Fatigue state detection system based on electroencephalogram signals

Abstract

The invention belongs to the technical field of electroencephalogram signal analysis, and particularly relates to an electroencephalogram signal-based fatigue state detection system, which comprises a signal acquisition module, a signal preprocessing module, a feature extraction module and a fatigue state judgment model, the signal acquisition module is used for establishing stable low-impedance connection between each electrode and scalp by wearing an electrode cap and using conductive paste, and recording an original EEG signal; the signal preprocessing module is used for carrying out band-pass filtering and self-adaptive wave trapping on an original EEG signal and then removing physiological artifacts through self-adaptive filtering and independent component analysis; the feature extraction module is used for extracting artificially designed frequency domain, time domain and nonlinear features from the preprocessed signals, automatically learning deep features through 1D-CNN and an attention mechanism, and splicing the two types of features to form a fusion feature vector; the fatigue state judgment model calculates a fatigue score based on the fusion feature vector, and obtains a comprehensive score through time smoothing and trend analysis.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of electroencephalogram signal analysis, and particularly relates to a fatigue state detection system based on electroencephalogram signals. BACKGROUND

[0002] Long hours of work, study or high-intensity tasks often lead to fatigue, especially in tasks that require full concentration, fatigue is a key factor that triggers accidents and errors. This state not only weakens the individual's cognitive and decision-making abilities, but also slows down the reaction speed, thereby increasing the probability of accidents. Therefore, timely and accurate monitoring of fatigue, especially in its early stages, is crucial for improving safety and work efficiency.

[0003] Compared with other methods, electroencephalogram (EEG) signals can directly map the neural activity of the brain, providing more objective and real-time data support, especially in the early detection of fatigue, it can give an early warning before the symptoms of fatigue appear, giving us valuable response time.

[0004] However, the above-mentioned technologies often have the following defects: most systems rely on a single filtering or artifact removal method, which is difficult to completely remove mixed physiological noise; feature engineering is mostly limited to a single approach of manual design or automatic extraction, failing to fully exploit the complementary advantages of both types of features; model decision is usually based on instantaneous analysis and fixed threshold, lacking consideration of the progressive nature of fatigue and individualized adaptation ability, resulting in limited detection accuracy, high false positive rate, and obvious shortcomings in real-time performance, stability and user experience.

[0005] Therefore, the present application provides a fatigue state detection system based on electroencephalogram signals. SUMMARY

[0006] In order to make up for the deficiencies of the prior art and solve at least one technical problem raised in the background art.

[0007] The technical solution adopted by the present application to solve its technical problems is: the fatigue state detection system based on electroencephalogram signals, comprising a signal acquisition module, a signal preprocessing module, a feature extraction module and a fatigue state judgment model.

[0008] The signal acquisition module establishes a stable low-impedance connection for each electrode and the scalp by wearing an electrode cap and using conductive paste, and records the original EEG signal with high quality, providing the system with original physiological data.

[0009] The signal preprocessing module is used to remove high and low frequency noise by band-pass filtering the original EEG signal, eliminate power frequency interference by adaptive notch, and remove physiological artifacts by adaptive filtering and independent component analysis, and finally reconstruct the pure electroencephalogram signal.

[0010] The feature extraction module is used to extract artificially designed frequency domain, time domain and nonlinear features from the preprocessed signal, and automatically learn deep features through 1D-CNN and attention mechanism, and the two types of features are spliced to form a fusion feature vector;

[0011] The fatigue state judgment model calculates a fatigue score based on the fusion feature vector, obtains a comprehensive score through time smoothing and trend analysis, and maps the continuous score to four levels of sober, mild fatigue, moderate fatigue and severe fatigue through threshold determination.

[0012] Preferably, the specific process of the signal acquisition module includes the following steps:

[0013] A1, guide the subject into an electromagnetic shielding room, keep a comfortable sitting posture, briefly explain the whole acquisition process to the subject to alleviate nervousness, and at the same time ensure that the indoor light is soft and the environment is quiet;

[0014] A2, use a flexible ruler to accurately measure the key head circumference reference points of the subject from the nasal root to the occipital protuberance, then put on a size-matched electrode cap and fix the chin strap;

[0015] A3, for each electrode point covered with hair, use a blunt injection needle to gently push the hair away from the scalp, and then inject a sufficient amount of conductive paste into the scalp surface through the electrode tube;

[0016] A4, continuously use the needle to perform small-range gentle circular massage on the scalp to remove air bubbles and fill the hair root gaps, and at the same time monitor the channel impedance in real time through the software until it is stably reduced to below 10kΩ;

[0017] A5, check the signal waveform of each channel comprehensively, confirm that the baseline is smooth and there is no severe drift or large motion artifact, start formal recording, and obtain the electroencephalogram signal x(t).

[0018] Preferably, the specific process of the signal preprocessing module includes the following steps:

[0019] B1, use a zero-phase Butterworth band-pass filter to remove DC drift and high-frequency noise and retain the effective EEG frequency band;

[0020] B2, use an adaptive notch filter to dynamically adjust the parameters to adapt to frequency changes and remove power line interference;

[0021] B3, eliminate physiological artifact interference in the EEG signal, including eye movement EOG, muscle activity EMG, electrocardiogram ECG, sweat EDA and skin potential GSR, first remove part of the artifact through adaptive filtering, and then separate the remaining artifact through ICA.

[0022] Preferably, the detailed implementation method of step B3 is as follows:

[0023] B31, track and remove ECG and EOG artifact components with strong periodicity through adaptive filtering;

[0024] B32, decompose the signal into statistically independent components through FastICA algorithm to identify and remove residual EMG, EDA, GSR artifact components;

[0025] B33, calculate the signal-to-noise ratio SNR and mean square error MSE of the signal before and after artifact removal, and ensure that the signal after artifact removal retains effective EEG information.

[0026] Preferably, the specific process of the feature extraction module includes the following steps:

[0027] C1, artificial feature extraction: purposefully select and calculate quantitative indicators with clear physiological significance from the electroencephalogram signal to capture specific pattern changes related to fatigue;

[0028] C2, deep learning automatic feature extraction: automatically learn and extract abstract feature representations in the electroencephalogram signal through a multi-layer neural network to capture complex fatigue-related patterns that are difficult for humans to design;

[0029] C3, feature fusion: concatenate artificial features and deep learning automatic features to form a fusion feature vector with more comprehensive and complementary information.

[0030] Preferably, the detailed implementation method of step C1 is as follows:

[0031] C11, extract frequency domain features, including frequency band power features, frequency band power ratio features, and spectral entropy features;

[0032] C12, extract time domain features, including statistical features and Hjorth parameters;

[0033] C13, extract nonlinear dynamic features, including sample entropy and fuzzy entropy.

[0034] Preferably, the detailed implementation method of step C2 is as follows:

[0035] C21, extract 1D-CNN features, and the network architecture contains three convolutional blocks;

[0036] C22, attention mechanism enhancement.

[0037] Preferably, the specific process of the fatigue state judgment model includes the following steps:

[0038] D1, calculate the instantaneous fatigue score SCORE of the current electroencephalogram signal using a nonlinear transformation formula ins(t), which preliminarily reflects the current fatigue state;

[0039] D2, current instantaneous fatigue score SCORE ins (t) is subjected to time smoothing processing, transient fluctuations are inhibited, and the stability and reliability of the fatigue score are improved;

[0040] D3, a short-term trend factor delta is introduced short , a deep analysis of fatigue trend is performed, a dynamic change process of fatigue degree is captured, and a comprehensive fatigue score SCORE fin (t) is calculated;

[0041] D4, a fatigue level standard is defined, including: level 0-wakefulness, level 1-mild fatigue, level 2-moderate fatigue, and level 3-severe fatigue;

[0042] D5, a determination function is set to determine the level state of the comprehensive fatigue score, which provides support for progressive grading response and risk control.

[0043] The beneficial effects of the present application are as follows:

[0044] 1. The fatigue state detection system based on the electroencephalogram signal provided by the present application closely integrates precise acquisition and multi-level algorithm processing to build a complete closed loop from signal acquisition to state judgment. The two-stage de-noising strategy of "adaptive filtering + ICA" effectively solves the problem of mixed physiological artifacts, and the artificial designed features and deep learning automatic features are fused, which not only retains the explainability of expert knowledge, but also plays the mining ability of neural network to complex patterns, significantly improving the accuracy and robustness of fatigue recognition.

[0045] 2. The fatigue state detection system based on the electroencephalogram signal provided by the present application can capture the progressive process of fatigue through the continuous scoring mechanism based on time smoothing and trend analysis, avoid sharp jumps in state determination, and be more in line with physiological laws, while ensuring detection sensitivity and effectively reducing false positive rate, thereby providing reliable technical support for fatigue monitoring in actual scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0046] The present application will be further described below in conjunction with the accompanying drawings.

[0047] Figure 1 is a system flowchart of the present application. DETAILED DESCRIPTION

[0048] In order to make the technical means, creative features, purposes and effects realized by the present application easy to understand, the present application will be further described below in conjunction with specific embodiments.

[0049] As Figure 1As shown, the fatigue state detection system based on electroencephalogram signals provided by the application comprises a signal acquisition module, a signal preprocessing module, a feature extraction module and a fatigue state judgment model.

[0050] The signal acquisition module establishes a stable low-impedance connection for each electrode and the scalp by wearing an electrode cap and using conductive paste, records the original EEG signal with high quality, and provides the original physiological data for the system.

[0051] The signal preprocessing module is used for band-pass filtering of the original EEG signal to remove high and low frequency noise, using adaptive notch to eliminate power frequency interference, and then removing physiological artifacts through adaptive filtering and independent component analysis to finally reconstruct the pure electroencephalogram signal.

[0052] The feature extraction module is used for extracting artificially designed frequency domain, time domain and nonlinear features from the preprocessed signal, and simultaneously automatically learning deep features through 1D-CNN and attention mechanism, and splicing the two types of features to form a fusion feature vector.

[0053] The fatigue state judgment model calculates the fatigue score based on the fusion feature vector, obtains the comprehensive score through time smoothing and trend analysis, and maps the continuous score to four levels of wakefulness, mild fatigue, moderate fatigue and severe fatigue through threshold determination.

[0054] As a specific embodiment of the application, the specific process of the signal acquisition module comprises the following steps:

[0055] A1, guide the subject into an electromagnetic shielding room, keep a comfortable sitting posture, briefly explain the whole acquisition process to the subject to alleviate nervousness, and at the same time ensure that the indoor light is soft and the environment is quiet;

[0056] A2, use a flexible ruler to accurately measure the key head circumference reference points of the subject from the nasal root to the occipital protuberance, and then wear a size-matched electrode cap and fix the chin strap;

[0057] A3, for each electrode point covered with hair, a blunt injection needle tube is used to gently push the hair away from the scalp, and then a sufficient amount of conductive paste is accurately injected onto the scalp surface through the electrode tube;

[0058] A4, continuously use the needle to perform small-range gentle circular massage on the scalp to remove air bubbles and fill the hair root gap, and at the same time monitor the channel impedance in real time through software until it is stably reduced to below 10kΩ;

[0059] A5, check the signal waveform of each channel comprehensively, confirm that the baseline is stable and there is no severe drift or large motion artifact, start formal recording, and obtain the electroencephalogram signal x(t).

[0060] As a specific embodiment of the present application, the specific process of the signal preprocessing module includes the following steps:

[0061] B1, a zero-phase Butterworth band-pass filter is used to remove DC drift and high-frequency noise and retain the effective EEG frequency band (0.5-50 Hz);

[0062] The filter transfer function is:

[0063]

[0064] The filtered signal is:

[0065] x bp (t)=BPF(x(t))

[0066] Where f c is the cutoff frequency (the low cutoff frequency is 0.5 Hz and the high cutoff frequency is 50 Hz), n=4 is the filter order, and BPF represents the band-pass filtering operation;

[0067] B2, an adaptive notch filter is used to dynamically adjust parameters to adapt to frequency changes and remove power line frequency interference (50 Hz);

[0068] The notch filter frequency response is:

[0069]

[0070] The filtered signal is:

[0071] x notch (t)=NotchFilter(x bp (t))

[0072] Where f0=50 Hz is the notch frequency, f s =250 Hz is the sampling rate, and p=0.95 is the control bandwidth, z is the Z-transform variable, and NotchFilter represents the notch filtering operation;

[0073] B3, physiological artifact interference in the EEG signal is removed, including eye movement EOG, muscle activity EMG, electrocardiogram ECG, sweat EDA and skin potential GSR, part of the artifacts are removed by adaptive filtering, and the remaining artifacts are separated by ICA.

[0074] As a specific embodiment of the present application, the detailed implementation method of step B3 is as follows:

[0075] B31, ECG and EOG artifact components with strong periodicity are removed by adaptive filtering tracking;

[0076] The adaptive filter uses the LMS least mean square adaptive algorithm:

[0077]

[0078] The error signal after removing artifacts is:

[0079] e(t) = x notch (t) - y af (t)

[0080] Weight update:

[0081] w(t+1) = w(t) + q e(t) x ref (t)

[0082] The output of this step is:

[0083] x af (t) = e(t)

[0084] where w(t) is the weight vector, is the vector transpose operator, x ref (t) is the reference signal (the reference signal for ECG and EOG can be obtained from an additional channel or estimated from the signal), q = 0.01 is the step size;

[0085] B32, decompose the signal into statistically independent components by FastICA algorithm to identify and remove residual EMG, EDA, GSR artifact components;

[0086] The ICA model is: X = AS

[0087] By maximizing non-Gaussianity, estimate the demixing matrix W = A -1 , get independent components U = WX;

[0088] Identify artifact components based on the spatiotemporal features of the components, including:

[0089] EOG component: high amplitude, high forehead channel correlation;

[0090] EMG component: high frequency activity (20-100 Hz);

[0091] ECG component: periodic spikes;

[0092] EDA / GSR component: low frequency drift (less than 0.5 Hz);

[0093] Set the artifact components to zero, remove the artifact components, and reconstruct the signal, let U artifacts be the subset of artifact components, and the reconstructed signal is:

[0094] X clean = A(U - U artifacts )

[0095] The output signal after pretreatment is:

[0096] y(t)=X clean

[0097] Wherein, X is a multi-channel signal matrix (each row is a channel x af (t)), A is a mixing matrix, and S is an independent component matrix;

[0098] B33, calculate the signal-to-noise ratio (SNR) and mean square error (MSE) of the signal before and after deartifacting, and ensure that the signal after deartifacting retains effective EEG information.

[0099] As a specific embodiment of the application, the specific process of the feature extraction module includes the following steps:

[0100] C1, artificial feature extraction: purposefully selecting and calculating quantitative indicators with clear physiological significance from the electroencephalogram signal to capture specific pattern changes related to fatigue;

[0101] C2, deep learning automatic feature extraction: automatically learning and extracting abstract feature representations in the electroencephalogram signal through a multi-layer neural network to capture complex fatigue-related patterns that are difficult for humans to design;

[0102] C3, feature fusion: concatenating the artificial features and the deep learning automatic features to form a fusion feature vector that is more comprehensive and complementary, in order to improve the accuracy and robustness of fatigue state recognition.

[0103] As a specific embodiment of the application, the detailed implementation method of step C1 is as follows:

[0104] C11, extract frequency domain features, including frequency band power features, frequency band power ratio features, and spectral entropy features;

[0105] The frequency band power features are specifically:

[0106] Four specific electroencephalogram frequency bands are defined, namely theta waves of 4-8 Hz, alpha waves of 8-15 Hz, beta waves of 15-30 Hz, and gamma waves of 30-50 Hz, and the power P θ , P α , P β , P γ of each frequency band is calculated.

[0107] Calculate the power of each specific frequency band:

[0108]

[0109] Wherein, P band represents the power P θ , Pα , P β , P γ , Y(f) is the Fourier transform of the pre-processed signal y(t), f low and f high are the upper and lower limit frequencies of each frequency band;

[0110] The frequency band power ratio feature is specifically:

[0111] R (α+θ) / β = (P α + P θ ) / P β

[0112] R α / β = P α / P β

[0113] R (θ+α) / (α+β) = (P θ + P α ) / (P α + P β )

[0114] The spectral entropy feature is specifically:

[0115]

[0116] wherein f min = 0.5 Hz, f max = 50 Hz, and p(f) is the normalized power spectrum;

[0117]

[0118] C12, extracting time domain features, including statistical features and Hjorth parameters;

[0119] The statistical features are specifically: mean μ, variance σ 2 , skewness SK, and kurtosis KU;

[0120]

[0121] wherein N is the number of sampling points in the time window;

[0122] The Hjorth parameters are specifically: activity B, movement M, and complexity Y;

[0123] B = σ 2

[0124] M = σ y ' / σ

[0125] Y = σ y " / σ y '

[0126] wherein σ, σ y , σ y are the standard deviation of the pre-processed signal y(t), the first derivative signal dy(t) / dt, the second derivative signal d 2 y(t) / dt 2 y(t) / dt

[0127] C13, extracting nonlinear dynamic features, including sample entropy and fuzzy entropy;

[0128] The sample entropy is specifically:

[0129] SampEn(m, r, N) = -ln[A m (r) / B m (r)]

[0130] wherein m is the template dimension, r is the similarity tolerance, B m (r) is the number of matched m-point templates, and A m (r) is the number of matched n+1-point templates.

[0131] The fuzzy entropy is specifically:

[0132]

[0133] wherein:

[0134]

[0135] wherein d ij m is the distance between vectors.

[0136] As a specific embodiment of the present application, the detailed implementation method of step C2 is as follows:

[0137] C21, extracting 1D-CNN features, the network architecture contains three convolutional blocks;

[0138] Convolutional block 1: 64 7x1 filters, ReLU activation, 2x1 max pooling;

[0139] Convolutional block 2: 128 5x1 filters, ReLU activation, 2x1 max pooling;

[0140] Convolutional block 3: 256 3x1 filters, ReLU activation, global average pooling;

[0141] The output deep feature vector is: f cnn ∈R 512 , R 512 represents a 512-dimensional real vector space;

[0142] C22, attention mechanism enhancement;

[0143] Channel attention is:

[0144]

[0145] Temporal attention is:

[0146] u t = tanh(W h ·h t +b h )

[0147]

[0148] Final attention feature is:

[0149]

[0150] Where h c (t) is the feature of the cth channel at the tth time point, W e , W h , W u are weight matrices, and b h is a bias term.

[0151] As a specific embodiment of the present application, the detailed implementation method of step C3 is as follows:

[0152]

[0153] Where F hand is the artificially designed feature vector (including all the artificially designed features extracted above), F deep is the automatically extracted feature vector by deep learning, F fused is the fusion feature vector, represents a d h -dimensional real vector space, represents a d d -dimensional real vector space, represents a d h +d d -dimensional real vector space.

[0154] As a specific embodiment of the present application, the specific process of the fatigue state judgment model includes the following steps:

[0155] D1, using a nonlinear transformation formula to calculate the instantaneous fatigue score SCORE ins (t) of the current electroencephalogram signal, which preliminarily reflects the current fatigue state;

[0156]

[0157] wherein E1, E2 are two-layer weight matrices, b1, b2 are bias terms, sigmoid is an activation function for compressing scores to the range of (0, 1);

[0158] D2, the current instantaneous fatigue score SCORE ins (t) is subjected to time smoothing processing, transient fluctuations are suppressed, and the stability and reliability of the fatigue score are improved;

[0159] SCORE smo (t) = τ·SCORE ins (t) + (1-τ)·SCORE smo (t-1)

[0160] wherein τ is a smoothing coefficient, and the value range is 0<τ≤1;

[0161] When τ is close to 1, the system has a higher weight for the current measurement value, the response speed is fast, but the noise resistance is weak; when τ is close to 0, the system has a higher weight for the historical value, the response speed is slow, but the noise resistance is strong;

[0162] D3, a short-term trend factor δ is introduced short , the fatigue trend is deeply analyzed, the dynamic change process of the fatigue degree is captured, and the comprehensive fatigue score SCORE fin (t) is calculated;

[0163]

[0164] SCORE fin (t) = SCORE smo (t) + ε·δ short

[0165] wherein k is the number of historical time points considered, usually 3, i=1, 2, …, k, and ε is a trend weight coefficient (usually 0.1-0.2).

[0166] As a specific embodiment of the present application, the specific process of the fatigue state judgment model further includes the following steps:

[0167] D4, define fatigue level standards, including: level 0-wake, level 1-mild fatigue, level 2-moderate fatigue, level 3-severe fatigue;

[0168] D5, set a judgment function to judge the level state of the comprehensive fatigue score, and provide support for gradual grading response and risk control;

[0169]

[0170]

[0171] The above-mentioned front, back, left, right, up, down are all based on the drawings in the specification Figure 1 as the standard, the side of the device facing the observer is defined as front, the left side of the observer is defined as left, and so on.

[0172] In the description of the present application, it should be understood that the terms "center", "longitudinal", "transverse", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the scope of protection of the present application.

[0173] The basic principles, main features and advantages of the present application are shown and described above. It should be understood by those skilled in the art that the present application is not limited by the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A fatigue state detection system based on electroencephalogram signals, characterized by: The system comprises a signal acquisition module, a signal preprocessing module, a feature extraction module and a fatigue state judgment model. The signal acquisition module establishes a stable and low-impedance connection for each electrode and the scalp by wearing an electrode cap and using conductive paste, and records the original EEG signal with high quality to provide the original physiological data for the system. The signal preprocessing module is used for band-pass filtering of the original EEG signal to remove high and low frequency noise, using adaptive notch filtering to eliminate power frequency interference, and then removing physiological artifacts through adaptive filtering and independent component analysis to finally reconstruct pure EEG signals. The feature extraction module is used for extracting artificial designed frequency domain, time domain and nonlinear features from the preprocessed signals, and simultaneously automatically learning deep features through 1D-CNN and attention mechanism, and splicing the two types of features to form a fusion feature vector. The fatigue state judgment model calculates the fatigue score based on the fusion feature vector, obtains the comprehensive score through time smoothing and trend analysis, and maps the continuous score to four levels of awake, mild fatigue, moderate fatigue and severe fatigue through threshold determination. 2.The brain electrical signal based fatigue state detection system of claim 1, wherein: The specific process of the signal acquisition module comprises the following steps: A1, guide the subject to enter the electromagnetic shielding room, keep a comfortable sitting posture, and briefly explain the whole acquisition process to the subject to alleviate nervousness, and at the same time ensure that the indoor light is soft and the environment is quiet; A2, use a flexible ruler to accurately measure the key head circumference reference points of the subject from the nasal root to the occipital protuberance, and then wear a size-matched electrode cap and fix the chin strap; A3, for each electrode point covered with hair, a blunt injection needle tube is used to gently push the hair away from the scalp, and then a sufficient amount of conductive paste is accurately injected onto the scalp surface through the electrode tube; A4, continuously use the needle to perform a small range of gentle circular massage on the scalp to remove air bubbles and fill the hair root gap, and at the same time, the impedance of the channel is monitored in real time through the software until it is stably reduced to below 10kΩ; A5, check the signal waveform of each channel comprehensively, confirm that the baseline is stable and free of severe drift or large motion artifacts, and start formal recording to obtain the EEG signal x(t). 3.The brain electrical signal based fatigue state detection system of claim 2, wherein: The specific process of the signal preprocessing module comprises the following steps: B1, a zero-phase Butterworth band-pass filter is used to remove DC drift and high-frequency noise and retain the effective EEG frequency band; The filter transfer function is: The filtered signal is: x bp (t) = BPF(x(t)) where f c is the cutoff frequency, n = 4 is the filter order, and BPF denotes a bandpass filtering operation. B2, an adaptive notch filter is used to dynamically adjust the parameters to adapt to frequency changes and remove power line interference; The notch filter frequency response is: The filtered signal is: x notch (t) = NotchFilter(x bp (t)) where f0= 50 Hz is the notch frequency, f s = 250 Hz is the sampling rate, p = 0.95 is the control bandwidth, fz is the Z-transform variable, and NotchFilter denotes the notch filtering operation. B3, physiological artifact interference in the EEG signal is removed, including eye movement EOG, muscle activity EMG, electrocardiogram ECG, sweat EDA and skin potential GSR, first through adaptive filtering to remove part of the artifacts, and then through ICA to separate the remaining artifacts.

4. The electroencephalogram-based fatigue state detection system according to claim 3, characterized in that: The detailed implementation method of step B3 is as follows: B31, track and remove the ECG and EOG artifact components with strong periodicity through adaptive filtering; The adaptive filter uses the LMS least mean square adaptive algorithm: y af (t) = w T (t) · x ref (t) The error signal after removing the artifacts is: e(t) = x notch (t) - y af (t) The weight update is: w(t + 1) = w(t) + q - e(t) - x ref (t) The output of this step is: x af (t) = e(t) where w(t) is a weight vector, T is a vector transposition operator, x ref (t) is a reference signal, q = 0.01 is a step size; B32, decompose the signal into statistically independent components by FastICA algorithm to identify and remove residual EMG, EDA, GSR artifact components; The ICA model is: X = AS By maximizing the non-Gaussianity, the unmixing matrix W = A is estimated -1 and the independent components U = WX are obtained. Identify artifact components based on the spatiotemporal features of components, including: EOG component: high amplitude, high forehead channel correlation; EMG component: high frequency activity; ECG component: periodic spikes; EDA / GSR component: low frequency drift; zeroing the artifact components, removing the artifact components, reconstructing the signal, let U artifacts be the reconstructed signal, and let A be the artifact components. The artifact components are a subset of the signal components, i.e. A is a subset of U. The artifact components are removed from the signal components to reconstruct the signal, i.e. U = U - A. X clean = A(U - U artifacts ) Then the output signal after preprocessing is: y(t) = X clean Wherein, X is a multi-channel signal matrix, A is a mixing matrix, S is an independent component matrix; B33, calculate the signal-to-noise ratio SNR and mean square error MSE of the signal before and after deartifacting to ensure that the signal after deartifacting retains effective EEG information.

5. The electroencephalogram-based fatigue state detection system according to claim 4, characterized in that: The specific process of the feature extraction module includes the following steps: C1, artificial feature extraction: purposefully select and calculate quantitative indicators with clear physiological significance from the electroencephalogram signal to capture specific pattern changes related to fatigue; C2, deep learning automatic feature extraction: automatically learn and extract abstract feature representations in the electroencephalogram signal through a multi-layer neural network to capture complex fatigue-related patterns that are difficult to design artificially; C3, feature fusion: concatenate artificial features and deep learning automatic features to form a fusion feature vector that is more comprehensive and complementary.

6. The electroencephalogram-based fatigue state detection system according to claim 5, characterized in that: The detailed implementation method of step C1 is as follows: C11, extract frequency domain features, including frequency band power features, frequency band power ratio features, and spectral entropy features; The frequency band power features are specifically: Four specific brain electrical frequency bands are defined, respectively, theta waves of 4-8 Hz, alpha waves of 8-15 Hz, beta waves of 15-30 Hz, and gamma waves of 30-50 Hz, and the power P of each frequency band is calculated θ , α , β , γ ; Calculate the power of each specific frequency band: where P band represents the power of a specific frequency band P θ , P α , P β , P γ , Y(f) is the Fourier transform of the pre-processed signal y(t), f low and f high are the upper and lower frequencies of each frequency band. The frequency band power ratio features are specifically: R (α+θ) / β = (P α + P θ ) / P β R α / β = P α / P β R (θ+α) / (α+β) = (P θ + P α ) / (P α + P β ) The spectral entropy features are specifically: where f min = 0.5 Hz, f max = 50 Hz, o(f) is the normalized power spectrum; C12, extract time domain features, including statistical features and Hjorth parameters; The specific statistical characteristics are: mean μ and variance σ. 2 Skewness (SK) and kurtosis (KU); Wherein, N is the number of sampling points in the time window; The Hjorth parameters are specifically: activity B, movement M, and complexity Y; B = σ 2 M = σ y ' / σ Y = σ y ” / σ y ′ wherein σ, σ y ′, σ y are the standard deviations of the pre-processed signal y(t), the first derivative signal dy(t) / dt, the second derivative signal d 2 y(t) / dt 2 y(t) / dt, respectively. C13, extract nonlinear dynamic features, including sample entropy and fuzzy entropy; The sample entropy is specifically: SampEn(m, r, N) = -ln[A m (r) / B m (r)] where m is the template dimension, r is the similarity tolerance, B m (r) is the number of templates matching m points, A m (r) is the number of templates matching m+1 points; The fuzzy entropy is specifically: Wherein: where d ij m is the distance between vectors.

7. The electroencephalogram-based fatigue state detection system according to claim 6, characterized in that: The detailed implementation method of step C2 is as follows: C21, extract 1D-CNN features, the network architecture contains three convolutional blocks; Convolutional block 1: 64 7x1 filters, ReLU activation, 2x1 max pooling; Convolutional block 2: 128 5x1 filters, ReLU activation, 2x1 max pooling; Convolutional block 3: 256 3x1 filters, ReLU activation, global average pooling; The output depth feature vector is: f cnn ∈R 512 C22, attention mechanism enhancement; Channel attention is: Time attention is: u t = tanh(W h · h t + b h ) The final attention feature is: where h c (t) is the feature of the cth channel at the tth time point, W e , W h , W u is the weight matrix, and b h is the bias term.

8. The electroencephalogram-based fatigue state detection system according to claim 7, characterized in that: The detailed implementation method of step C3 is as follows: wherein F hand is a feature vector extracted by manual design, deep is a feature vector extracted by deep learning, fused is a fusion feature vector.

9. The electroencephalogram-based fatigue state detection system according to claim 8, characterized in that: The specific process of the fatigue state judgment model includes the following steps: D1. Calculate the instantaneous fatigue score SCORE of the current electroencephalogram signal using a non-linear transformation formula ins (t), which initially reflects the current fatigue state; Wherein, E1, E2 are two layers of weight matrices, b1, b2 are bias terms, sigmoid is an activation function used to compress the score to the range (0, 1); D2, the current instantaneous fatigue score SCORE ins (t) time smoothing is performed to suppress transient fluctuations and improve the stability and reliability of the fatigue score; SCORE smo (t) = τ · SCORE ins (t) + (1 - τ) · SCORE smo (t - 1) Wherein, τ is a smoothing coefficient, the value range is 0 < τ ≤ 1; D3, introduce short-term trend factor δ short , in-depth analysis of fatigue trend, capture the dynamic change process of fatigue degree, and calculate the comprehensive fatigue score SCORE fin (t); SCORE fin (t) = SCORE smo (t) + ε·δ short Wherein, K is the number of historical time points considered, usually take 3, i = 1, 2, …, K, ε is the trend weight coefficient.

10. The electroencephalogram-based fatigue state detection system of claim 9, wherein: The specific process of the fatigue state judgment model also includes the following steps: D4, define the fatigue level standard, including: level 0-wake, level 1-mild fatigue, level 2-moderate fatigue, level 3-severe fatigue; D5, set the decision function to determine the level state of the comprehensive fatigue score, and provide support for the progressive grading response and risk control;

Images