in this public space, the variance value of each target frequency component can be differentiated to the maximum, so as to obtain the target feature.

Further, in step S6, the classifier in the neural network model adopts an error back-propagation neural network.

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

1. The present invention proposes a strategy of pre-determining frequency bands based on power spectral density, extracting the most relevant components of the EEG signal for feature extraction and feature classification operations, thereby enhancing the traditional motor imagery EEG using co-spatial mode and neural network. The classification performance of the signal processing algorithm ultimately improves the decoding accuracy of the motor imagery EEG signal.

2. The present invention can effectively classify and identify the motor imagery EEG signal, judge the movement intention of the subject, convert the intention into a corresponding control command and transmit it to the relevant peripheral device, and the peripheral device can perform the corresponding operation after receiving the command.

Description of drawings

FIG. 1 is a schematic structural diagram of a brain-computer interface system of the present invention.

FIG. 2 is a schematic flowchart of a method for classifying and identifying EEG signals according to the present invention.

Detailed ways

The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented on the premise of the technical solution of the present invention, and provides a detailed implementation manner and a specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

This embodiment provides an EEG signal classification and identification method based on a predetermined frequency band based on power spectral density, which is applied to a certain type of brain-computer interface system. The classification and recognition of EEG signals is the core part of the entire brain-computer interface system. The recognition accuracy determines the subject's ability to directly interact with the outside world through EEG, which affects the performance of the entire system. However, EEG signals are characterized by instability, low signal-to-noise ratio, and high complexity. The precise analysis and processing of EEG signals often encounter many difficulties, and the recognition accuracy is difficult to maintain a high value. The decoding process is still a challenging research work. To solve this problem, this embodiment proposes an EEG signal classification and identification method based on power spectral density pre-determined frequency bands, which improves the traditional identification and classification method combining the co-space pattern and neural network classifier, and improves the Classification accuracy of EEG signals.

As shown in Figure 1, the brain-computer interface system can generally be divided into three parts: an EEG signal acquisition module, a signal processing and identification module, and a control signal output/external device execution module.

The signal processing and identification module is the core module in the brain-computer interface system, which includes a series of processing on the original EEG signals obtained in the previous step. This stage includes three steps: preprocessing, feature extraction, and feature classification. The noise contained in the original signal is filtered out as much as possible, and the key components reflecting brain activity are left. After feature extraction, the features related to the subject's movement intention are obtained. Components, such as motor imagery (MI) spontaneous EEG energy information/rhythm, evoked EEG amplitude and phase and other characteristic information. Finally, these features are classified by a classification and recognition algorithm, so as to identify the intentions of the subjects, and further convert them into corresponding control signals and transmit them to the application and external interaction module. This embodiment focuses on the signal processing module, and uses the power spectral density to improve the traditional identification and classification algorithm combining the co-space pattern and the neural network classifier to improve the system's ability to identify EEG signals.

In the signal processing and identification module, the EEG signal classification and identification method based on the power spectral density predetermined frequency band is specifically executed by software, as shown in Figure 2, including the following steps:

1. Signal preprocessing:

Obtain the original signal, and use the band-pass filter to filter and preprocess the original signal.

2. Feature extraction:

Using the method of pre-determining the frequency band by the power spectral density, extract the frequency components of a small interval near the average value of the frequencies with the maximum energy displayed in the respective power spectral density of all channels of the preprocessed signal, and the extracted frequency components are the target frequency components. , as the input for the next feature extraction;

The feature extraction operation is performed on the target frequency components in the spatial domain by using the co-space pattern algorithm to obtain the target features.

Three, feature classification:

Use the neural network model to classify the target features, and output the EEG signal recognition results.

In the process of signal preprocessing, a 1-30Hz Butterworth band-pass digital filter is used to filter the original signal. The main function is to remove the baseline drift of the signal and the interference components such as the 50Hz power frequency signal. A Butterworth bandpass digital filter can be represented by the square of the amplitude as a function of frequency as follows:

Among them, N represents the order of the digital filter, ω represents the signal frequency, ω _c represents the cutoff frequency, and H(*) represents the amplitude.

The principle of the classification and identification method in this embodiment is that the motor imagery EEG signal is closely related to the movement-related μ rhythm (8-13 Hz) and β-rhythm (13-30 Hz) in the alpha rhythm. Imagining a specific limb movement, the motor-sensory area in the cerebral cortex related to the movement is in an excited state, and its μ/β rhythm energy decreases, this phenomenon is called Event-Related Desynchronization (ERD); in a few seconds After that, the region will return to a resting state, and the μ/β rhythm energy will rise, a phenomenon called Event-Related Synchronization (ERS). Studies have shown that ERD and ERS phenomena do not occur independently, and often occur concomitantly when performing or imagining specific limb movements. Taking left-hand/right-hand motor imagery as an example, when subjects performed left-hand motor imagery tasks, the motor sensory area of the right hemisphere, that is, the area near the C4 electrode, showed obvious ERD phenomenon; In the sensory area, that is, the area near the C3 electrode, there is a clear ERS phenomenon. When subjects performed a right-hand motor imagery task, the ERD/ERS phenomenon occurred in the opposite region. Therefore, the power spectral density can be used to determine the frequency value of the maximum energy in the channel signals such as C3 and C4 which are highly related to the motor imagery task. Then, take a frequency band near this value, and perform a second band-pass filter on the preprocessed signal to extract the signal component most relevant to motor imagery in the EEG signal, and then perform feature extraction and feature extraction on the signal component. The classification operation realizes the decoding of the subject's movement intention. In the process of feature extraction, the co-space mode algorithm is used. The key idea of the algorithm is to analyze the signal in a co-space, and fully consider the spatial characteristics of each channel. To sum up, the motion imagery EEG signal analysis and processing algorithm proposed in this embodiment can significantly improve the decoding accuracy of the signal.

In the EEG signal classification and identification method:

The feature extraction process specifically includes the following steps:

Step 1. Perform power spectral density estimation on the preprocessed signal, and draw a spectrogram.

Step 2: Determine the frequency band according to the frequency corresponding to the maximum power in the spectrogram, that is, take the frequency corresponding to the maximum value as the center, and determine the frequency band according to the set length. The set length of the frequency band is determined to be 3-6 Hz, and 4 Hz is preferably used in this embodiment.

Step 3: Perform secondary band-pass filtering on the preprocessed signal under the frequency band to obtain the target frequency component. Secondary bandpass filtering is a filtering operation on the preprocessed signal using a Butterworth bandpass digital filter in the range of the frequency band.

Step 4. Use the co-space pattern algorithm to perform feature extraction operation on the target frequency components in the air domain to obtain target features.

During the execution of the feature extraction algorithm, the frequency corresponding to the maximum power value in the power spectral density analysis of each channel needs to be determined. The corresponding formula is:

f _max =argmaxPSD(f)

When determining the frequency range of the second bandpass filter, it is necessary to extend the range with f _max as the center, and the corresponding formula is:

f _band = [f _max -2, f _max +2]

The digital filters used in the second bandpass filtering operation are the same as in the preprocessing step.

The signal component after the second band-pass filtering will have a high correlation with motor imagery, and inputting the signal component into the subsequent feature extraction step is conducive to extracting more effective features. In the feature extraction module, we use the common spatial pattern algorithm, which can effectively extract the spatial features in the signal and is widely used in the identification and classification of EEG signals. The detailed principles and steps of the algorithm are as follows:

Co-spatial pattern (CSP) is a spatial filter that is widely used in feature extraction of motor imagery EEG signals. The key idea of the algorithm is to analyze the signal in co-space. In the classification scenario of left/right-handed motor imagery EEG signals, use two classes of signals, diagonalize two covariance matrices, generate a set of optimal spatial filters, find a projection matrix, and project the two classes of signals onto the same Common space, in this space, the variance value of the two types of signals can be maximized. If it is in two or more motor imagery task scenarios, a cascaded method can be used, and multiple binary classification modules can be used to realize multi-classification processing of signals. The detailed steps of the algorithm are as follows:

For example, for the classification scene of left/right-handed motor imagery EEG signals, it is assumed that the target frequency component obtained from a single EEG experimental data is E∈R ^N*M , where N is the number of electrode channels, and M is the number of single data sampling points. Perform the covariance matrix normalization operation on each experimental data:

Among them, C represents the result matrix obtained by normalizing the covariance matrix of each experimental data, and T represents the transpose operation of the matrix.

表示左手平均，

表示右手平均，然后对两者进行相加操作，得到合成的空间协方差矩阵：Then within the class, the left-hand and right-hand cases are averaged separately,

represents the left-hand average,

Represents the right-hand average, and then adds the two to obtain the composite spatial covariance matrix:

表示左手平均归一化后空间协方差矩阵，

表示右手平均归一化后空间协方差矩阵，C_C表示由

和

合成的空间协方差矩阵。in,

represents the left-hand average normalized spatial covariance matrix,

represents the right-hand average normalized spatial covariance matrix, C _C represents the

and

The composite spatial covariance matrix.

Then perform eigenvalue decomposition on C _C , and arrange the eigenvectors according to the eigenvalues in descending order:

Among them, Uc represents the eigenvector matrix, and λ _C represents the eigenvalue matrix

对

和

进行白化变换，可得到白化后的左手/右手平均协方差矩阵为：According to the whitening matrix

right

and

After whitening transformation, the left-hand/right-hand average covariance matrix after whitening can be obtained as:

表示左手平均归一化后空间协方差矩阵，

表示右手平均归一化后空间协方差矩阵，P是白化矩阵，S_L和S_R分别为白化后的左手和右手平均协方差矩阵。in,

represents the left-hand average normalized spatial covariance matrix,

represents the right-hand average normalized spatial covariance matrix, P is the whitening matrix, and _SL and _SR are the whitened left-hand and right-hand average covariance matrices, respectively.

According to the characteristics of the whitening transformation, the whitened matrices _SL and _SR can be represented by a common eigenvector A:

S _L =Aλ _L A ^T

S _R =Aλ _R A ^T

λ _L +λ _R =I

Among them, A is the common eigenvector, λ _L is the eigenvalue matrix in the eigenvalue decomposition of the whitened matrix _SL , λ _R is the eigenvalue matrix in the eigenvalue decomposition of the whitened matrix _SR , and I is the identity matrix.

According to the above formula, when the left-hand whitened matrix _SL takes the largest eigenvalue, the right-hand whitened matrix _SR must take the smallest eigenvalue, so two types of signals can be distinguished. The projection matrix W can be obtained by multiplying the transpose of the common eigenvector A by the whitening matrix P, and then using the projection matrix to project the single EEG experimental data E, and finally calculating the variance value to obtain the CSP-processed features. The eigenvalues calculated by the CSP can be used as the input of the feedforward neural network classification model. The detailed calculation steps are as follows:

W= ^ATP

Z _p = WE, p = 1, 2, 3, ..., 2m

Among them, A is the common eigenvector, P is the whitening matrix, W is the projection matrix, E is the target frequency component obtained from the single EEG experimental data, p is the feature serial number, Z is the projected result matrix of E, var(*) In order to calculate the variance value of the vector, f _p is the final calculated eigenvalue.

In this embodiment, the classifier in the neural network model selects an error back propagation neural network, referred to as a BP (Back Propagation) neural network for short. BP neural network has the characteristics of simple structure, nonlinear and strong adaptability, and is widely used in pattern recognition, regression analysis and other fields.

In the model structure of BP neural network, there are three basic components, namely input layer, hidden layer and output layer. In the classification model of this embodiment, the input layer includes 3 neurons, the middle layer is a single hidden layer, including 10 neurons, and the output layer includes 2 neurons.

The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make many modifications and changes according to the concept of the present invention without creative efforts. Therefore, all technical solutions that can be obtained by those skilled in the art through logical analysis, reasoning or limited experiments on the basis of the prior art according to the concept of the present invention shall fall within the protection scope determined by the claims.

Claims (7)

1. a kind of EEG signal classification and identification method of predetermined frequency band based on power spectral density, is characterized in that, comprises: S1. Obtain the original signal, and use a band-pass filter to filter and preprocess the original signal; S2. Perform power spectral density estimation on the preprocessed signal to obtain the frequency corresponding to the maximum power value on the spectrum; S3. Take the frequency corresponding to the maximum value as the center, and determine the frequency band according to the set length; S4. Perform secondary band-pass filtering on the preprocessed signal under the frequency band to obtain the target frequency component; S5, use the common space mode algorithm to perform feature extraction operation on the target frequency components in the air domain to obtain the target features; S6. Use the neural network model to classify the target features, and output the EEG signal recognition result. 2. A kind of EEG signal classification and identification method based on power spectral density predetermined frequency band according to claim 1, it is characterized in that, in step S1, use 1～30Hz Butterworth bandpass digital filter to original signal Perform a filtering operation to remove the baseline drift of the original signal and the interference components of the power frequency signal. 3. a kind of EEG signal classification and identification method based on power spectral density predetermined frequency band according to claim 1, is characterized in that, in step S2, carries out power spectral density calculation to each channel of the preprocessed signal , to get the frequency corresponding to the maximum power among them. 4 . The method for classifying and identifying EEG signals based on a predetermined frequency band based on power spectral density according to claim 1 , wherein in step S3 , the set length of the determined frequency band is 3-6 Hz. 5 . 5. a kind of EEG signal classification and identification method based on power spectral density predetermined frequency band according to claim 1, is characterized in that, in step S4, secondary band-pass filtering is to use Butterworth within the scope of frequency band A bandpass digital filter performs a filtering operation on the preprocessed signal. 6. a kind of EEG signal classification and identification method based on power spectral density predetermined frequency band according to claim 1, is characterized in that, in step S5, common space pattern algorithm comprises: the covariance matrix of diagonalization target frequency component , generate a set of optimal spatial filters, and then find a projection matrix to project the target frequency components into the same common space, in which the variance value of the target frequency components can be maximized to distinguish, thus obtaining target characteristics. 7 . The method for classifying and identifying EEG signals based on a predetermined frequency band based on power spectral density according to claim 1 , wherein in step S6 , the classifier in the neural network model adopts an error back-propagation neural network. 8 .

Images