A method and system for classifying fatigue driving based on electroencephalogram multi-scale fuzzy entropy features

Abstract

The application belongs to the field of electroencephalogram signal processing, and is a fatigue driving classification method and system based on electroencephalogram multi-scale fuzzy entropy features of CEEMDAN. The method comprises the following steps: collecting electroencephalogram data, and labeling part of the data; pre-processing the electroencephalogram signal to remove artifacts in the electroencephalogram signal; extracting features from the pre-processed electroencephalogram signal, constructing and training an SVM classifier, classifying unmarked data using the trained SVM classifier to obtain pseudo-labeled data; performing CEEMDAN processing and scale transformation processing on the pseudo-labeled data and the previously obtained labeled data to obtain electroencephalogram multi-scale fuzzy entropy features based on CEEMDAN, and establishing a fatigue state classification model to classify the electroencephalogram signal. The collection method is simple, is more suitable for application in an intelligent driving system, and has a high fatigue state recognition rate.

Description

Technical Field

[0001] This invention belongs to the field of electroencephalogram (EEG) signal processing within the field of biological signal processing, specifically relating to a fatigue driving classification method and system based on CEEMDAN multi-scale fuzzy entropy features. Background Technology

[0002] Studies have shown that fatigue is one of the main causes of frequent traffic accidents. Fatigue significantly affects a driver's attention and decision-making abilities, thereby reducing driver alertness. Therefore, monitoring driver fatigue and providing early warnings is crucial. Driver fatigue detection is mainly divided into two methods: subjective fatigue detection and objective fatigue detection. The former is often used as an auxiliary means of assessing fatigue levels, such as the Stanford Drowsiness Scale (SSS) and the Karolinska Drowsiness Scale (KSS). The latter mainly includes detection methods based on vehicle driving conditions, driver behavioral characteristics, and driver physiological signal parameters. Vehicle driving conditions include the distance of the vehicle from the predetermined track under normal driving conditions, the reaction time required to adjust the vehicle, and the steering wheel angle. Driver behavioral characteristics include blinking frequency, percentage of eye closure, and duration of eye closure. Driver physiological signals, such as electroencephalography (EEG), electrocardiography (ECG), and electromyography (EMG), are currently effective means of driver fatigue detection.

[0003] EEG (electroencephalography), as a direct reflection of brain activity, is widely used in current fatigue driving detection due to its good temporal resolution and non-invasive signal acquisition. However, traditional multi-channel EEG acquisition methods are complex and not portable, which is not conducive to the practical application of intelligent vehicle driving systems. Furthermore, machine learning-based fatigue recognition methods require a large amount of labeled information as input for training the classifier; otherwise, serious overfitting problems will occur. However, various state annotation methods are complex, unreliable, and sometimes even expensive. Therefore, researching a driver fatigue classification method based on single-channel EEG data and self-trained semi-supervised learning is more in line with practical needs. Summary of the Invention

[0004] To address the aforementioned problems with existing brainwave-based fatigue driving classification methods, this invention proposes a fatigue driving classification method and system based on multi-scale fuzzy entropy features of brainwaves. It processes brainwave signals based on CEEMDAN, eliminating the need for multi-lead brainwave signals. The brainwave acquisition method is simple and portable, and the state labeling method for brainwave samples is more practical.

[0005] The technical solution adopted in this invention is as follows: A fatigue driving classification method based on EEG multi-scale fuzzy entropy features, comprising the following steps:

[0006] We conducted a simulated driving experiment to collect EEG data and labeled some of the data.

[0007] Preprocessing of EEG signals removes artifact interference.

[0008] Feature extraction was performed on the preprocessed EEG signals, an SVM classifier was constructed and trained, and the trained SVM classifier was used to classify the unlabeled data to obtain pseudo-labeled data.

[0009] Both pseudo-labeled data and pre-obtained labeled data were processed using CEEMDAN and scale transformation to obtain CEEMDAN-based multi-scale fuzzy entropy features of EEG, and a fatigue state classification model was established to classify EEG signals for fatigue states.

[0010] Preferably, the EEG signal is preprocessed to remove non-physiological artifacts; non-physiological artifacts include power frequency interference and noise caused by improper electrode contact with the scalp; the steps for removing non-physiological artifacts include:

[0011] First, discard the first segment of the EEG signal and remove abnormal data segments from the EEG signal; then filter the signal to eliminate DC bias and power frequency interference.

[0012] Preferably, the EEG signal is preprocessed by using discrete wavelet transform based on statistical threshold and coif3 wavelet function to remove physiological artifact interference from the EEG signal.

[0013] More preferably, the physiological artifacts include electrooculogram (EOG) artifacts, and the steps for removing EOG artifacts include:

[0014] The EEG signal was decomposed into multiple levels using a discrete wavelet transform based on the coif3 wavelet function, yielding an approximate coefficient set and a detail coefficient set {C}. k The statistical threshold ST for each coefficient group (k = 1, 2, ..., n) is calculated as follows:

[0015] ST = 1.5 * STD(C k )

[0016] Among them, STD(C) k ) is the standard deviation of the k-th level wavelet coefficient set;

[0017] If the absolute value of the coefficients in the wavelet coefficient set is greater than the statistical threshold, then the standard deviation of the wavelet coefficient set is set to 0.

[0018] Finally, the EEG signal with oculomotor artifacts removed is obtained through inverse wavelet transform.

[0019] Preferably, the steps for obtaining pseudo-labeled data include:

[0020] An SVM classifier is trained using the feature information of labeled data obtained through prior experiments to establish a self-trained model. The trained SVM classifier is then used to predict unlabeled data to obtain pseudo-labeled data.

[0021] Preferably, the step of feature extraction from the preprocessed EEG signal includes:

[0022] Subband signals of the original EEG signal (Delta, Theta, Alpha, Beta) were obtained through wavelet packet transform. Temporal and frequency domain features of the EEG signal were extracted from these subband signals. Mobility (HM) and complexity (HC) were selected to describe the temporal features of the EEG signal and its subband signals, completing the temporal feature extraction. Fast Fourier Transform was used to calculate the relative energy and energy ratio of the original EEG signal and each subband signal. The Welch average periodogram method was used to calculate the power spectral density of each subband signal, completing the frequency domain feature extraction.

[0023] Extract entropy features from EEG signals, including Renyi entropy, spectral entropy, and fuzzy entropy.

[0024] The technical solution adopted by the system of this invention is: a fatigue driving classification system based on EEG multi-scale fuzzy entropy features, comprising the following modules:

[0025] The data acquisition module conducts simulated driving experiments to collect EEG data and annotates some of the data;

[0026] The preprocessing module preprocesses the EEG signals to remove artifact interference.

[0027] The feature extraction module extracts features from the preprocessed EEG signals, constructs and trains an SVM classifier, and uses the trained SVM classifier to classify unlabeled data to obtain pseudo-labeled data.

[0028] The fatigue state classification module performs CEEMDAN processing and scale transformation on both pseudo-labeled data and pre-obtained labeled data to obtain CEEMDAN-based EEG multi-scale fuzzy entropy features, establishes a fatigue state classification model, and classifies EEG signals for fatigue states.

[0029] This invention classifies fatigue states based on electroencephalogram (EEG) signals, distinguishing between the awake and fatigued states of a driver. Compared with existing technologies, the beneficial effects include:

[0030] First, most current methods for identifying driver fatigue are based on multi-lead EEG signals, which is not conducive to practical applications. This invention, however, identifies driver fatigue based on CEEMDAN's multi-scale fuzzy entropy features and self-trained pseudo-labels. It can analyze data using only EEG data from the left prefrontal cortex (FP1). Its EEG acquisition method is simple and portable, and its EEG sample state labeling method is more practical. Second, a self-trained semi-supervised method is used to convert unlabeled data into pseudo-labeled data, improving information utilization. Attached Figure Description

[0031] Figure 1 This is a flowchart of feature extraction and classification model construction in the fatigue driving classification method of this invention. Detailed Implementation

[0032] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments, but the specific implementation of the present invention is not limited thereto.

[0033] Example 1

[0034] This embodiment provides a fatigue driving classification method based on EEG multi-scale fuzzy entropy features, including the following steps:

[0035] Step 1: Conduct a simulated driving experiment to collect EEG data and annotate some of the data.

[0036] A simulated driving environment was established and a simulated driving experiment was conducted. Several qualified subjects were selected, and simulated driving tasks were provided to the subjects by running the Euro Truck Simulator 2 software platform.

[0037] During the simulated driving experiment, the channel positions were determined according to the 10-20 international standard lead system. A Mangold-10 multi-channel physiological device was used to collect single-channel EEG signals from the subject's left forehead, specifically the FP1 channel. Ag-AgCl patch electrodes were selected as the contact electrodes. The positive electrode was placed at the driver's left forehead FP1, and the negative electrode was placed below the driver's left ear as a reference electrode. A ground electrode was placed at the back of the neck to reduce artifact interference. The multi-channel physiological device recorded the EEG signals at a sampling rate of 256 Hz and simultaneously uploaded them to the Biotrace software platform. Most of the driving scenarios were highway scenarios.

[0038] The Karolinska Drowsiness Scale (KSS) and a state recording button were used to record the subject's fatigue level and state transition time during driving. The 10 different fatigue states on the KSS scale were divided into three levels: alert (KSS <= 5), mild fatigue (KSS = 6 or KSS = 7), and severe fatigue (KSS >= 8). In this embodiment, both mild and severe fatigue were considered fatigue. A score of 1 to 5 on the KSS scale was considered a normal driver state, while a score of 6 to 10 was considered a fatigued driver state. Data from severe fatigue was used as much as possible, rather than mild fatigue data, in the final dataset.

[0039] This embodiment uses a subjective scale as an auxiliary tool, primarily employing a state recording button annotation method. Subjects are instructed to press the state recording button on the steering wheel when they perceive a change in state. This triggers the screenshot function of the Euro Truck Simulator 2 software platform, recording the time point of the subject's state change via the timestamp on the screenshot. By combining the states displayed on the scale before and after driving, the corresponding time periods for alertness and fatigue can be determined, thus annotating the EEG data.

[0040] Step 2: Preprocess the EEG signals to remove artifact interference.

[0041] This step removes two types of artifacts from EEG signals: physiological artifacts and non-physiological artifacts. Non-physiological artifacts include power line interference and noise caused by improper electrode contact with the scalp. These non-physiological artifacts can be processed using MATLAB's open-source toolbox EEGLAB. The steps for removing non-physiological artifacts include:

[0042] First, the initial segment (e.g., the first 3 minutes) of the EEG signal is discarded, and abnormal data segments with high amplitude or excessive noise are removed through manual inspection. Then, EEGLAB's filtering function is used to eliminate DC bias and power frequency interference, for example, using a 1Hz low-pass filter and a 30Hz high-pass filter. After removing non-physiological artifacts, the EEG signal is segmented into 5-second segments with no overlap.

[0043] Since electrooculogram (EOG) artifacts are particularly prominent physiological artifacts in electroencephalogram (EEG) signals, this embodiment employs discrete wavelet transform based on statistical thresholds and the coif3 wavelet function for artifact removal. The specific steps for removing EOG artifacts are as follows:

[0044] First, the EEG signal is decomposed into 9 levels using a discrete wavelet transform based on the coif3 wavelet function, yielding an approximation coefficient set and a detail coefficient set {C}. k The statistical threshold ST for each coefficient group (k = 1, 2, ..., n) is calculated as follows:

[0045] ST = 1.5 * STD(C k (1)

[0046] Among them, STD(C) k ) is the standard deviation of the k-th wavelet coefficient group.

[0047] If the absolute value of the coefficients in the wavelet coefficient set is greater than the statistical threshold, the standard deviation of the wavelet coefficient set is set to 0, and finally the EEG signal with oculomotor artifacts removed is obtained by inverse wavelet transform.

[0048] Step 3: Extract features from the preprocessed EEG signals, construct and train an SVM classifier, and use the trained SVM classifier to classify the unlabeled data to obtain pseudo-labeled data.

[0049] This step utilizes the feature information of pre-labeled data (i.e., initial training samples) obtained through prior experiments to train an SVM classifier, establish a self-trained model, and predict unlabeled data to obtain pseudo-labeled data. To obtain more reliable pseudo-labeled data, as much feature information as possible is extracted from the EEG signals during the feature extraction process. The extracted feature information includes time-domain, frequency-domain, and entropy features of the EEG signals.

[0050] like Figure 1 As shown, to extract the time-domain, frequency-domain, and entropy features of EEG signals, this embodiment first uses wavelet packet transform to extract sub-band signals. Since the sampling frequency of the acquired EEG signals is 256 Hz, the original EEG signals are decomposed into 7 levels. The decomposed coefficient array is arranged into a binary tree according to the frequency range, where each node represents the frequency range to which the coefficient array belongs. Sub-band signals of theta, delta, alpha, and beta are extracted. Based on this, the Hjorth parameter, power spectral density, spectral entropy, sample entropy, and fuzzy entropy features of each sub-band are extracted, for a total of 53 features. Then, principal component analysis (PCA), non-negative matrix factorization (NMC), and T-SNE are used to reduce the dimensionality of the features. Finally, a self-trained semi-supervised model based on SVM is used to obtain pseudo-labeled data. The probability threshold is set to 0.8, and after several rounds of training iterations, until the predicted probability of the sample is less than 0.8 or no labeled data is retained.

[0051] In the feature extraction process, to increase information diversity, three types of features (time domain, frequency domain, and entropy) were extracted from the EEG signal and its subband signals, specifically including Hjorth parameter, power spectral density, spectral entropy, sample entropy, and fuzzy entropy. The specific steps include:

[0052] Step 31: Obtain the sub-band signals of the original EEG signal through wavelet packet transform: Delta (1-3hz), Theta (4-7hz), Alpha (8-13hz), Beta (14-30hz), and extract the time domain and frequency domain features of the EEG signal through the sub-band signals.

[0053] In this embodiment, Hjorth mobility (HM) and Hjorth complexity (HC) are selected to describe the temporal characteristics of EEG signals and their subband signals, thus completing the temporal feature extraction of EEG signals. For the time series {x(n)|n=1,2,…,n}, the calculation formulas for mobility HM and complexity HC are as follows:

[0054]

[0055]

[0056] Where x′(n) represents the difference between time series x(n+1) and x(n), and x″(n) represents the difference between x′(n+1) and x′(n). Let x(n) represent the mean. Let x′(n) represent the mean. Let x″(n) represent the mean.

[0057] The Fast Fourier Transform (FFT) was used to calculate the relative energy and energy ratio of the original EEG signal and each sub-band signal, and the Welch average periodogram method was used to calculate the power spectral density of each sub-band signal, thus completing the frequency domain feature extraction of the EEG signal. The energy expression of each sub-band signal is as follows:

[0058]

[0059] in, Let m be the set of sub-band signals, and s be the frequency of each sub-band signal. i The amplitude of the corresponding frequency is obtained by performing an FFT transform on the original EEG signal.

[0060] In addition, the energy ratios of the subband signals are also extracted, such as the energy ratio of theta subband signal to the alpha subband signal power(theta) / power(alpha), and the energy ratio of delta subband signal to the alpha subband signal power(delta) / power(alpha).

[0061] Step 32: Extract the entropy features of the EEG signal.

[0062] Entropy features include Renyi entropy, spectral entropy, and fuzzy entropy. Based on the histogram density estimation method, the EEG signal is divided into k equally wide bins according to the value range. The width of each bin is determined by w = (max... 1≤n≤N x(n)-min 1≤n≤N The expression for renyi entropy is determined by x(n)) / k.

[0063]

[0064] Where n i This represents the number of data points falling into the i-th partition. In this embodiment, k is a constant, which can be set to 20; α is set to 2, i.e., the quadratic entropy is calculated.

[0065] The general definition of spectral entropy is:

[0066]

[0067] Where f0 and f n These are the upper and lower frequencies of the frequency band, respectively, P f This represents the normalized power spectral density at frequency f in this frequency band.

[0068] For a given time series {x(n)|n=1,2,...,N}, calculating the fuzzy entropy requires reconstructing a new m-dimensional vector, X. m (i) = {x(i), x(i+1), ..., x(i+m-1)} - x0(i), where i = 1, 2, ..., n-m+1; where x0(i) is the average of the sequence calculated by the following formula:

[0069]

[0070] In this embodiment, m is set to 2. Furthermore, the parameters n and r in the fuzzy membership function are set to 4 and 0.2 × STD, respectively, where STD is the standard deviation of the time series.

[0071] Step 33: Use principal component analysis (PCA), nonnegative matrix factorization (NMF), and t-distributed random nearest neighbor embedding (T-SNE) to reduce the feature dimension.

[0072] Step 34: Build and train the SVM classifier model, and use the trained SVM classifier to classify unlabeled samples to obtain pseudo-labeled data.

[0073] In this embodiment, an SVM classifier is used to classify unlabeled samples. Among these predicted class labels, those with a predicted probability higher than a set threshold are considered "pseudo-labels." The training set is updated by adding these types of samples to the original training data. Then, the SVM classifier is retrained. The retrained SVM classifier is used again to classify the unlabeled samples. This process is repeated until all unlabeled samples with predicted probabilities greater than the threshold are obtained, resulting in the final training set.

[0074] In self-training, a small number of labeled samples are used to train the classifier; then, the classifier is used to classify unlabeled samples. Among these predicted class labels, those with a predicted probability higher than a set threshold are considered "pseudo-labels." The training set is updated by adding these "pseudo-label" samples to the original training data, and the classifier is retrained. The trained classifier is then used again to classify the unlabeled samples. This process is repeated until all unlabeled samples with predicted probabilities greater than the threshold are obtained, resulting in the final training set. In this embodiment, the posterior probability output of the support vector machine is used to obtain the predicted probability of pseudo-labeled samples, with the probability threshold set to 0.8.

[0075] Step 4: Perform CEEMDAN processing and scaling transformation on both the pseudo-labeled data and the pre-obtained labeled data to obtain CEEMDAN-based EEG multi-scale fuzzy entropy features, establish a fatigue state classification model, classify EEG signals for fatigue states, and distinguish between the driver's awake and fatigued states.

[0076] This step adds pseudo-labeled data to the initial training samples (i.e., labeled data) to construct a new training set; then, this new training set is used to train a fatigue state classification model. First, the pseudo-labeled data and the previously obtained labeled data are processed using CEEMDAN, selecting six IMF sequence components. Then, these IMF sequence components are scaled to extract the fuzzy entropy features of the new sequences. Finally, classification algorithms such as Random Forest, GBDT, LR, K-Nearest Neighbors, or SVM are used to construct a fatigue state classification model for classification.

[0077] After processing the pseudo-labeled EEG signal data using CEEMDAN, each IMF sequence component x(n) is obtained, with a sequence length of N. A moving window with a step size of l = 10 is determined as the sequence scaling factor, and the window is moved sequentially without overlap. The mean of the time series under each moving window is calculated as the new sequence value x. l (j), the formula is as follows:

[0078]

[0079] Given a new time series {x(n)|n=1,2,...,N}, where N is the dimension of the given time series, the calculation of fuzzy entropy requires reconstructing a new m-dimensional vector X. m (i):

[0080] X m (i)={x(i),x(i+1),...,x(i+m-1)}-x0(i) (9)

[0081] Where i = 1, 2, ..., n-m+1, x0(i) is the average value of the sequence calculated by the following formula:

[0082]

[0083] In this embodiment, m is a constant, set to 2. Furthermore, the parameters n and r in the fuzzy membership function are set to 4 and 0.2 × STD, respectively, where STD is the standard deviation of the time series.

[0084] The fuzzy entropy of new sequence values ​​is extracted to construct a fatigue state classification model. The constructed fatigue state classification model is used to classify the EEG signal into two states: awake and fatigued, to obtain the driver's state.

[0085] Example 2

[0086] This embodiment, based on the same inventive concept as Embodiment 1, provides a fatigue driving classification system using EEG multi-scale fuzzy entropy features, comprising the following modules:

[0087] The data acquisition module conducts simulated driving experiments to collect EEG data and annotates some of the data;

[0088] The preprocessing module preprocesses the EEG signals to remove artifact interference.

[0089] The feature extraction module extracts features from the preprocessed EEG signals, constructs and trains an SVM classifier, and uses the trained SVM classifier to classify unlabeled data to obtain pseudo-labeled data.

[0090] The fatigue state classification module performs CEEMDAN processing and scale transformation on both pseudo-labeled data and pre-obtained labeled data to obtain CEEMDAN-based EEG multi-scale fuzzy entropy features, establishes a fatigue state classification model, and classifies EEG signals for fatigue states.

[0091] In this embodiment, the feature extraction process of the feature extraction module includes:

[0092] Subband signals of the original EEG signal (Delta, Theta, Alpha, Beta) were obtained through wavelet packet transform. Temporal and frequency domain features of the EEG signal were extracted from these subband signals. Mobility (HM) and complexity (HC) were selected to describe the temporal features of the EEG signal and its subband signals, completing the temporal feature extraction. Fast Fourier Transform was used to calculate the relative energy and energy ratio of the original EEG signal and each subband signal. The Welch average periodogram method was used to calculate the power spectral density of each subband signal, completing the frequency domain feature extraction.

[0093] Extracting entropy features from EEG signals, including Renyi entropy, spectral entropy, and fuzzy entropy;

[0094] The process of obtaining CEED multi-scale fuzzy entropy features based on CEEMDAN in the fatigue state classification module includes:

[0095] The pseudo-labeled data and the pre-obtained labeled data are processed by CEEMDAN, multiple IMF sequence components are selected, and then the scaling transformation is performed on these IMF sequence components to extract the fuzzy entropy features of the new sequence.

[0096] In this process, the pseudo-labeled data is processed by CEEMDAN to obtain each IMF sequence component x(n), with a sequence length of N. A moving window with a step size of l is determined as the scaling factor of the sequence, and the windows are moved sequentially without overlap. The mean of the time series under each moving window is calculated as the new sequence value x. l (j), the formula is as follows:

[0097]

[0098] Where N is the dimension of the given time series {x(n)|n=1,2,...,N}.

[0099] Each module in this embodiment is used to implement each step of Embodiment 1. The detailed implementation process can be further referred to Embodiment 1, and will not be repeated here.

[0100] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope disclosed in the present invention, based on the technical solution and inventive concept of the present invention, shall fall within the scope of protection of the present invention.

Claims

1. A fatigue driving classification method based on EEG multi-scale fuzzy entropy features, characterized in that, Includes the following steps: A simulated driving experiment was conducted to collect single-channel EEG signals from the left frontal lobe of the subjects, and some of the data were labeled. Preprocessing of EEG signals removes artifact interference. Feature extraction was performed on the preprocessed EEG signals, an SVM classifier was constructed and trained, and the trained SVM classifier was used to classify the unlabeled data to obtain pseudo-labeled data. Both pseudo-labeled data and pre-obtained labeled data were processed by CEEMDAN and scale transformation to obtain CEEMDAN-based EEG multi-scale fuzzy entropy features, and a fatigue state classification model was established to classify EEG signals for fatigue states. The steps to obtain pseudo-labeled data include: An SVM classifier is trained using the feature information of labeled data obtained through prior experiments to establish a self-trained model. The trained SVM classifier is then used to predict unlabeled data to obtain pseudo-labeled data. The steps for feature extraction from preprocessed EEG signals include: The EEG signal subbands, Delta, Theta, Alpha, and Beta, were obtained through wavelet packet transform. Temporal and frequency domain features of the EEG signal were then extracted from these subbands. Mobility (HM) and complexity (HC) were selected to describe the temporal features of the EEG signal and its subbands, thus completing the temporal feature extraction. Fast Fourier Transform was used to calculate the relative energy and energy ratio of the EEG signal and each subband, and the Welch average periodogram method was used to calculate the power spectral density of each subband, completing the frequency domain feature extraction. Entropy features of EEG signals are extracted, including Renyi entropy, spectral entropy, and fuzzy entropy. The steps for obtaining CEED multiscale fuzzy entropy features based on CEEMDAN include: CEEMDAN processing was performed on pseudo-labeled data and pre-obtained labeled data, and multiple IMF sequence components were selected. Then, for multiple IMF sequence components Perform scaling transformation and extract fuzzy entropy features of the new sequence; The pseudo-labeled data, after being processed by CEEMDAN, yielded multiple IMF sequence components. Given a sequence of length N, a moving window with a step size of l is determined as the scaling factor of the sequence. The window is moved sequentially, ensuring that all moving windows do not overlap. The mean of the sequence under each moving window is calculated as the new sequence value. The formula for obtaining the answer is as follows: 。 2. The fatigue driving classification method according to claim 1, characterized in that, Preprocessing of EEG signals removes non-physiological artifacts from the EEG signals; Non-physiological artifacts include power frequency interference and noise caused by improper electrode contact with the scalp; The steps to remove non-physiological artifacts include: First, discard the first segment of the EEG signal and remove any abnormal data segments from the EEG signal. Then, the DC bias and power frequency interference are eliminated by filtering.

3. The fatigue driving classification method according to claim 2, characterized in that, The EEG signal was preprocessed, and physiological artifacts were removed from the EEG signal by using statistical thresholding and discrete wavelet transform of the coif3 wavelet function.

4. A fatigue driving classification system based on EEG multi-scale fuzzy entropy features for implementing the fatigue driving classification method based on EEG multi-scale fuzzy entropy features according to claim 1, characterized in that, include: The data acquisition module conducts a simulated driving experiment to collect single-channel EEG signals from the left frontal lobe of the subjects and labels some of the data. The preprocessing module preprocesses the EEG signals to remove artifact interference. The feature extraction module extracts features from the preprocessed EEG signals, constructs and trains an SVM classifier, and uses the trained SVM classifier to classify unlabeled data to obtain pseudo-labeled data. The fatigue state classification module performs CEEMDAN processing and scale transformation on both pseudo-labeled data and pre-obtained labeled data to obtain CEEMDAN-based EEG multi-scale fuzzy entropy features, establishes a fatigue state classification model, and classifies EEG signals for fatigue state.

Images