Brain fatigue rapid detection method based on electroencephalogram signals

Abstract

The invention discloses a brain fatigue rapid detection method based on electroencephalogram signals, which comprises the following steps: firstly, extracting neural oscillation characteristics related to brain fatigue, such as ERP (Event-Related Potential), PSD (Power Spectral Density), CFC (Cross-Frequency Coupling), waveform statistics and the like, at each electrode, sorting the importance of the neural oscillation characteristics by adopting a recursive feature cancellation (RFE) scheme, and then, extracting the neural oscillation characteristics related to brain fatigue, such as ERP (Event-Related Potential), PSD (Power Spectral Density), CFC (Cross-Frequency Coupling), CFC (Cross-Frequency Coupling), waveform statistics and the like; the method comprises the following steps of: performing quantitative evaluation on the correlation degree of each electrode and brain fatigue in sequence, screening out a plurality of brain fatigue related electrodes to be respectively used for constructing a support vector machine (SVM) classifier, and performing weighted fusion on the SVM classifier according to an area under a receiver operating characteristic curve (ROC) to form a brain fatigue rapid detection model. Besides, through samples correctly classified by the SVM, change intervals of neural oscillation characteristics are respectively calculated, 95% of boundary lines of the change intervals are used as threshold values to construct a brain fatigue grade evaluation model, and the severity of brain fatigue is evaluated according to the number of the characteristics falling into the threshold value range.

Description

1. Technical field

[0001] The present invention relates to the field of signal analysis technology, and in particular to a method for quickly detecting mental fatigue based on electroencephalogram signals. 2. Background technology

[0002] With the development of manned space technology, the average time humans stay in space has increased from a few days to several months. This long-term in-orbit flight will bring huge challenges to the cognitive and behavioral functions of the brain. The main influencing factors include: the microgravity state in space, the rapid change between day and night, the closed and narrow cabin environment, noise, radiation, and a strong workload, etc., which will strongly affect the sleep quality and emotional state of astronauts, and thus cause astronauts to easily experience mental fatigue during in-orbit flight. Since the physiological state of humans in space is different from that on the ground, it will have a unique impact on brain function, which may make the mental fatigue detection and level assessment model based on the ground simulation experiment design invalid. There is currently no effective solution.

[0003] To this end, this applicant has also conducted relevant experiments on astronauts in the Chinese space station. Based on the characteristics of neural oscillations such as event-related potentials, power spectral density and cross-frequency coupling, a rapid detection method for mental fatigue was constructed as a strategy to solve this problem.

[0004] Chinese patent CN114403897A uses noisy EEG signals to obtain clean EEG signals through preprocessing, and performs wavelet packet transform decomposition to obtain rhythmic waves, calculate the energy ratio of rhythmic waves, and input them into a pre-trained artificial neural network to judge whether fatigue is present. Chinese patent CN116327218A indirectly coordinates high-frequency activities through the combination of long-distance, low-frequency phase synchronization and local cross-frequency PAC to achieve mental fatigue judgment and degree measurement. Existing EEG-based mental fatigue detection methods all require the use of traditional 64-lead EEG signals to collect dozens of scalp EEG signals, which cannot achieve rapid detection and convenient operation, and the possibility of actual application in space environments is severely limited.

[0005] Based on the analysis of the mechanism of changes in mental fatigue during long-term on-orbit flight, the present application extracts neural oscillation features such as event-related potential (ERP), power spectral density (PSD), cross-frequency coupling (CFC) and waveform statistics at each electrode, and uses a recursive feature elimination (RFE) scheme to sort their importance. The degree of association between each electrode and mental fatigue is quantitatively evaluated in turn, and several electrodes related to mental fatigue are selected to construct a support vector machine (SVM) classifier, and the SVM classifier is weighted and fused according to the area under the receiver operating characteristic curve (ROC) (area under curve, AUC) value to form a rapid detection model for mental fatigue. In addition, the change interval of the neural oscillation feature is calculated respectively through the samples correctly classified by SVM, and the mental fatigue level evaluation model is constructed with its 95% boundary line as the threshold, and the severity of mental fatigue is evaluated according to the number of features falling within the threshold range. 3. Summary of the invention

[0006] The purpose of this application is to design a rapid detection method for mental fatigue based on EEG signals, which can realize rapid and accurate detection of mental fatigue by collecting EEG from frontal lobe Fz, F3, F4 and occipital lobe Oz, O1, O2 electrodes for analysis.

[0007] To achieve the above object, the present invention provides a method for rapid detection of mental fatigue based on EEG signals, comprising:

[0008] 30 neural oscillation features including ERP, PSD, CFC and waveform statistics were extracted at each electrode of the EEG signal;

[0009] A recursive feature elimination (RFE) scheme was used to rank the importance of 30 neural oscillation features;

[0010] The correlation between each electrode and mental fatigue was quantitatively evaluated in turn, and the frontal lobe Fz, F3, F4 and occipital lobe Oz, O1, O2 electrodes were selected;

[0011] Six SVM classifiers were weighted and fused according to the area under the receiver operating characteristic (ROC) curve (AUC) value to form a rapid detection model for mental fatigue. The change ranges of 30 neural oscillation features were calculated using the samples correctly classified by SVM, and a mental fatigue level assessment model was constructed with the 95% boundary line as the threshold. The severity of mental fatigue was assessed according to the number of features falling within the threshold range.

[0012] Preferably, 30 neural oscillation features such as ERP, PSD, CFC and waveform statistics are extracted at each electrode, including:

[0013] Through EEGLAB, we performed operations such as electrode positioning, invalid electrode deletion, re-referencing, signal filtering, trial segmentation, interpolation of bad leads, removal of bad segments, ICA analysis, and repair of eye movement components, and obtained the average ERP of each valid electrode and trial;

[0014] The three ERP-like neural oscillation features extracted at each electrode are N100 and P300 of the response inhibition Go / Nogo paradigm and P200 of the short-term memory Twoback paradigm;

[0015] Discrete Fourier calculation is realized by fast Fourier transform to obtain the power spectral density PSD of the EEG signal. PSD describes the distribution of the power of the EEG signal at different frequencies. The present invention uses the periodogram method to perform resting state EEG power spectrum analysis, and displays the joint time-frequency distribution (TFD) of the task state EEG signal through time-frequency analysis (TFA) to perform task state power spectrum analysis.

[0016] The 12 PSD-like neural oscillation features extracted at each electrode are the resting state delta, theta, alpha, beta and gamma power, the average power of ROI 1, ROI 2 and ROI 3 in the Go / Nogo task state, and the average power of ROI 1, ROI 2, ROI 3 and ROI4 in the Twoback task state;

[0017] The original EEG signal is decomposed by complex Morlet wavelet transform, and the rhythmic waves of different frequency bands of the EEG signal are decomposed to obtain the gamma amplitude locked by theta phase. The cross-correlation maximum analysis method is used to quantify and evaluate the phase-amplitude coupling.

[0018] Each electrode corresponded to the FLC, PFC, CPC, TPC, and OPC regions, and five CFC-like neural oscillation features were obtained;

[0019] A total of 10 waveform statistical neural oscillation features were extracted from the resting-state EEG signal, including sample entropy, time-domain Hjorth parameter, autoregressive (AR) spectrum and histogram.

[0020] Preferably, a recursive feature elimination (RFE) scheme is used to rank the importance of 30 neural oscillation features, including:

[0021] Construct an SVM in the training sample set, calculate the importance of each feature according to the evaluation criteria, use the weight coefficient to sort the features, remove the feature with the smallest weight from the current feature space, limit the training set to the remaining feature space, repeat the above operations, sort the importance of the features, and remove irrelevant or redundant features in turn.

[0022] Preferably, the correlation between each electrode and mental fatigue is quantitatively evaluated in turn, and the frontal lobe Fz, F3, F4 and occipital lobe Oz, O1, O2 electrodes are selected, including:

[0023] The performance of each SVM classifier was evaluated using the area under the receiver operating characteristic curve (ROC) (AUC). The probability value of the test sample was obtained according to the decision function of the SVM classifier. The closer the probability value is to 1, the greater the possibility of mental fatigue. The ROC was plotted using the probability value output by the decision function to find the demarcation threshold between the normal state and mental fatigue. After obtaining the probability value of mental fatigue of the test sample using the leave-one-out crossover method, different thresholds were used to determine whether mental fatigue occurred in the test sample.

[0024] Preferably, six SVM classifiers are weighted and fused according to the area under the receiver operating characteristic curve (AUC) value to form a rapid detection model for mental fatigue, including:

[0025] Assuming that there are N mental fatigue-related electrodes, the specific steps of weighted fusion of these N SVM classifiers are as follows: (1) The neural oscillation features of the N electrodes are recursively selected and the SVM classifier is trained. After leave-one-out cross-validation, N groups of probability values ​​are obtained, denoted as [P n (i)|n∈1,2,…,N; i=1,2,…,S], where S is the number of samples; (2) Calculate the AUC of each of the N probability values, denoted as [AUC n |n∈1,2,…,N]; (3) N groups of probability values ​​P n (i) AUC n After weighting, the fusion probability value is formed Recalculate The AUC value can be used to evaluate the performance of the mental fatigue detection model:

[0026] 4. Description of the drawings

[0027] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings required for use in the implementation mode. Obviously, the drawings described below are only some implementation modes of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0028] Figure 1 This is a model diagram for rapid detection and level assessment of mental fatigue according to an embodiment of the present invention;

[0029] Figure 2 This is the SVM-RFE algorithm process described in the embodiment of the present invention;

[0030] Figure 3 It is a weighted fusion method of the SVM classifier of the electrode related to mental fatigue described in the embodiment of the present invention;

[0031] Figure 4 are the ROC and AUC values ​​corresponding to the mental fatigue detection model at the frontal lobe Fz electrode according to the embodiment of the present invention;

[0032] Figure 5 are the ROC and AUC values ​​after the frontal lobe Fz, F3, F4 and occipital lobe Oz, O1, O2 electrodes are fused as described in the embodiment of the present invention;

[0033] Figure 6 It is the variation range of P300 of the Go / Nogo paradigm of the mental fatigue sample described in the embodiment of the present invention. 5. Specific implementation methods

[0034] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the scope of protection of the present invention.

[0035] See also Figure 1 The embodiment of the present invention provides a method for rapid detection of mental fatigue based on EEG signals, comprising:

[0036] 30 neural oscillation characteristics related to mental fatigue, including ERP, PSD, CFC and waveform statistics, were extracted at each electrode.

[0037] In this embodiment, two experiments, long-term cognitive tasks and sleep deprivation, were designed to induce mental fatigue. Subjective data and EEG signals of 28 subjects were collected, and whether the subjects entered a state of mental fatigue was determined from the perspective of subjective questionnaires and objective behavioral performance.

[0038] Through EEGLAB, we performed operations such as electrode positioning, invalid electrode deletion, re-referencing, signal filtering, trial segmentation, interpolation of bad leads, removal of bad segments, ICA analysis, and repair of eye movement components, and obtained the average ERP of each valid electrode and trial;

[0039] The brain regulates response inhibition and short-term memory functions through different pathways. The neural oscillation characteristics related to human response inhibition include N100 and P300, and the neural oscillation characteristics related to short-term memory include N100, P200 and P300. The mental fatigue induced by complete sleep deprivation, sleep-wake disorder and long-term cognitive tasks not only leads to a decrease in the amplitude of N100 and P300 in the Go / Nogo paradigm, but also leads to a decrease in the amplitude of P200 in the Twoback paradigm. Therefore, in this embodiment, the three ERP-like neural oscillation characteristics extracted at each electrode are N100 and P300 in the response inhibition Go / Nogo paradigm and P200 in the short-term memory Twoback paradigm.

[0040] Discrete Fourier calculation is realized by fast Fourier transform to obtain the power spectral density PSD of the EEG signal. PSD describes the distribution of the power of the EEG signal at different frequencies. The present invention uses the periodogram method to perform resting state EEG power spectrum analysis, and displays the joint time-frequency distribution (TFD) of the task state EEG signal through time-frequency analysis (TFA) to perform task state power spectrum analysis.

[0041] The resting power spectrum analysis showed that theta rhythmic activity was a potential indicator of the net synaptic strength of the cerebral cortex. The theta wave power increased abnormally after sleep deprivation. Mental fatigue destroyed the long-term synaptic plasticity of central neurons, leading to an imbalance in the net synaptic strength of the cerebral cortex and entering an oversaturated state. Long-term cognitive tasks also led to abnormal neuronal activity in the cerebral cortex at rest, and the power of delta, theta, alpha, beta and gamma rhythmic waves generally increased. In addition, the task power spectrum analysis showed that the ERS / ERD of the Go / Nogo and Twoback paradigms was different in the brain under fatigue. The brain regulates response inhibition and short-term memory function through different pathways. Therefore, the 12 PSD-like neural oscillation features extracted at each electrode in this embodiment were the resting delta, theta, alpha, beta and gamma power, the average power of ROI 1, ROI 2 and ROI 3 in the Go / Nogo task state, and the average power of ROI 1, ROI 2, ROI 3 and ROI4 in the Twoback task state.

[0042] The original EEG signal is decomposed by complex Morlet wavelet transform, and the rhythmic waves of different frequency bands of the EEG signal are decomposed to obtain the gamma amplitude locked by theta phase. The cross-correlation maximum analysis method is used to quantify and evaluate the phase-amplitude coupling.

[0043] The cortical communication that relies on theta-gamma PAC in the resting state of the brain does not form an absolute advantage. When the human body performs the reaction inhibition Go / Nogo task in the awake state, the occipital lobe gamma amplitude is phase-locked by the frontal lobe theta trough. When the human body performs the short-term memory Twoback task in the awake state, the temporal lobe gamma amplitude is phase-locked by the frontal lobe theta trough. However, mental fatigue has a time accumulation effect, and the phase encoding mechanism between the frontal cortexes continues to evolve to theta-gamma PAC decoupling, and the reaction inhibition and short-term memory functions are gradually impaired. When calculating the CFC-like neural oscillation characteristics at each electrode, the gamma amplitude is provided by a certain electrode itself, and the theta phase comes from a certain brain lobe. For example, the CFC-like neural oscillation characteristics of the Fz electrode refer to the average gamma amplitude of the electrode locked by the theta phase of the FLC, PFC, CPC, TPC and OPC regions. Therefore, in this embodiment, each electrode corresponds to the FLC, PFC, CPC, TPC and OPC regions, and 5 CFC-like neural oscillation characteristics are obtained respectively;

[0044] A total of 10 waveform statistical neural oscillation features were extracted from the resting-state EEG signal, including sample entropy, time-domain Hjorth parameter, autoregressive (AR) spectrum and histogram.

[0045] A recursive feature elimination (RFE) scheme was used to rank the importance of 30 neural oscillation features.

[0046] The recursive feature selection algorithm SVM-RFE is used to eliminate redundant information in the high-dimensional feature space and generate a simpler and more refined feature subset. After eliminating features according to the discriminant function and related criteria, the SVM is retrained. The coefficients of the linear SVM weight vector in this process are used as the feature sorting criteria. The specific process is as follows: Figure 2 The following are four steps: (1) constructing an SVM in the training sample set; (2) calculating the importance of each feature according to the evaluation criteria and sorting the features using the weight coefficient; (3) removing the feature with the smallest weight from the current feature space; (4) limiting the training set to the remaining feature space and repeating steps 1-4. Each cycle will sort the importance of the features and remove irrelevant or redundant features in turn.

[0047] The SVM-RFE algorithm sorts the importance of features through the kernel function and uses the weight value ω to calculate the evaluation criterion C i , Ci =(ω i ) 2 , where ω i is the i-th element of the weight vector ω, if C i If the value is the smallest, the i-th feature is eliminated, C i =(ω i ) 2 It is an approximate measure of the change in class interval in SVM. This study uses the class interval as the objective function of the SVM-RFE algorithm, and the calculation formula is as follows:

[0048]

[0049] When the i-th feature is removed, according to Taylor expansion, the change in class interval ΔJ is given by:

[0050]

[0051] In the optimal J, the first-order partial derivative is close to zero, so the above formula is equal to:

[0052] ΔJ=(Δω) 2

[0053] If the i-th feature is removed, Δω=ω i , C i =(ω i ) 2 Approximately to the above formula, when the present invention uses a nonlinear kernel function:

[0054]

[0055] Where ω -(i) is the weight vector after removing the i-th feature.

[0056] The correlation between each electrode and mental fatigue was quantitatively evaluated in turn, and the frontal lobe Fz, F3, F4 and occipital lobe Oz, O1, O2 electrodes were selected.

[0057] refer to Figure 3 , according to the area (AUC) value under the receiver operating characteristic curve (ROC), six SVM classifiers were weighted and fused to form a rapid detection model for mental fatigue. The change intervals of 30 neural oscillation features were calculated through the samples correctly classified by SVM, and the mental fatigue level assessment model was constructed with the 95% boundary line as the threshold. The severity of mental fatigue was assessed according to the number of features falling within the threshold range.

[0058] In this embodiment, according to the above method, 30 neural oscillation features were extracted at each electrode, and they were recursively selected to train SVM classifiers corresponding to 32 electrodes in turn. The leave-one-out cross-validation results showed that the most relevant electrode for mental fatigue was the frontal central Fz, such as Figure 4 As shown, its AUC value is 0.7836. According to the optimal classification threshold corresponding to the upper left corner of the ROC, it is calculated that the accuracy of the SVM classifier is 0.7843, the precision is 0.8376, the recall rate is 0.7967, and the F1 score is 0.8167.

[0059] According to the construction method of the mental fatigue detection model, starting from the electrode most relevant to mental fatigue, i.e., the central frontal lobe Fz, the SVM classifiers corresponding to the electrodes secondarily relevant to mental fatigue were weighted and fused one by one. The AUC value after each fusion was compared with the previous one to determine the number of electrodes N required to construct the mental fatigue detection model. The statistical bootstrap strategy was used to compare the AUCs of the two models: (1) bootstrapped resampling of the test samples and output probability values ​​of the two models; (2) calculating the AUC values ​​of the bootstrapped resampled data; (3) performing a paired T test on the bootstrapped resampled AUCs of the two models, and judging the significance of the difference between the two models by whether the P value was less than 0.05.

[0060] According to the correlation between each electrode and mental fatigue, that is, the AUC value of each electrode corresponding to the SVM classifier, 32 electrodes were fused in sequence starting from Fz. The first 10 fusion models of this process were evaluated. When the number of weighted fused electrodes increased to 6, the AUC value tended to stabilize and no longer increased significantly. Figure 5 As shown, the weighted fusion of the SVM classifier corresponding to the six electrodes of the frontal lobe Fz, F3, F4 and the occipital lobe Oz, O1, O2 significantly improved the performance of the mental fatigue detection model, with an AUC of 0.9299, an accuracy of 0.9338, a precision of 0.9660, a recall rate of 0.9228, and an F1 score of 0.9439. Continuing to fuse new electrodes will not have the same effect. Therefore, in order to improve the efficiency of EEG signal processing and the possibility of on-track application, the present invention finally selects the six electrodes of the frontal lobe Fz, F3, F4 and the occipital lobe Oz, O1, O2 to construct the mental fatigue detection model.

[0061] Through the rapid detection model of mental fatigue constructed by the frontal lobe Fz, F3, F4 and occipital lobe Oz, O1, O2 electrodes, 227 out of 246 samples marked as mental fatigue were correctly classified. In order to further evaluate the degree of mental fatigue, the present invention uses these 227 samples to construct a grade evaluation model. The mental fatigue grade evaluation model is consistent with the rapid detection model. It continues to use 30 neural oscillation features, with 95% of its change range as the threshold, and evaluates the mental fatigue level according to the number of features that fall within the threshold range. When calculating the change range of a neural oscillation feature, the feature values ​​of the 6 electrodes are averaged. Figure 6 The variation range of P300 in the response inhibition Go / Nogo paradigm is displayed. If the P300 of a test sample in the Go / Nogo paradigm falls within this variation range, the sample is recorded as having one mental fatigue-related feature. According to this method, 30 neural oscillation features are counted in turn to obtain the total number of mental fatigue-related features possessed by the test sample. The more mental fatigue-related features there are, the higher the level of brain fatigue.

[0062] When the subjects did not feel mental fatigue subjectively, the proportion of mental fatigue-related characteristics was less than 50%, which was a normal state, which confirmed the hierarchical evaluation model from a subjective perspective; when the subjects felt fatigue subjectively but no fatigue manifestations were shown in the objective task performance, or only fatigue manifestations were shown in the response inhibition paradigm, the proportion of mental fatigue-related characteristics was between 50% and 70%, which was mild fatigue; when the subjects felt fatigue subjectively and fatigue manifestations were shown in the objective task performance, the proportion of mental fatigue-related characteristics was between 70% and 100%, which was moderate fatigue or severe fatigue, which confirmed the hierarchical evaluation model from an objective perspective; in addition, after 36 hours of complete sleep deprivation, the proportion of mental fatigue-related characteristics exceeded 90%, which was severe fatigue, and the difference between severe fatigue and moderate fatigue could not be distinguished by subjective feelings and objective task performance alone, which further demonstrated the effectiveness of the hierarchical evaluation model.

[0063] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for rapid detection of mental fatigue based on EEG signals, characterized in that: include: 30 neural oscillation features including ERP, PSD, CFC and waveform statistics were extracted at each electrode of the EEG signal; A recursive feature elimination (RFE) scheme was used to rank the importance of 30 neural oscillation features; The correlation between each electrode and mental fatigue was quantitatively evaluated in turn, and the frontal lobe Fz, F3, F4 and occipital lobe Oz, O1, O2 electrodes were selected; Six SVM classifiers were weighted and fused according to the area under the receiver operating characteristic (ROC) curve (AUC) value to form a rapid detection model for mental fatigue. The change ranges of 30 neural oscillation features were calculated using the samples correctly classified by SVM, and a mental fatigue level assessment model was constructed with the 95% boundary line as the threshold. The severity of mental fatigue was assessed according to the number of features falling within the threshold range.

2. According to claim 1, a method for rapid detection of mental fatigue based on EEG signals, characterized in that: 30 neural oscillation features such as ERP, PSD, CFC and waveform statistics are extracted from each electrode of the EEG signal: Through EEGLAB, we performed operations such as electrode positioning, invalid electrode deletion, re-referencing, signal filtering, trial segmentation, interpolation of bad leads, removal of bad segments, ICA analysis, and repair of eye movement components, and obtained the average ERP of each valid electrode and trial; The three ERP-like neural oscillation features extracted at each electrode are N100 and P300 of the response inhibition Go / Nogo paradigm and P200 of the short-term memory Twoback paradigm; Discrete Fourier calculation is realized by fast Fourier transform to obtain the power spectral density PSD of the EEG signal. PSD describes the distribution of the power of the EEG signal at different frequencies. The present invention uses the periodogram method to perform resting state EEG power spectrum analysis, and displays the joint time-frequency distribution (TFD) of the task state EEG signal through time-frequency analysis (TFA) to perform task state power spectrum analysis. The 12 PSD-like neural oscillation features extracted at each electrode are the resting state delta, theta, alpha, beta and gamma power, the average power of ROI 1, ROI 2 and ROI 3 in the Go / Nogo task state, and the average power of ROI 1, ROI2, ROI 3 and ROI4 in the Twoback task state; The original EEG signal is decomposed by complex Morlet wavelet transform, and the rhythmic waves of different frequency bands of the EEG signal are decomposed to obtain the gamma amplitude locked by theta phase. The cross-correlation maximum analysis method is used to quantify and evaluate the phase-amplitude coupling. Each electrode corresponded to the FLC, PFC, CPC, TPC, and OPC regions, and five CFC-like neural oscillation features were obtained; A total of 10 waveform statistical neural oscillation features were extracted from the resting-state EEG signal, including sample entropy, time-domain Hjorth parameter, autoregressive (AR) spectrum and histogram.

3. According to claim 1, a method for rapid detection of mental fatigue based on EEG signals, characterized in that: A recursive feature elimination (RFE) scheme was used to rank the importance of 30 neural oscillation features: Construct an SVM in the training sample set, calculate the importance of each feature according to the evaluation criteria, use the weight coefficient to sort the features, remove the feature with the smallest weight from the current feature space, limit the training set to the remaining feature space, repeat the above operations, sort the importance of the features, and remove irrelevant or redundant features in turn.

4. According to claim 3, a method for rapid detection of mental fatigue based on EEG signals, characterized in that: A recursive feature elimination (RFE) scheme was used to rank the importance of 30 neural oscillation features: The SVM-RFE algorithm sorts the importance of features through the kernel function and uses the weight value ω to calculate the evaluation criterion C i , C i =(ω i ) 2 , where ω i is the i-th element of the weight vector ω, if C i If the value is the smallest, the i-th feature is eliminated, C i =(ω i ) 2 It is an approximate measure of the change in class interval in SVM. This study uses the class interval as the objective function of the SVM-RFE algorithm, and the calculation formula is as follows: When the i-th feature is removed, according to Taylor expansion, the change in class interval ΔJ is given by: In the optimal J, the first-order partial derivative is close to zero, so the above formula is equal to: ΔJ=(Dω) 2 If the i-th feature is removed, Δω=ω i , C i =(ω i ) 2 Approximately to the above formula, when the present invention uses a nonlinear kernel function: Where ω -(i) is the weight vector after removing the i-th feature.

5. According to claim 1, a method for rapid detection of mental fatigue based on EEG signals, characterized in that: The correlation between each electrode and mental fatigue was quantitatively evaluated in turn, and the frontal lobe Fz, F3, F4 and occipital lobe Oz, O1, O2 electrodes were selected: The performance of each SVM classifier was evaluated using the area under the receiver operating characteristic curve (ROC) (AUC). The probability value of the test sample was obtained according to the decision function of the SVM classifier. The closer the probability value is to 1, the greater the possibility of mental fatigue. The ROC was plotted using the probability value output by the decision function to find the demarcation threshold between the normal state and mental fatigue. After obtaining the probability value of mental fatigue of the test sample using the leave-one-out crossover method, different thresholds were used to determine whether mental fatigue occurred in the test sample.

6. According to claim 1, a method for rapid detection of mental fatigue based on EEG signals, characterized in that: Six SVM classifiers were weighted and fused according to the area under the receiver operating characteristic curve (AUC) value to form a rapid detection model for mental fatigue: Assuming that there are N mental fatigue-related electrodes, the specific steps of weighted fusion of these N SVM classifiers are as follows: (1) The neural oscillation features of the N electrodes are recursively selected and the SVM classifier is trained. After leave-one-out cross-validation, N groups of probability values ​​are obtained, denoted as [P n (i)|n∈1,2,…,N; i=1,2,…,S], where S is the number of samples; (2) Calculate the AUC of each of the N probability values, denoted as [AUC n |n∈1,2,…,N]; (3) N groups of probability values ​​P n (i) AUC n After weighting, the fusion probability value is formed Recalculate The AUC value can be used to evaluate the performance of the mental fatigue detection model:

Images