阅读说明：本技术 一种基于eemd的疲劳脑电特征提取方法 (EEMD-based fatigue electroencephalogram feature extraction method ) 是由 马玉良 郑赟 孙明旭 高云园 佘青山 于 2021-06-25 设计创作，主要内容包括：本发明公开了一种基于EEMD的疲劳脑电特征提取方法。本发明首先使用脑电采集设备采集疲劳驾驶信号,选取大脑皮层上能显著反映疲劳状态的电极通道并预处理；对信号进行EEMD分解并获得前3个IMF分量；对每个IMF分量进行STFT,并且计算每个IMF分量的PSD；最后使用多种分类器分别对数据进行二分类,验证了特征的鲁棒性。本发明减少了使用的电极数,减少了冗余信息,大大增加了实用性,并且使用了EEMD-IMF-PSD特征提高了分类准确率,优于传统的5个子频带的PSD特征。(The invention discloses a fatigue electroencephalogram feature extraction method based on EEMD. Firstly, collecting fatigue driving signals by using electroencephalogram collecting equipment, selecting and preprocessing an electrode channel on cerebral cortex capable of obviously reflecting fatigue state; EEMD decomposition is carried out on the signal to obtain the first 3 IMF components; performing STFT on each IMF component, and calculating PSD of each IMF component; and finally, performing secondary classification on the data by using various classifiers respectively, and verifying the robustness of the characteristics. The invention reduces the number of used electrodes, reduces redundant information, greatly increases the practicability, and improves the classification accuracy by using the EEMD-IMF-PSD characteristic, which is superior to the traditional PSD characteristic of 5 sub-bands.)

1. A fatigue electroencephalogram feature extraction method based on EEMD is characterized by comprising the following steps:

step 1, collecting fatigue simulation driving electroencephalogram signals by using equipment;

step 2, selecting 7 electrode channels according to the relevant areas of the fatigue electroencephalogram signals in the cerebral cortex, and preprocessing the signals;

step 3, carrying out empirical mode decomposition on the preprocessed signals to obtain the first 3 IMF components of each segment;

step 4, performing STFT on each IMF component, calculating a corresponding PSD, and constructing a PSD feature space;

and 5, putting the data of the optimal channel combination obtained in the step into a plurality of classifiers for training to respectively obtain training models, and inputting the test data of the optimal channel combination into each training model for classification to obtain a classification result.

2. The EEMD-based fatigue electroencephalogram feature extraction method as claimed in claim 1, which is characterized in that: in the step 2, the 16 fixed electrodes and the pretreatment process specifically comprise:

the fatigue brain electrical signals in the relevant area in the cerebral cortex select FC1, FC2, Cz, C3, C4, CP1 and CP2 as processing signals, the signals are subjected to 200Hz down-sampling and 0.1Hz-30Hz band-pass filtering, and the obtained data of each subject are as follows: 7x240000x 2.

3. The EEMD-based fatigue electroencephalogram feature extraction method as claimed in claim 1, which is characterized in that:

in the step 3, EEMD decomposition of EEMD data is carried out,

EEMD decomposition is performed for each segment:

(1) in the original signal [ x (t), t ═ 0, 1., N]White noise n (t), E [ n (t)]New signal x' (t) ═ x (t) + n (t) is obtained, and r is also recorded_1,1(t)＝x′(t)；

(2) Calculating a signal r_1,1(t) forming the upper and lower envelope e of the signal using cubic spline interpolation_upper(t) and e_lower(t)；

(3) Calculate the mean m of the upper and lower envelopes₁(t):

(4)r_2,1(t)＝r_1,1(t)-m₁(t) and then returning to the step (2);

(5) repeating the steps (2) to (4) until the number N of the extreme points in the sequence is reached after the k time of iteration_eAnd the number of zero points N_zThe following relationship is satisfied:

Nz-1≤Ne≤Nz+1

m_k(t)＝0

(6) let IMF₁(t)＝r_k,1(t) calculating a first IMF component, and then letting r_1,2(t)＝x′(t)-IMF₁(t), returning to the step (2);

(7) repeating steps (2) - (6) until the first time, the standard deviation SD (l) is satisfied:

e represents a threshold value of standard deviation, and EMD decomposition is completed for one time;

(8) returning to the step (1), repeating the steps (1) - (7) M times, and averaging the sum of the corresponding IMF components to obtain a final IMF component;

in the above equation, r_i,j(t) represents the i-th result of iterating the j-th IMF component at time t, K is the amplitude coefficient in the added white noise sequence, σ_n(t)Is the standard deviation, σ, of the noise_n(t)Smaller values of (c) indicate higher decomposition accuracy.

4. The EEMD-based fatigue electroencephalogram feature extraction method as claimed in claim 1, which is characterized in that: in step 4, PSD features are extracted from the obtained IMF component, and the specific steps are as follows:

time-frequency analysis was performed using STFT:

wherein x [ n ]]Is a sequence of signals that is to be transmitted,is the angular frequency, k ═ 0, 1.., N-1, w [ N ]]The window function is represented by using a Hanning window, and the formula is as follows:

calculating the PSD of each IMF component:

wherein f is_nRepresenting the IMF component, F (ω) is the fourier transform of the signal and F (ω) is the conjugate of F (ω).

Technical Field

The invention relates to a driving fatigue electroencephalogram feature extraction method, in particular to a method for extracting features of fatigue electroencephalogram signals after EEMD decomposition.

Background

Driving fatigue can lead to serious accidents, particularly when the driver is often tired or drowsy while driving for long distances. Therefore, fatigue detection has become an important issue in the field of driving safety. At present, the main fatigue detection methods are of the following three types:

(1) behavior characteristics: including a change in lateral lane position, which is one of the most common indicators for detecting fatigue driving through vehicle behavior, and a change in vehicle travel direction difference. These metrics are conceptually easy to understand, but the selection of these metrics has not yet formed a uniform standard, which makes their implementation challenging.

(2) Facial expression characteristics such as degree of resting eye closure or nodding frequency. While these features are simple and easy to use, the detection of these features is affected by a variety of factors, including image angle and illumination level, among others. These disadvantages reduce their overall recognition accuracy.

(3) Physiological features including Electrocardiogram (ECG), heart rate, Electromyogram (EMG), electroencephalogram (EEG) based features. Unlike behavioral and facial features, these features are objective indicators of physiological changes in the human body. Among these signals, electroencephalography is considered to be one of the most direct, effective, and promising ones for detecting driving fatigue. The electroencephalogram-based fatigue driving detection is generally divided into the following 4 steps: data acquisition, data preprocessing, feature extraction and classification.

In the fatigue detection based on the electroencephalogram signals, different feature selection methods directly influence the classification accuracy. A common method is to divide electroencephalogram signals into an Alpha band, a Beta band, a Gamma band, a Delta band, and a Theta band, and then calculate the power spectral density of each band as a signal characteristic. Some researchers have improved the choice of frequency bands and have obtained more suitable features through different combinations of frequency bands. Although this simple feature extraction method is effective, the accuracy is limited because the feature extraction based on the frequency domain only ignores a large amount of hidden time domain information. Therefore, the classification accuracy of the driving fatigue detection can be further improved by adopting a proper time-frequency analysis method to extract the electroencephalogram signal characteristics. A short-time fourier transform (STFT) is used for time-frequency domain analysis of driving fatigue detection. The limitation of short-time fourier transform for driving fatigue detection is that the time window length is fixed. For non-stationary signals, such as electroencephalography (EEG), a small window is required for high-band analysis, and a large window is required for low-band analysis.

Empirical Mode Decomposition (EMD) is an adaptive time-frequency analysis method that adaptively performs signal Decomposition according to the time scale characteristics of data itself, overcoming the limitation of fixed time window length in short-time fourier transform (STFT). However, a limitation of empirical mode decomposition is the frequent occurrence of modal aliasing, resulting in a spurious time-frequency distribution. In order to further improve the performance of EMD, researchers have proposed an integrated Empirical Mode Decomposition (EEMD). EEMD involves a comprehensive screening of the white noise signal and its mean value as the final true value. The white noise is added to improve the signal distribution precision of the extreme point, so that the problem of mode aliasing of the original EMD can be solved more effectively, and the signal-to-noise ratio of the signal is obviously improved.

Disclosure of Invention

The invention provides a fatigue electroencephalogram feature extraction method based on EEMD (ensemble empirical mode decomposition), aiming at the defects of the prior art, and the method comprises the following steps:

step 1, collecting fatigue simulation driving electroencephalogram signals by using equipment;

step 2, selecting 7 electrode channels according to the relevant areas of the fatigue electroencephalogram signals in the cerebral cortex, and preprocessing the signals;

step 3, carrying out empirical mode decomposition on the preprocessed signals to obtain the first 3 IMF components of each segment;

step 4, performing STFT on each IMF component, calculating a corresponding PSD, and constructing a PSD feature space;

and 5, putting the data of the optimal channel combination obtained in the step into a plurality of classifiers for training to respectively obtain training models, and inputting the test data of the optimal channel combination into each training model for classification to obtain a classification result.

According to the technical scheme provided by the invention, a series of Intrinsic Mode Functions (IMFs) containing time-frequency characteristics are obtained by decomposing fatigue electroencephalogram signals through EEMD. The Power Spectral Density of each IMF component is then calculated by STFT, constructing a Power Spectral Density (PSD) feature space. Compared with the traditional PSD characteristic directly using 5 sub-bands, the characteristic difference between the normal state and the fatigue state is increased. And the common optimal channel combination among different subjects is found, the number of electrodes in the experiment and the data processing time are greatly reduced, and the classification accuracy is improved.

Preferably, in step 2, the 16 fixed electrodes and the pretreatment process specifically include:

the fatigue brain electrical signals in the relevant area in the cerebral cortex select FC1, FC2, Cz, C3, C4, CP1 and CP2 as processing signals, the signals are subjected to 200Hz down-sampling and 0.1Hz-30Hz band-pass filtering, and the obtained data of each subject are as follows: 7x240000x 2.

Preferably, in the step 3, EEMD decomposition of the electroencephalogram data is performed, and EEMD decomposition is performed on each segment, and the main steps are as follows:

(1) in the original signal [ x (t), t ═ 0, 1., N]Adding different white noises n (t), E [ n (t)]New signal x' (t) ═ x (t) + n (t) is obtained, and r is also recorded_1，1(t)＝x′(t)；

(2) Calculating a signal r_1，1(t) forming the upper and lower envelope e of the signal using cubic spline interpolation_upper(t) and e_lower(t)；

(3) Calculate the mean m of the upper and lower envelopes₁(t)：

(4)r_2，1(t)＝r_1，1(t)-m₁(t) and then returning to the step (2);

(5) repeating the steps (2) to (4) until the k time of iteration is reached and the number N of extreme points in the sequence is reached_eAnd the number of zero points N_zThe following relationship is satisfied:

Nz-1≤Ne≤Nz+1

m_k(t)＝0

(6) let IMF₁(t)＝r_k，1(t), by this step, the first IMF component is calculated. Then let r_1，2(t)＝x′(t)-IMF₁(t), returning to the step (2);

(7) repeating steps (2) - (6) until the first time, the standard deviation SD (l) is satisfied:

e represents a threshold value of standard deviation, and EMD decomposition is completed for one time;

(8) returning to the step (1), repeating the steps (1) - (7) M times, and averaging the sum of the corresponding IMF components to obtain a final IMF component;

in the above equation, r_i，j(t) represents the i-th result of iterating the j-th IMF component at time t, K is the amplitude coefficient in the added white noise sequence, σ_n(t)Is the standard deviation with noise, and smaller values indicate higher resolution accuracy.

Preferably, in step 4, the PSD features of the obtained IMF components are extracted, and the specific steps include:

STFT is adopted for time-frequency analysis, and the calculation formula is as follows:

wherein x [ n ]]Is a sequence of signals that is to be transmitted,is the angular frequency, k ═ 0, 1. W [ n ]]The window function is represented by using a Hanning window, and the formula is as follows:

calculating the PSD of each IMF component, wherein the specific formula is as follows:

wherein f is_nRepresenting the IMF component, F (ω) is the fourier transform of the signal and F (ω) is the conjugate of F (ω).

The invention has the following beneficial effects:

the method comprises the steps of performing signal decomposition in a self-adaptive mode according to time scale characteristics of fatigue electroencephalogram data by adopting an EEMD decomposition algorithm to obtain corresponding IMF components, further obtaining PSD characteristics from the IMF components through STFT calculation, and constructing a PSD characteristic space, wherein the difference between a normal state and a fatigue state can be reflected better than the traditional PSD characteristic space constructed on the basis of 5 sub-bands; the simulated driving data set is used as a sample, and data of 7 electrodes are selected from relevant areas in cerebral cortex for processing, so that the classification accuracy is improved while the number of fixed electrodes is greatly reduced, and the practical degree of equipment and an algorithm is improved to a great extent.

Drawings

FIG. 1 is a graph comparing fatigue and normal states of EEMD-IMF-PSD and conventional PSD;

FIG. 2 is a diagram of a basic structure for obtaining EEMD-IMF-PSD;

FIG. 3 is a general flowchart of fatigue driving detection;

Detailed Description

The present invention is further illustrated by the following specific examples. The following description is exemplary and explanatory only and is not restrictive of the invention in any way.

Step 1, collecting fatigue driving simulation electroencephalogram signals by using equipment, wherein the collected electroencephalogram signals are obtained from an international 10-20 lead system worn by 6 subjects, and the sampling frequency is 1000 Hz;

and 2, selecting 7 electrode channels according to the relevant areas of the fatigue electroencephalogram signals in the cerebral cortex, wherein the electrode channels are respectively as follows: FC1, FC2, Cz, C3, C4, CP1 and CP2, and 200Hz down-sampling and 0.1-30 Hz band-pass filtering are carried out on the obtained signals, and the obtained data format is 7x240x 2000;

and 3, carrying out EEMD decomposition on the preprocessed signal to obtain the first 3 IMF components, wherein the specific steps are as follows:

3-1, slicing the preprocessed data, wherein each 10s segment is a total of 240 segments (including 120 normal segments and 120 fatigue segments), and the data specifically comprises: 32x240x 2000;

3-2, carrying out EEMD decomposition on each segment;

3-3. get the first 3 IMF components of each fragment, for a total of 7x240x3 IMF components.

Step 4, performing STFT on each IMF component, calculating a corresponding PSD, and constructing a PSD feature space, wherein the specific steps are as follows:

4-1, performing time-frequency analysis by adopting the STFT of a Hanning window to obtain the STFT of a signal sequence;

4-2. calculating PSD of each IMF component through the result of STFT.

And 5, training the data of the optimal channel combination in a plurality of classifiers to respectively obtain training models, and inputting the test data of the optimal channel combination into each training model for classification to obtain a classification result.

In the step 3, EEMD decomposition of the electroencephalogram data is performed, EEMD decomposition is performed on each segment, and the main steps are as follows:

(1) in the original signal [ x (t), t ═ 0, 1., N]Adding different white noises n (t), E [ n (t)]New signal x' (t) ═ x (t) + n (t) is obtained, and r is also recorded_1，1(t)＝x′(t)；

(2) Calculating a signal r_1，1(t) forming the upper and lower envelope e of the signal using cubic spline interpolation_upper(t) and e_lower(t)；

(3) Calculate the mean m of the upper and lower envelopes₁(t)：

(4)r_2，1(t)＝r_1，1(t)-m₁(t) and then returning to the step (2);

(5) repeating the steps (2) to (4) until the k time of iteration is reached and the number N of extreme points in the sequence is reached_eAnd the number of zero points N_zThe following relationship is satisfied:

Nz-1≤Ne≤Nz+1

m_k(t)＝0

(6) let IMF₁(t)＝r_k，1(t), by this step, the first IMF component is calculated. Then let r_1，2(t)＝x′(t)-IMF₁(t), returning to the step (2);

(7) repeating steps (2) - (6) until the first time, the standard deviation SD (l) is satisfied:

e represents a threshold value of standard deviation, and EMD decomposition is completed for one time;

(8) returning to the step (1), repeating the steps (1) - (7) M times, and averaging the sum of the corresponding IMF components to obtain a final IMF component;

in the above equation, r_i，j(t) represents the i-th result of iterating the j-th IMF component at time t, K is the amplitude coefficient in the added white noise sequence, σ_n(t)Is the standard deviation with noise, and smaller values indicate higher resolution accuracy.

In step 4, PSD features are extracted from the obtained IMF component, and the specific steps are as follows:

STFT is adopted for time-frequency analysis, and the calculation formula is as follows:

wherein x [ n ]]Is a sequence of signals that is to be transmitted,is angular frequency，k＝0，1，...，N-1，w[n]The window function is represented by using a Hanning window, and the formula is as follows:

calculating the PSD of each IMF component, wherein the specific formula is as follows:

wherein f is_nRepresenting the IMF component, F (ω) is the fourier transform of the signal and F (ω) is the conjugate of F (ω).

The flow chart of the whole work is shown in fig. 3, wherein 6 subjects are taken as experimental objects, each subject takes 1 experiment, 7 channels are recorded in each experiment, the method is adopted for feature extraction, 50% of ideal channels are randomly selected as training data, 50% of ideal channels are taken as test data for comparison, and finally, a plurality of classifiers are used for carrying out secondary classification on fatigue brain electricity, wherein the classes are a normal state and a fatigue state respectively. The table records pairs of SVM method, SVM (RBF) method, KNN method, H-ELM method, PSO-E-ELM method: the feature extraction algorithm provided by the invention and the traditional PSD feature are used for comparing the accuracy rate (accuracycacy) of the two-classification (see table 1).

TABLE 1 comparison of recognition rates for EEMD-IMF-PSD and conventional PSD classifications using multiple classifiers

Compared with the PSD characteristic of the traditional sub-band, the PSD characteristic of the IMF component after EEMD decomposition is carried out on the signal, the obtained identification rate is obviously improved, the effectiveness of the IMF component in the fatigue driving detection field based on the EEMD-IMF-PSD characteristic and the PSO-H-ELM classifier is verified, the characteristic selection method is suitable for different subjects, the classification effect is good, and the accuracy is as high as 99.17%. It can be seen that the algorithm provided by the invention has certain improvement on the capability of detecting fatigue driving.