A method and system for detecting fatigue state across subjects based on multi-scale EEG features

Abstract

The present invention proposes a method and system for detecting fatigue status across subjects based on multi-scale EEG features, studies a multi-scale fusion EEG feature extraction method, and establishes a cross-subject fatigue detection model based on an improved FLDANN network. In the cross-subject fatigue detection stage, the fatigue EEG data of existing subjects are used to identify the fatigue status of target subjects through the improved FLDANN network. Not only is the cost of acquiring fatigue data of target subjects used for training models reduced, but solutions are also proposed for the one-sidedness and information missing problems of extracted fatigue EEG features, as well as the domain imbalance problem caused by multi-source domains and single target domains faced in cross-subject detection, effectively improving the classification performance.

Description

Technical Field

[0001] The present invention relates to the technical field of electroencephalogram (EEG) signal processing, and in particular to a method and system for cross-subject fatigue detection based on multi-scale EEG features. Background Art

[0002] The rapid development of science and technology has greatly promoted the improvement and development of my country's economic level and productivity, and promoted the gradual transformation of my country's industrial situation from labor-intensive industries that mainly consume physical strength to service-oriented industries that mainly rely on brain power and wisdom; on the other hand, with the rapid development of the economy, people's pace of life is getting faster and faster, and the competitive pressure they face is getting greater and greater. Overtime work has become the norm in many industries. Long-term and intensive use of the brain can easily cause brain fatigue, which has become a sub-healthy state that seriously affects people's normal life. Long-term brain fatigue can cause physiological diseases such as heart disease and palpitations, and may also cause mental diseases such as insomnia, anxiety, and depression, which seriously affect physical and mental health. Some people even die from excessive brain fatigue. Brain fatigue has become one of the important issues that seriously threaten people's health and safety. Therefore, it is very important and necessary to detect brain fatigue status in a timely manner and give feedback.

[0003] EEG signals have high temporal resolution and are rich in information that can objectively reflect changes in brain states. At present, most EEG-based fatigue detection models are targeted at a single target subject, that is, the personal data of the target subject is used to train the model and test. However, the establishment of a traditional personal model requires a large amount of training data for the target subject, and for each subject, the induced fatigue EEG data must be obtained after the same experimental duration, and then the data must be labeled. If fatigue detection is carried out in this way, data collection will take a lot of time, and the long duration of fatigue induction will also have a negative emotional impact on the subjects participating in the experiment, and in severe cases, it will affect the quality of the EEG signal. Due to the unstable characteristics of EEG signals and large individual differences, the recognition accuracy of cross-subject fatigue detection is low and cannot meet the needs of the online fatigue assessment system. Therefore, it is very important to use the EEG data of existing subjects to establish a cross-subject model. Summary of the invention

[0004] The present invention proposes a method and system for detecting fatigue state across subjects based on multi-scale EEG features. By studying multi-scale fused EEG features and using an improved adversarial domain adaptation network FLDANN, the problem of decreased classification effect of DANN in multi-source domains and single target domains is solved.

[0005] To achieve the above purpose, the present invention proposes a method for detecting fatigue state across subjects based on multi-scale EEG features, which is special in that it includes the following steps:

[0006] Step S1: Collecting EEG data of the subject in fatigue state and non-fatigue state through EEG acquisition equipment;

[0007] Step S2: filtering the EEG signal to extract the sub-band EEG signal between 0-30 Hz;

[0008] Step S3: segmenting the sub-band EEG signal data to obtain EEG signal samples under different fatigue states;

[0009] Step S4: extracting multi-scale time domain and frequency domain features of the EEG signal sample, and fusing them through a feature sorting selection method to calculate feature fusion data;

[0010] Step S5: Based on the feature fusion data, a cross-subject fatigue detection model is established to identify and detect the fatigue state of the target subject.

[0011] Preferably, the different fatigue states in step S1 are induced by a 2-back experiment, and the steps of the 2-back experiment are specifically as follows:

[0012] Step S11: The subject takes a pre-practice screening test, and when the accuracy rate reaches 80%, he / she is determined as a formal subject, and then formal training is carried out;

[0013] Step S12: the formal subject performs at least two rounds of 2-back training, the fatigue state of the formal subject is determined according to the results of the 2-back training, and EEG data under different fatigue states are collected.

[0014] Preferably, in step S3, the sub-band EEG signal collected for a long time is divided into one EEG signal sample every 500 sampling points at a sampling frequency of 500 Hz, and there is no overlap between the samples.

[0015] Preferably, step S4 specifically includes:

[0016] Step S41: Calculate the time domain features of the EEG signal samples Y={y(1), y(2), ..., y(t)} by using Hjorth parameters, and extract the activity, mobility and complexity of the EEG signal samples under different states;

[0017]

[0018]

[0019]

[0020] Where T represents the number of EEG signal samples, represents the average number of EEG signal samples, Activity' represents the first derivative of Activity, and Activity" represents the second derivative of Activity;

[0021] Step S42: extracting four rhythmic waves of the EEG signal sample: 0-3 Hz delta wave, 4-7 Hz theta wave, 8-12 Hz alpha wave, 13-30 Hz beta wave, and using the Welch-PSD algorithm to extract the power spectrum density of the rhythmic waves to determine the frequency domain characteristics, the window function of the Welch-PSD algorithm is set to a Hamming window, and the overlap rate is set to 20%;

[0022] Step S43: using the FisherRatio algorithm to calculate the Fisher ratio of the time domain features and the frequency domain features, and selecting features that are effective for classification for fusion.

[0023] Preferably, step S43 is specifically:

[0024] Step S431: Calculate the Fisher ratios φ(i) of all dimensions of the EEG signal samples, and sort all the Fisher ratios in descending order to obtain an ordered Fisher ratio set φ;

[0025] Step S432: selecting a subset of the Fisher ratio set φ and features corresponding to the subset according to a certain feature selection ratio to obtain a feature subset, and inputting the feature subset into a classifier to observe the effect;

[0026] Step S433: continuously adjusting the feature selection ratio to obtain an optimal feature selection ratio effective for classification, and performing feature fusion on the corresponding feature subsets.

[0027] Preferably, step S5 specifically includes:

[0028] Step S51: using the feature fusion data as source domain data S, and using the subject data that needs to be fatigue tested as target domain data T;

[0029] Step S52: input the source domain data S and the single target domain data T into the adversarial domain adaptation network FLDANN to obtain the classification result;

[0030] Step S53: Calculate the fatigue detection result according to the classification result.

[0031] Preferably, the pre-discriminator of the FLDANN network adopts the loss function F l (p t )=-α t (1-p t ) γ log(pt ), where p t is the predicted value of the model, α t ∈[0,1] is the weight factor, (1-p t ) γ is the modulation factor.

[0032] The present invention also proposes a cross-subject fatigue state detection system based on multi-scale EEG features, the system comprising a data acquisition module, a data filtering module, a data processing module, a feature extraction module, a feature fusion module and a detection module;

[0033] The data acquisition module is used to collect EEG data under different fatigue states and input the data into the data filtering module;

[0034] The data filtering module is used to filter the EEG data, remove the power frequency signal and high-frequency electromyographic signal in the environmental noise, and retain the 0-30 Hz EEG main frequency signal to input into the data processing module;

[0035] The data processing module is used to process the EEG main frequency signal in sections to obtain EEG signal samples in different states and input them into the feature extraction module;

[0036] The feature extraction module is used to extract the time domain and frequency domain features of the EEG signal sample and input them into the feature fusion module;

[0037] The feature fusion module is used to select and fuse the time domain and frequency domain features to obtain fused feature values, and input the fused feature values ​​into the detection module;

[0038] The detection module is used to calculate the fused feature value and the data to be detected to obtain a fatigue assessment result.

[0039] Preferably, the feature extraction module uses Hjorth to calculate the time domain features of the EEG signal samples, and uses Welch-PSD algorithm to extract the frequency domain features of the EEG signal samples.

[0040] Preferably, the feature fusion module uses Fisher Ratio algorithm to perform feature selection and fusion.

[0041] Preferably, the detection module adopts the adversarial domain adaptation network FLDANN.

[0042] The present invention studies a multi-scale fusion EEG feature extraction method, establishes a cross-subject fatigue detection model based on an improved FLDANN network, and uses the fatigue EEG data of existing subjects to identify the fatigue state of the target subject. The present invention explores a fatigue detection method based on EEG signals, which not only reduces the cost of obtaining fatigue data of the target subject for training the model, but also proposes solutions to the problems of one-sidedness and information missing of the extracted fatigue EEG features, as well as the domain imbalance caused by multi-source domains and single target domains faced in cross-subject detection, effectively improving the classification performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Flow chart of the method of the present invention.

[0044] Figure 2 Flow chart of the 2-back experiment of the present invention.

[0045] Figure 3 This is a hardware schematic diagram of the signal acquisition module in the present invention.

[0046] Figure 4 Flow chart of the FisherRatio algorithm of the present invention.

[0047] Figure 5 It is a schematic diagram of the characteristic matrix of the present invention.

[0048] Figure 6 This is a schematic diagram of the FLDANN network proposed in the present invention.

[0049] Figure 7 This is the experimental result data of the present invention. DETAILED DESCRIPTION

[0050] The present invention proposes a method and system for detecting fatigue status across subjects based on multi-scale EEG features. The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0051] like Figure 1 As shown, the present invention proposes a method for detecting fatigue state across subjects based on multi-scale EEG features, comprising the following steps:

[0052] Step S1: Collecting EEG data of the subject in fatigue state and normal state through EEG acquisition equipment;

[0053] This embodiment uses a 2-back experiment to induce eight subjects to produce EEG signals under different fatigue states; the 2-back task is a brain cognitive training task for training brain memory. The subjects compare the current stimulus signal with the previous two stimulus signals and make corresponding responses. In the experimental task of the present invention, the duration of each stimulus signal is 0.5 seconds, and the time interval between the two stimulus signals is 3 seconds. If the current stimulus signal is consistent with the previous two stimulus signals, the subject presses the ← key on the keyboard, otherwise presses the → key; Figure 2 As shown, the subjects repeatedly practiced and trained for five minutes until the test accuracy reached 80%, and then entered formal training to induce EEG signals under different fatigue states;

[0054] The formal training consisted of two rounds of 2-back training for 25 minutes each. The Chalder Fatigue Scale was filled in five minutes before and after each round of 2-back task. The fatigue status was calculated based on the results of the fatigue scale. In different states, the fatigue status was calculated by Figure 3 The EEG acquisition device shown collects EEG data corresponding to the subject, and the sampling frequency is 500HZ.

[0055] Step S2: Filter the EEG signal to extract the sub-band EEG signal between 0-30 Hz;

[0056] The effective bandwidth of the EEG signal is between 0-30 Hz. The EEG data obtained in step S1 is filtered by a 0-30 Hz filter to remove the power frequency signal and high-frequency electromyographic signal in the environmental noise and obtain a useful sub-band EEG signal.

[0057] Step S3: segmenting the sub-band EEG signal data to obtain EEG signal samples under different fatigue states;

[0058] The sub-band EEG signal data in the normal state and fatigue state were respectively cut into 120 segments of 1 s, and the segments did not overlap, resulting in 8x2x120=1920 EEG signal samples, 8 represents eight subjects, 2 represents two different stages, namely the normal state and the fatigue state, and each EEG signal sample includes 16 channels and 500 sampling points.

[0059] Step S4: extracting multi-scale time domain and frequency domain features of the EEG signal samples, and fusing them through a feature sorting selection method to calculate feature fusion data;

[0060] The set of EEG signal samples is recorded as Y = {y(1), y(2), ..., y(t)}, and the time domain characteristics of each EEG signal sample are calculated by the Hjorth parameter. The amplitude changes of EEG signal samples are different under normal and fatigue conditions, and the Hjorth parameter can be used to extract the activity, mobility and complexity of EEG signals under different states;

[0061]

[0062]

[0063]

[0064] Where T represents the number of EEG signal samples, which is 1920 in this embodiment. represents the average number of EEG signal samples, Activity' represents the first derivative of Activity, and Activity" represents the second derivative of Activity;

[0065] When the brain changes from a normal state to a fatigue state, the power changes of each rhythmic wave of the EEG signal are different. The power spectrum extracted by the Welch-PSD algorithm has a wide main lobe width, good convergence, and a small power energy variance, and the result is more accurate. First, the Welch-PSD algorithm is used to extract the four rhythmic waves of the EEG signal sample: 0-3Hz delta wave, 4-7hz theta wave, 8-12hz alpha wave, 13-30hz beta wave, and then the power spectrum density of each rhythmic wave is extracted by the Welch-PSD algorithm;

[0066] The specific method of the Welch-PSD algorithm is as follows:

[0067] Let a signal of length N be represented as {x(0), x(1)...x(N-1)}, and split it into k segments, each of which has a length of L, where L corresponds to the number of sampling points. If there is no overlap between segments, then N = k x L;

[0068] If the overlap length between segments is LD, then for the i-th segment (i=0, 1, ..., K-1), the following relationship holds:

[0069] x i (n) = x(n+iD)

[0070] Where n represents the nth sampling point in the ith segment, n = 0, 1, ..., L-1, X i (n) represents the value of the nth sampling point in the i-th segment;

[0071] The periodic spectrum of each segment is then corrected by the window function w(n). The correction process uses the following formula:

[0072]

[0073] In the formula, f represents frequency and j represents imaginary unit;

[0074] Finally, all fragments are averaged using the following formula,

[0075]

[0076] In this embodiment, the window function w(n) is set to a Hamming window, and the overlap rate is set to 20%.

[0077] Each EEG signal sample includes 16 channels, each channel includes 3 Hjorth parameters and 4 Welch-PSD features, then each EEG signal sample has 16x7=112-dimensional features. Since the feature dimension of the EEG signal itself is high and the fused features have redundant information, the present invention calculates the value of each feature and sorts them through the Fisher Ratio algorithm, selects the most effective features for classification and fuses them, realizes feature dimensionality reduction while retaining effective time-frequency domain feature information, reduces the negative impact of redundant features, and improves the classification accuracy of brain fatigue detection.

[0078] The specific process of the Fisher Ratio algorithm is as follows Figure 4 As shown, first enter Figure 5 The 240-sample x 112-dimensional feature matrix and the fatigue state corresponding to each EEG signal sample are shown. Calculate the Fisher ratio of the i-th dimension feature in the sample set Until each dimension feature is calculated, a Fisher ratio set of all dimensional features is obtained Sort all the parameters in this set in descending order to get an ordered Fisher ratio set In the feature importance evaluation stage, select according to a certain feature ratio The Fisher ratio in the original feature set and its corresponding feature in the original feature set are used to obtain a feature subset. The feature subset is input into the classifier to observe the effect. If the effect is not ideal, the selected feature ratio is adjusted until the optimal feature subset is found and the feature subset is subjected to data fusion to obtain feature fusion data. In this embodiment, through screening, the feature dimension is retained at 57% of the original number of features, that is, 64 dimensions of the 112-dimensional feature are retained.

[0079] In the present invention, the Fisher ratio of the i-th dimension feature is calculated using the following formula:

[0080]

[0081] In the formula, p represents the sample set under fatigue state, q represents the sample set under normal state, and n p Indicates the number of fatigue state samples, n q Indicates the number of normal state samples, u i Represents the mean of the i-th dimension feature in the entire data set; Represents the mean of the i-th dimension feature on the p-type data set; is the mean value of the i-th dimension feature in the q-type data set. Represents the eigenvalue of the i-th dimension feature of the k-th p-class sample; Represents the eigenvalue of the i-th dimension feature of the k-th q-class sample.

[0082] Step S5: Based on the feature fusion data, a cross-subject fatigue detection model is established to identify and detect the fatigue state of the target subject.

[0083] The feature fusion data is used as the source domain data S, and the subject data that needs to be fatigue detected is used as the target domain data T. The source domain data S and the target domain data T are input into the adversarial domain adaptation network FLDANN, and the weight coefficient and number of iterations of FLDANN are set. The specific structure and parameters of the FLDANN network are as follows: Figure 6 As shown, it includes a feature extractor, a domain discriminator and a label predictor. In the domain discriminator, a weighted loss function is used to perform weighted adjustment of domain samples; the feature extractor compresses the extracted features by a fully connected layer, first passing through a linear activation function layer, then through a batch normalization layer, and finally through an activation function layer to obtain the extracted features. The above activation function is the ELU activation function.

[0084] The domain discriminator identifies samples from different domains to determine whether the samples entering the model are from the source domain data or the target domain data. In the domain discriminator, the weights between samples from different domains in the domain discriminator are adjusted by the loss function. In this embodiment, the loss function of the domain discriminator adopts the loss function F l (p t )=-α t (1-p t ) γ log(p t ), where p t is the predicted value of the model, α t ∈[0,1] is the weight factor, (1-p t ) γ is the modulation factor. By adjusting the weight factor α t The value of controls the weight of source domain samples and target domain samples in the loss function of the domain discriminator.

[0085] There is a gradient flip layer between the domain discriminator and the feature extractor, which is used to reverse the gradient during back propagation; the domain discriminator is first composed of a fully connected layer, followed by a linear activation function and then a BN. The dropout layer is used to improve the generalization ability of the domain discriminator and prevent overfitting, and the parameter is set to 0.5; finally, it passes through a fully connected layer for classification.

[0086] The label classifier consists of a fully connected layer and a softmax activation function to classify the input samples and output the classification results. The loss function used in the label classifier is the cross entropy loss function.

[0087] Finally, statistics are performed based on the classification results to obtain fatigue assessment results. When the model predicts that the number of fatigue samples accounts for more than half of the total number of samples, it is in a fatigue state. When the number of fatigue samples predicted by the model accounts for more than 80% of the total number of samples, it is in a severe fatigue state.

[0088] like Figure 7 As shown, the method proposed in the present invention was applied to eight subjects, and the classification accuracy, precision and recall were statistically analyzed and displayed. The prediction accuracy reached 87.9%, the precision reached 92.5%, and the recall reached 84.7%.

[0089] The present invention also proposes a cross-subject fatigue state detection system based on multi-scale EEG features, including a data acquisition module, a data filtering module, a data processing module, a feature extraction module, a feature fusion module and a detection module;

[0090] The data acquisition module is used to collect EEG data under different fatigue states and input them into the data filtering module;

[0091] The data filtering module is used to filter the EEG data, remove the power frequency signal and high-frequency electromyographic signal in the environmental noise, and retain the 0-30 Hz EEG main frequency signal to input into the data processing module;

[0092] The data processing module is used to process the EEG main frequency signal in segments to obtain EEG signal samples in different states and input them into the feature extraction module;

[0093] The feature extraction module is used to extract the time domain and frequency domain features of the EEG signal samples and input them into the feature fusion module;

[0094] The feature fusion module is used to select and fuse the time domain and frequency domain features to obtain fused feature values, and input the fused feature values ​​into the detection module;

[0095] The detection module is used to calculate the fused feature value and the data to be detected to obtain a fatigue assessment result.

[0096] The feature extraction module uses Hjorth to calculate the time domain features of the EEG signal samples and uses the Welch-PSD algorithm to extract the frequency domain features of the EEG signal samples.

[0097] The feature fusion module uses the Fisher Ratio algorithm to select and fuse features.

[0098] The detection module adopts the adversarial domain adaptation network FLDANN.

[0099] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for detecting fatigue state across subjects based on multi-scale EEG features, characterized in that: The following steps are involved: Step S1: Collecting EEG data of the subject in fatigue state and non-fatigue state through EEG acquisition equipment; Step S2: filtering the EEG signal to extract the sub-band EEG signal between 0-30 Hz; Step S3: segmenting the sub-band EEG signal data to obtain EEG signal samples under different fatigue states; Step S4: extracting multi-scale time domain and frequency domain features of the EEG signal sample, and fusing them through a feature sorting selection method to calculate feature fusion data; Step S5: Based on the feature fusion data, a cross-subject fatigue detection model is established to identify and detect the fatigue state of the target subject; Step S4 specifically includes: Step S41: Calculate the time domain features of the EEG signal samples Y={y(1), y(2), ..., y(t)} by using Hjorth parameters, and extract the activity, mobility and complexity of the EEG signal samples under different states; Where T represents the number of EEG signal samples, represents the average number of EEG signal samples, Activity' represents the first derivative of Activity, and Activity" represents the second derivative of Activity; Step S42: extracting four rhythmic waves of the EEG signal sample: 0-3 Hz delta wave, 4-7 Hz theta wave, 8-12 Hz alpha wave, 13-30 Hz beta wave, and using the Welch-PSD algorithm to extract the power spectrum density of the rhythmic waves to determine the frequency domain characteristics, the window function of the Welch-PSD algorithm is set to a Hamming window, and the overlap rate is set to 20%; Step S43: using the FisherRatio algorithm to calculate the Fisher ratio of the time domain features and the frequency domain features, and selecting the features that are effective for classification for fusion; Step S43 is specifically as follows: Step S431: Calculate the Fisher ratios φ(i) of all dimensions of the EEG signal samples, and sort all the Fisher ratios in descending order to obtain an ordered Fisher ratio set φ; Step S432: selecting a subset of the Fisher ratio set φ and features corresponding to the subset according to a certain feature selection ratio to obtain a feature subset, and inputting the feature subset into a classifier to observe the effect; Step S433: continuously adjusting the feature selection ratio to obtain an optimal feature selection ratio effective for classification, and performing feature fusion on the corresponding feature subsets.

2. The method for detecting fatigue state across subjects based on multi-scale EEG features according to claim 1, characterized in that: The different fatigue states in step S1 are induced by a 2-back experiment, and the steps of the 2-back experiment are specifically as follows: Step S11: The subject takes a pre-practice screening test, and when the accuracy rate reaches 80%, he / she is determined as a formal subject, and then formal training is carried out; Step S12: the formal subject performs at least two rounds of 2-back training, the fatigue state of the formal subject is determined according to the results of the 2-back training, and EEG data under different fatigue states are collected.

3. The method for detecting fatigue state across subjects based on multi-scale EEG features according to claim 1, characterized in that: In step S3, the sub-band EEG signal collected for a long time is divided into an EEG signal sample at every 500 sampling points according to the sampling frequency of 500 Hz, and there is no overlap between the samples.

4. The method for detecting fatigue state across subjects based on multi-scale EEG features according to claim 1, characterized in that: Step S5 specifically includes: Step S51: using the feature fusion data as source domain data S, and using the subject data that needs to be fatigue tested as target domain data T; Step S52: input the source domain data S and the single target domain data T into the adversarial domain adaptation network FLDANN to obtain the classification result; Step S53: Calculate the fatigue detection result according to the classification result.

5. The method for detecting fatigue state across subjects based on multi-scale EEG features according to claim 4 is characterized in that: The pre-discriminator of the FLDANN network adopts the loss function F l (p t )=-α t (1-p t ) γ log(p t ), where p t is the predicted value of the model, α t ∈[0,1] is the weight factor, (1-p t ) γ is the modulation factor.

6. A cross-subject fatigue state detection system based on multi-scale EEG features, characterized in that: The system includes a data acquisition module, a data filtering module, a data processing module, a feature extraction module, a feature fusion module and a detection module; The data acquisition module is used to collect EEG data under different fatigue states and input the data into the data filtering module; The data filtering module is used to filter the EEG data, remove the power frequency signal and high-frequency electromyographic signal in the environmental noise, and retain the 0-30 Hz EEG main frequency signal to input into the data processing module; The data processing module is used to process the EEG main frequency signal in sections to obtain EEG signal samples in different states and input them into the feature extraction module; The feature extraction module is used to extract the time domain and frequency domain features of the EEG signal sample and input them into the feature fusion module; The feature fusion module is used to select and fuse the time domain and frequency domain features to obtain fused feature values, and input the fused feature values ​​into the detection module; The detection module is used to calculate the fused feature value and the data to be detected to obtain a fatigue assessment result; The feature extraction module calculates the time domain features of the EEG signal samples Y={y(1), y(2), ..., y(t)} through Hjorth parameters, and extracts the activity, mobility and complexity of the EEG signal samples under different states; Where T represents the number of EEG signal samples, represents the average number of EEG signal samples, Activity' represents the first derivative of Activity, and Activity" represents the second derivative of Activity; Extracting four rhythmic waves of the EEG signal sample: 0-3 Hz delta wave, 4-7 Hz theta wave, 8-12 Hz alpha wave, 13-30 Hz beta wave, and using the Welch-PSD algorithm to extract the power spectrum density of the rhythmic waves to determine the frequency domain characteristics, the window function of the Welch-PSD algorithm is set to a Hamming window, and the overlap rate is set to 20%; The feature fusion module uses the FisherRatio algorithm to calculate the Fisher ratio of the time domain features and the frequency domain features, and selects the features that are effective for classification for fusion: Calculating the Fisher ratios φ(i) of all dimensions of the EEG signal samples, and sorting all the Fisher ratios in descending order to obtain an ordered Fisher ratio set φ; Select a subset of the Fisher ratio set φ and features corresponding to the subset according to a certain feature selection ratio to obtain a feature subset, and input the feature subset into a classifier to observe the effect; The feature selection ratio is continuously adjusted to obtain an optimal feature selection ratio that is effective for classification, and the corresponding feature subsets are subjected to feature fusion.

7. The cross-subject fatigue state detection system based on multi-scale EEG features according to claim 6 is characterized in that: The detection module adopts the adversarial domain adaptation network FLDANN.

Images