A mental fatigue detection dataset construction method based on a pvt experiment paradigm

Abstract

The application discloses a mental fatigue detection dataset construction method based on a PVT experiment paradigm, uses the PVT experiment paradigm to induce mental fatigue of a subject, synchronously collects behavior performance and corresponding electroencephalogram signal data of the subject, carries out noise reduction pretreatment on the collected electroencephalogram signal, cuts out fixed length segments corresponding to each test from the electroencephalogram signal, formulates a plurality of derived task performance indexes based on a reaction time task performance index of the PVT, uses the task performance indexes and combinations thereof to represent a mental fatigue state, sets a classification threshold, a classification benchmark and a cleaning method based on the task performance index and the mental fatigue state represented thereby, classifies and labels and cleanses the electroencephalogram signal segments, extracts features of the labeled and retained electroencephalogram signal segments, compiles an input feature matrix and a label vector, and constructs a dataset used for training and testing a machine learning model. The dataset construction method can improve mental fatigue detection performance of the machine learning model.

Description

Technical Field

[0001] This invention belongs to the fields of cognitive neuroscience and computer science, and specifically relates to a method for constructing a mental fatigue detection dataset based on the PVT experimental paradigm. Background Technology

[0002] Currently, mental fatigue is generally considered to be a gradual accumulation during mental work, typically manifesting as an inability to continue performing tasks or a loss of the will to do so, accompanied by negative effects such as poor concentration, poor decision-making, reduced efficiency, or memory decline. Mental fatigue is a challenge faced by workers in many industries, and severe mental fatigue can even lead to serious safety accidents.

[0003] To scientifically measure people's mental fatigue, experimental psychologists have proposed numerous cognitive-behavioral experimental paradigms. The Psychomotor Vigilance Test (PVT) is one of the most widely used methods, largely due to its high sensitivity to mental fatigue. Furthermore, the PVT has many advantages over other mental fatigue tests, such as low time cost (a single test only takes 10 minutes) and immunity to learning effects and differences in ability among subjects (the task requirements are extremely simple). Specifically, the standard version of the PVT lasts 10 minutes. During this time, visual stimuli are presented multiple times at random stimulus intervals (Inter-Stimulus Interval, ISI) of 2–10 seconds. Subjects are required to respond by pressing buttons as quickly as possible after perceiving the stimulus, and the system records their reaction time (RT) and other behavioral performance. Unlike the simplest stimulus-response model, the PVT, in addition to relying on specific stimulus forms (usually visual) and response methods (usually button presses), is also influenced by many factors such as the test duration and ISI parameter design.

[0004] Electroencephalography (EEG) signals reflect the sum of postsynaptic inhibitory and excitatory potentials in nerve cells. Due to its ability to directly and promptly reflect neural activity, EEG signals are considered a reliable indicator for detecting mental fatigue. However, EEG signals acquired by electromechanical devices are actually highly random, exhibiting nonlinear, non-Gaussian, and non-stationary properties. Therefore, effectively analyzing brain neural activity often requires extracting suitable features from EEG signals, including time-domain features, frequency-domain features, spatial-domain features, nonlinear features, network features, and fusion features. Among these, commonly used features suitable for training machine learning models for mental fatigue detection include power (frequency-domain feature) and entropy (nonlinear feature). For example, when a subject changes between alert and fatigued states, the EEG power in different brain regions and frequency bands will show significant differences; entropy features are also very effective neural markers of mental fatigue, such as differential entropy (DE), which can demonstrate superior performance compared to frequency band power in some applications.

[0005] During prolonged or high-intensity work, the accumulation of mental fatigue accelerates significantly, necessitating timely detection and early warning through technological means to mitigate risks. Machine learning models capable of detecting mental fatigue can be trained using EEG signals and classification labels generated by subjects in fatigue and alert states. Support Vector Machine (SVM) models, suitable for solving small-sample, high-dimensional, and non-linear pattern recognition problems, are commonly used for detecting mental fatigue. To effectively filter out important features, tree models such as Random Forest (RF) can be selected. The importance ranking generated by the RF model is used to recursively eliminate features, achieving better mental fatigue detection results with a smaller number of important features. XGBoost (eXtreme Gradient Boosting) is an efficient gradient boosting algorithm that has gradually become a popular choice in the field of mental fatigue detection in recent years due to its excellent ease of use, accuracy, and scalability.

[0006] To obtain a machine learning model capable of accurately detecting mental fatigue, it is necessary to train, evaluate, and optimize the model using a suitable dataset. In the recent wave of artificial intelligence, machine learning (especially deep learning) technology has advanced rapidly, with better methods constantly being proposed, validated, and promoted. However, good datasets are not easy to obtain because constructing a mental fatigue detection dataset often faces many difficulties. First, the methods for inducing mental fatigue in existing datasets are relatively homogeneous, mostly employing simulated driving paradigms, which to some extent limits the model's generalization ability outside of driving scenarios. Second, the true degree of mental fatigue corresponding to each EEG sample is difficult to define and label: taking commonly used labeling benchmarks such as surface bioactivity and task performance indicators as examples, the former cannot reflect changes in external behavioral performance caused by mental fatigue, while the latter cannot be correlated with the internal physiological changes caused by mental fatigue. Furthermore, selecting from numerous features of EEG signals is also highly uncertain; features that perform well in certain experimental paradigms or natural scenarios may not be applicable to other experimental paradigms or natural scenarios, while features with strong universality often fail to provide the model with the best performance. Summary of the Invention

[0007] This invention provides a method for constructing a mental fatigue detection dataset based on the PVT experimental paradigm to solve the technical problems existing in the prior art.

[0008] The technical solution adopted by this invention to solve the technical problems existing in the prior art is as follows:

[0009] A method for constructing a mental fatigue detection dataset based on the PVT experimental paradigm includes the following steps:

[0010] Step 1: Use the PVT experimental paradigm to induce mental fatigue in the subjects and simultaneously collect their behavioral performance and corresponding EEG signal data.

[0011] Step 2: Perform noise reduction preprocessing on the collected EEG signal data;

[0012] Step 3: Segment the EEG signal into fixed-length segments corresponding to each trial.

[0013] Step 4: Based on the reaction time task performance index of PVT, develop a variety of derived task performance indicators, and use these task performance indicators and their combinations to characterize mental fatigue.

[0014] Step 5: Based on task performance indicators and the mental fatigue state they represent, set classification thresholds, classification benchmarks, and cleaning methods to classify, label, and clean EEG signal segments.

[0015] Step 6: Extract features from the cleaned and retained EEG signal segments with classification labels, compile the input feature matrix and label vector, and thus construct a dataset for training and testing machine learning models.

[0016] Step 7: Validate the constructed input feature matrix and label vector dataset using one or more of the following combinations: validate the validity of the data annotation, validate the differences in feature extraction, and validate the applicability of multiple models.

[0017] Further, step 1 includes the following methodological steps: During the PVT task, the subject always gazes at the fixation point pattern displayed on the monitor and presses a button immediately after the fixation point pattern changes to the stimulus pattern; after receiving the button information, the time difference between stimulus presentation and button response is calculated and recorded as the reaction time of the current trial; the end of the response is defined as the stimulus pattern switching back to the fixation point pattern, and the interval from the end of the response to the next stimulus presentation is randomly selected by the experimental procedure within the range of 2 to 10 seconds; in a complete experiment, each subject needs to perform 3 to 5 PVT tasks, and rest quietly for 5 to 15 minutes after each task; the duration of the first PVT task is set to 20 to 40 minutes, and the others are 10 to 20 minutes; the experimental procedure generates digital markers at the stimulus presentation time and button response time of each trial and sends them to the EEG signal acquisition device in a timely manner; the temporal resolution of EEG signal acquisition and behavioral marker generation is 1000 Hz.

[0018] Furthermore, in step 2, the noise reduction preprocessing method includes one or more of the following methods in combination: deleting useless segments, deleting useless channels, configuring channel localization, low-pass filtering, high-pass filtering, removing power frequency interference, channel rereference, reducing sampling rate, eliminating bad channels, artifact subspace reconstruction, missing channel reconstruction, independent component decomposition, independent component selection, and bandpass filtering.

[0019] Furthermore, in step 3, a fixed-duration segment of EEG signal is extracted before the stimulus presentation time in each trial.

[0020] Further, in step 4, the actual reaction time of each trial is defined as its local reaction time, and the weighted sum of the local reaction times of each trial and its sequentially adjacent trials is taken as its global reaction time, with the weights of each trial summed to 1. Local reaction time and / or global reaction time are used to characterize the subject's level of mental fatigue and serve as task performance indicators reflecting mental fatigue status. Let i be the trial number, l1 represent the number of adjacent trials before the i-th trial included in the weighted average, l2 represent the number of adjacent trials after the i-th trial included in the weighted average, and the sum of the number of sequentially adjacent trials is l1 + l2 + 1. The formula for calculating the global reaction time is as follows:

[0021]

[0022] In the formula:

[0023] w k The weight for the k-th trial;

[0024] LRT k Let be the local reaction time of the k-th trial;

[0025] GRT i Let be the global reaction time for the i-th trial.

[0026] Further, in step 5, potential reaction time, alert reaction time, and fatigue reaction time are set as classification thresholds for mental fatigue. Assuming each subject experienced N trials in the experiment, resulting in N local reaction times, all local reaction times are sorted from smallest to largest. The local reaction time with a ranking of 5% × N is defined as the potential reaction time. The alert reaction time is set to 1.20–1.30 times the potential reaction time; the fatigue reaction time is set to 1.50–2.00 times the potential reaction time. When constructing the same dataset, the ratio of alert reaction time to fatigue reaction time is fixed. For each sample, if both its local and global reaction times are less than its alert reaction time, it is labeled as an alert sample; if both its local and global reaction times are greater than its fatigue reaction time, it is labeled as a fatigue sample. Samples that are neither alert nor fatigue samples are removed.

[0027] Furthermore, if the ratio of fatigue samples to alert samples in the labeled sample set from the same subject is greater than 5 or less than 0.2, it is determined that the subject's samples have a class imbalance problem, and all samples of that subject are removed and not included in the construction of the dataset.

[0028] Furthermore, in step 6, the method for extracting features of EEG signal segments with classification labels includes one or more of the following combinations: setting the length of the EEG signal segment corresponding to the sample to 1000-2000 milliseconds; selecting one or more of the following combinations of features to extract the EEG signal segment: time domain features, frequency domain features, spatial domain features, nonlinear features, network features, or fusion features.

[0029] Furthermore, in step 7, XGBoost is used as the benchmark machine learning model. The differential entropy of each frequency band and each channel extracted from the EEG signal segments is used as the input features of the benchmark model. The effectiveness of data annotation is verified by comparing the performance of models with different sample annotation methods in detecting mental fatigue. The benchmark features are used to test the performance of three machine learning models, namely XGBoost, RF and SVM, with the goal of evaluating the performance of mental fatigue detection. Based on the classification ability of the three models across the mental fatigue state of the subjects, the generalizability of the constructed dataset among different models is verified.

[0030] Further, in step 7, the method for verifying the difference in feature extraction includes the following methods and steps: taking out labeled samples of two categories, calculating the EEG signal features of each channel and frequency band for each sample, and drawing a brain topography map of the mean difference and a brain topography map of the significance test; in the brain topography map of the mean difference, the value of each point is equal to the average value of the EEG signal features of one type of labeled sample in that channel and frequency band minus the corresponding average value of the other type of labeled sample; each point on the brain topography map of the significance test represents the significance test result of the EEG signal feature value of the two types of labeled samples in that channel and frequency band, i.e., the p value; using XGBoost as the benchmark machine learning model, and conducting a comparative experiment on the model's mental fatigue detection performance by using different EEG signal features as the benchmark model input features to verify the difference in feature extraction.

[0031] The advantages and positive effects of this invention are:

[0032] 1. Based on the cognitive behavioral principles of PVT and the need to construct a dataset for detecting mental fatigue, this invention improves some design details and implementation procedures in the standard PVT experimental paradigm, thereby enhancing its effectiveness in inducing mental fatigue in subjects to a certain extent.

[0033] 2. This invention uses the PVT experimental paradigm to induce mental fatigue in subjects and constructs a dataset using synchronously collected behavioral performance and corresponding EEG signals, which can expand the versatility of fatigue detection machine learning methods in more scenarios beyond driving.

[0034] 3. Based on reaction time, which can be directly measured in PVT task performance indicators, this invention proposes a series of derivative task performance indicators that are more correlated with mental fatigue state, namely local reaction time, global reaction time, potential reaction time, alert reaction time, and fatigue reaction time.

[0035] 4. By comprehensively utilizing multiple novel task performance indicators, this invention proposes an effective data labeling and cleaning method. This method can naturally and simultaneously clean transitional samples and label fatigued and alert samples. Notably, this method can also construct a dataset with a relatively balanced distribution of categories and subjects, avoiding bias in machine learning models towards specific categories or subjects during training and testing. Furthermore, the effectiveness of this innovative data labeling and cleaning method has been verified through comparative experiments, and the results confirm its significant advantages over methods using only a single task performance indicator.

[0036] 5. For EEG signal data, this invention proposes four benchmark feature extraction methods, namely, combinations of two segment lengths and two feature types, thereby further transforming the labeled EEG segments into an input feature matrix form more suitable for training and testing machine learning models. Related validation experiments demonstrate differences in the representation of mental fatigue states among these feature extraction methods. Using the preferred feature extraction method when constructing the dataset can improve the mental fatigue detection performance of machine learning models.

[0037] 6. Validation experiments on different machine learning methods demonstrate that the mental fatigue detection dataset constructed in this invention is applicable to three common machine learning models: XGBoost, RF, and SVM. In other words, this dataset has good versatility for multiple machine learning models. Attached Figure Description

[0038] Figure 1 This is a flowchart of the construction process for the mental fatigue detection dataset.

[0039] Figure 2 This is a schematic diagram of the PVT experimental paradigm.

[0040] Figure 3 This is a diagram illustrating task performance indicators.

[0041] Figure 4 This is a statistical chart of the number of labeled samples from all subjects.

[0042] Figure 5 This is a schematic diagram of the task performance indicators of a subject who failed due to mental fatigue.

[0043] Figure 6 It is a brain topography map showing the average difference in the logarithmic power characteristics across the entire frequency band between fatigue and alert samples.

[0044] Figure 7 It is a brain topography map that tests the significance of the full-band logarithmic power characteristics of fatigue and alert samples.

[0045] Figure 8It is a brain topography map showing the average difference in the differential entropy features across the entire frequency band between fatigue and alert samples.

[0046] Figure 9 It is a brain topography map that tests the significance of the full-band differential entropy features of fatigue and alert samples.

[0047] In the picture:

[0048] Figure 4 An asterisk (*) marks the subject IDs, indicating that the sample of the corresponding subject has a class imbalance problem.

[0049] Figures 6 to 9 In this context, delta represents the δ band; theta represents the θ band; alpha represents the α band; beta represents the β band; and gamma represents the γ band.

[0050] Figure 6 and Figure 8 The value below the color band in the image represents the average difference value corresponding to the color gradient of the color band; Figure 7 and Figure 9 The values ​​below the color bands in the diagram represent the p-values ​​for the significance test corresponding to the color gradient of the color bands. Detailed Implementation

[0051] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0052] The following are the Chinese definitions of English words, abbreviations, and phrases:

[0053] PVT: refers to the psychomotor alertness test, a cognitive-behavioral test that can assess mental fatigue.

[0054] ISI: refers to the random stimulus interval, which is a variable experimental parameter in the PVT experimental paradigm;

[0055] RT: refers to reaction time, which is a task performance indicator in the PVT experimental paradigm;

[0056] EEG: refers to electroencephalogram (EEG) signals, which are electrophysiological signals generated by the human brain.

[0057] SVM stands for Support Vector Machine, which is a machine learning method.

[0058] RF stands for Random Forest, a machine learning method.

[0059] XGBoost: Extreme Gradient Boosting, is a machine learning method.

[0060] Please see Figures 1 to 9A method for constructing a mental fatigue detection dataset based on the PVT experimental paradigm includes the following steps:

[0061] Step 1: Use the PVT experimental paradigm to induce mental fatigue in the subjects and simultaneously collect their behavioral performance and corresponding EEG signal data.

[0062] Step 2: Perform noise reduction preprocessing on the collected EEG signal data.

[0063] Step 3: Segment the EEG signal into fixed-length segments corresponding to each trial.

[0064] Step 4: Based on the reaction time task performance index of PVT, develop a variety of derived task performance indicators, and use these task performance indicators and their combinations to characterize mental fatigue.

[0065] Step 5: Based on task performance indicators and the mental fatigue state they represent, set classification thresholds, classification benchmarks, and cleaning methods to classify, label, and clean EEG signal segments.

[0066] Step 6: Extract features from the cleaned and retained EEG signal segments with classification labels, compile the input feature matrix and label vector, and thus construct a dataset for training and testing machine learning models.

[0067] Step 7: Validate the constructed input feature matrix and label vector dataset using one or more of the following combinations: validate the validity of the data annotation, validate the differences in feature extraction, and validate the applicability of multiple models.

[0068] Preferably, step 1 may include the following steps: During the PVT task, the subject always gazes at the fixation point pattern displayed on the monitor and presses a button immediately after the fixation point pattern changes to the stimulus pattern; after receiving the button information, the time difference between stimulus presentation and button response is calculated and recorded as the reaction time of the current trial; the end of the response is defined as the stimulus pattern switching back to the fixation point pattern, and the interval from the end of the response to the next stimulus presentation is randomly selected by the experimental program within the range of 2 to 10 seconds; in a complete experiment, each subject needs to perform 3 to 5 PVT tasks, and rest quietly for 5 to 15 minutes after each task; the duration of the first PVT task is set to 20 to 40 minutes, and the others are 10 to 20 minutes; the experimental program generates digital markers at the stimulus presentation time and button response time of each trial and sends them to the EEG signal acquisition device in a timely manner; the temporal resolution of EEG signal acquisition and behavioral marker generation is 1000 Hz.

[0069] Preferably, in step 2, the noise reduction preprocessing method may include one or more of the following methods in combination: deleting useless segments, deleting useless channels, configuring channel localization, low-pass filtering, high-pass filtering, removing power frequency interference, channel rereference, reducing sampling rate, eliminating bad channels, artifact subspace reconstruction, missing channel reconstruction, independent component decomposition, independent component selection, and bandpass filtering.

[0070] Preferably, in step 3, a fixed-duration segment of EEG signal can be extracted before the stimulus presentation time in each trial.

[0071] Preferably, in step 4, the actual reaction time of each trial can be defined as its local reaction time, and the weighted sum of the local reaction times of each trial and its sequentially adjacent trials can be used as its global reaction time, with the weights of each trial being 1. Local reaction time and / or global reaction time can be used to characterize the subject's level of mental fatigue and serve as a task performance indicator reflecting mental fatigue. Let i be the trial number, l1 represent the number of adjacent trials before the i-th trial included in the weighted average, and l2 represent the number of adjacent trials after the i-th trial included in the weighted average, with the sum of the number of sequentially adjacent trials being l1 + l2 + 1. The formula for calculating the global reaction time is as follows:

[0072]

[0073] In the formula:

[0074] w k Let k be the weight of the k-th trial; k is the trial number.

[0075] LRT k Let be the local reaction time of the k-th trial;

[0076] GRT i Let be the global reaction time for the i-th trial.

[0077] Preferably, in step 5, potential reaction time, alert reaction time, and fatigue reaction time can be set as classification thresholds for mental fatigue state; assuming each subject experienced N trials in the experiment, i.e., generated N local reaction times, all local reaction times are sorted from smallest to largest, and the local reaction time with a ranking of 5% × N can be defined as potential reaction time; alert reaction time can be set to 1.20 to 1.30 times the potential reaction time; fatigue reaction time can be set to 1.50 to 2.00 times the potential reaction time; when constructing the same dataset, the ratio of alert reaction time to fatigue reaction time is fixed; for each sample, if its local reaction time and global reaction time are both less than its alert reaction time, it can be labeled as an alert sample; if its local reaction time and global reaction time are both greater than its fatigue reaction time, it can be labeled as a fatigue sample; samples that are neither alert nor fatigue samples can be removed.

[0078] Preferably, if the ratio of fatigue samples to alert samples in the labeled sample set from the same subject is greater than 5 or less than 0.2, it can be determined that the subject's samples have a class imbalance problem, and all samples of the subject can be removed and not included in the construction of the dataset.

[0079] Preferably, in step 6, the method for extracting features of EEG signal segments with classification labels may include one or more of the following combinations: the length of the EEG signal segment corresponding to the sample may be set to 1000-2000 milliseconds; one or more of the following combinations of features may be extracted from the EEG signal segment: time domain features, frequency domain features, spatial domain features, nonlinear features, network features, or fusion features.

[0080] Preferably, in step 7, XGBoost can be used as the benchmark machine learning model, and the differential entropy of each frequency band and each channel extracted from the EEG signal segment can be used as the input features of the benchmark model. The effectiveness of data annotation can be verified by comparing the performance of models with different sample annotation methods in detecting mental fatigue. The benchmark features can be used to test the performance of XGBoost, RF and SVM machine learning models with the goal of evaluating mental fatigue detection. Based on the classification ability of the three models across the mental fatigue state of the subjects, the generalizability of the constructed dataset among different models can be verified.

[0081] Preferably, in step 7, the method for verifying the difference in feature extraction may include the following methods and steps: taking out labeled samples of two categories, calculating the EEG signal features of each channel and frequency band for each sample, and drawing a brain topography map of the mean difference and a brain topography map of the significance test; in the brain topography map of the mean difference, the value of each point is equal to the mean value of the EEG signal features of one type of labeled sample in that channel and frequency band minus the corresponding mean value of the other type of labeled sample; each point on the brain topography map of the significance test represents the significance test result of the EEG signal feature value of the two types of labeled samples in that channel and frequency band, i.e., the p value; XGBoost can be used as a benchmark machine learning model, and the difference in feature extraction can be verified by comparing the performance of the model in detecting mental fatigue by using different EEG signal features as the benchmark model input features.

[0082] The p-value is a parameter used to determine the result of a hypothesis test. The p-value indicates the probability that a more extreme outcome than the observed sample result will occur when the null hypothesis is true. The smaller the p-value, the more significant the result.

[0083] The workflow and working principle of the present invention will be further described below with reference to a preferred embodiment:

[0084] The specific process for constructing a mental fatigue detection dataset based on the PVT experimental paradigm proposed in this invention is as follows: Figure 1 As shown. With Figure 1 The specific invention steps related to the content are as follows:

[0085] A method for constructing a mental fatigue detection dataset based on the PVT experimental paradigm includes the following steps:

[0086] Step 1: Based on the need to construct a mental fatigue detection dataset, improve some design details and implementation procedures in the standard PVT experimental paradigm; recruit subjects to participate in the experiment, use the PVT experimental paradigm to induce mental fatigue in them, and simultaneously collect their behavioral performance and corresponding EEG signal data.

[0087] Step 2: Denoise reduction and bandpass filtering are performed on the raw EEG signal data collected from the experiment to obtain EEG signal data of different frequency bands with most noise removed.

[0088] Step 3: Extract fixed-length segments from the complete EEG signals of each frequency band, corresponding to each test.

[0089] Step 4: Based on the reaction time task performance index of PVT, develop various derived task performance indices and their combinations to characterize mental fatigue. Using reaction time, which can be directly measured within the PVT task performance index, invent derived task performance indices that better characterize the mental fatigue state of subjects, namely, local reaction time, global reaction time, potential reaction time, alert reaction time, and fatigue reaction time. These derived task performance indices are sets of computational indicators generated based on the existing reaction time measurement index through certain computational rules, such as logical operations.

[0090] Step 5: Based on the task performance indicators and the mental fatigue state they represent, set classification thresholds, classification benchmarks, and cleaning methods to classify, label, and clean EEG signal segments, retaining the EEG signal segments that are strongly correlated with the mental fatigue state as samples for constructing the dataset.

[0091] Step 6: Extract features from the cleaned and retained EEG signal segments with classification labels, compile the input feature matrix and label vector, and thus construct a dataset for training and testing machine learning models.

[0092] Step 7 involves validating the constructed input feature matrix and label vector dataset using one or more of the following methods: validating the effectiveness of data annotation, validating the differences in feature extraction, and validating the applicability of various models. Specifically, the samples with completed feature extraction and category labeling are divided into training, validation, and test sets to conduct validation experiments on the performance of the machine learning model for detecting mental fatigue.

[0093] The working process and working principle of the present invention will be explained below in conjunction with the working method of the present invention:

[0094] Data acquisition based on the PVT experimental paradigm:

[0095] In order to stably and efficiently induce mental fatigue in subjects within a limited time and to expand the versatility of the mental fatigue detection machine learning model in more scenarios, this invention adopts a PVT-based experimental paradigm and recruits 20 subjects to participate in the experiment.

[0096] This invention applies an improved visual PVT experimental paradigm, the specific details and implementation process of which are as follows: Figure 2 As shown. During the PVT task, the subject needs to keep their gaze fixed on the white cross-shaped fixation point displayed on the monitor, and press a designated key on the keyboard as soon as the fixation point changes to a red circular stimulus. After receiving the key input, the computer will re-fixate the stimulus to the fixation point.

[0097] The time difference between each stimulus presentation and the corresponding button response was calculated and recorded as the reaction time for each trial. The end of the response was defined as the stimulus pattern switching back to the fixation point pattern. The interval between the end of the response and the next stimulus presentation was randomly selected by the experimental procedure within the range of 2–10 seconds. In a complete experiment, each subject performed the PVT task three times, resting quietly for 10 minutes after each task. The first PVT task lasted 40 minutes, and the subsequent two PVT tasks each lasted 20 minutes.

[0098] This invention eliminates the reaction time feedback step included in the standard version of PVT in the experimental paradigm's program design, and uses static graphics as visual stimuli instead, which can further enhance the mental fatigue-inducing effect of the PVT task.

[0099] According to the preliminary experimental results of this invention, longer PVT durations can better induce mental fatigue in subjects; however, subjects are also more likely to exhibit significant negative emotions and make large-scale disruptive movements in the later stages of prolonged PVT, making it difficult to guarantee the quality of EEG data acquisition. Therefore, this invention sets the duration of the first PVT task to 40 minutes to induce mental fatigue in subjects as quickly as possible; while the 20-minute duration of subsequent PVT tasks is a result of balancing the effect of inducing mental fatigue and the quality of EEG data acquisition.

[0100] According to the experimental paradigm designed in this invention, while subjects perform the PVT task, specialized electromechanical equipment is needed to collect their scalp electroencephalogram (EEG) signals. To synchronize and match the subjects' behavioral performance with their EEG signals in time, this invention develops a computer experimental program to generate digital markers at each stimulus presentation and key press response moment, and promptly transmits these markers to the EEG signal acquisition device via a computer hardware communication interface. Specifically, the temporal resolution for both EEG signal acquisition and behavioral marker generation is 1000Hz, meaning the system has millisecond-level accuracy.

[0101] Data labeling and cleaning based on task performance:

[0102] For stimulus-response tests like the PVT, the subject's responses in each trial are the direct and fundamental basis for measuring their behavioral performance. Commonly used task performance indicators include reaction time and error frequency. Because indicators such as reaction time or reaction speed can be measured in each trial, this invention uses them as a behavioral performance indicator to indirectly characterize the subject's level of mental fatigue. However, from the perspective of the entire experiment, reaction time often exhibits frequent and drastic fluctuations; that is, there may be significant differences in reaction time between adjacent trials. Figure 3The green line in the diagram illustrates this. Since the physiological state of mental fatigue in subjects does not actually exhibit such frequent fluctuations, this task performance indicator, reaction time, obtained by directly calculating the time difference between stimulus and response, is not suitable as a label representing the mental fatigue state corresponding to a trial or EEG segment.

[0103] To address the aforementioned problems, this invention utilizes the concepts of Local-RT and Global-RT: the actual reaction time of each trial is defined as its local reaction time; the weighted sum of the local reaction times of each trial and its sequentially adjacent trials is used as its global reaction time, with the weights of each trial totaling 1; local reaction time and / or global reaction time are used to characterize the subject's level of mental fatigue and serve as task performance indicators reflecting mental fatigue status. Let i be the trial number, l1 represent the number of adjacent trials before the i-th trial included in the weighted average, l2 represent the number of adjacent trials after the i-th trial included in the weighted average, and the sum of the number of sequentially adjacent trials is l1 + l2 + 1; then the formula for calculating the global reaction time is as follows:

[0104]

[0105] In the formula:

[0106] w k The weight for the k-th trial;

[0107] LRT k Let be the local reaction time of the k-th trial;

[0108] GRT i Let be the global reaction time for the i-th trial.

[0109] Based on the results of the preliminary experiments, the specific calculation method selected in this invention is a two-sided equal-weighted average, that is:

[0110]

[0111] In the formula:

[0112] l represents the number of adjacent trials before and after the i-th trial included in the weighted average.

[0113] The average number of adjacent trials within 30 seconds before and after each trial is taken as l for calculation, and the global reaction time obtained from this is as follows: Figure 3 As shown by the purple broken line in the figure. By combining the local reaction time, which represents the degree of short-term (about 5 seconds) mental fatigue, and the global reaction time, which represents the degree of long-term (about 1 minute) mental fatigue, a more stable sample labeling benchmark for characterizing the degree of mental fatigue can be obtained.

[0114] In PVT, a stimulus and its corresponding response constitute a trial. However, the randomness of the stimulus interval makes the time scale of each trial unclear. To construct the trials in a uniform form as samples in the dataset, a fixed-length EEG segment before the stimulus presentation time can be extracted from each trial. Taking the segment before the stimulus is to avoid the model being limited to specific event-related neural activity induced by PVT.

[0115] To classify samples corresponding to different states of mental fatigue into fatigue and alertness using machine learning models, it is necessary to determine the threshold for labeling sample categories based on local reaction time and global reaction time. Therefore, this invention proposes three concepts: Potential Reaction Time (Potential-RT), Vigilant Reaction Time (Vigilant-RT), and Fatigue Reaction Time (Fatigue-RT), as described below. Figure 3 The yellow, blue, and red lines in the diagram illustrate this. Assuming each subject experienced N trials in the experiment, resulting in N local reaction times, and all local reaction times are sorted from smallest to largest, the local reaction time ranking 5% × N is defined as the potential reaction time. Potential reaction time reflects the upper limit of a subject's reaction ability during the experiment and varies significantly among individuals. Based on potential reaction time, two classification thresholds can be defined: alert reaction time and fatigue reaction time. Alert reaction time is set at 1.25 times the potential reaction time, and fatigue reaction time is set at 1.50 times the potential reaction time. For each sample, if both its local and global reaction times are less than its alert reaction time, it is labeled as an alert sample; if both its local and global reaction times are greater than its fatigue reaction time, it is labeled as a fatigue sample.

[0116] This annotation method excludes some samples that do not meet the constraints from the dataset because some samples always fall into the transition region between the two thresholds, and some samples have contradictory local and global reaction times. These excluded samples generally correspond to relatively moderate task performance at the behavioral level, but correspond to complex and difficult-to-determine states at the neural activity level, making it impossible for the model to learn relatively deterministic alertness or fatigue neural patterns. Therefore, samples that are neither alertness samples nor fatigue samples are removed.

[0117] Based on the above annotation method, we can obtain the following: Figure 4 The chart shows the statistical statistics of the labeled sample size for all subjects. From Figure 4It can be observed that the number of labeled samples for most subjects is much smaller than their total sample size (approximately 900-1000), indicating that the data labeling method proposed in this invention can naturally clean up interfering data that is not strongly correlated with mental fatigue. Furthermore, 8 out of the 20 subjects exhibited significant sample class imbalance, with the 10th subject's actual performance ( Figure 5 Taking [example 1] as an example, it can be observed that the local reaction time and global reaction time of each trial remained relatively stable throughout the task, and most samples met the labeling requirements for the alertness category. This indicates that the PVT experiment did not sufficiently induce mental fatigue in this subject. For the eight subjects with this situation, the measures taken in this invention are to perform data cleaning: if the ratio of fatigue samples to alertness samples in the labeled sample set from the same subject is greater than 5 or less than 0.2, it is determined that the subject's samples have a class imbalance problem; all samples of this subject are removed and not included in the construction of the dataset to avoid the model losing classification performance due to excessive learning of neural patterns of either alertness or fatigue.

[0118] To further improve the performance of the classification model through data cleaning and avoid bias towards specific categories or subjects during training and testing, this invention, in a preferred embodiment, constructs a dataset where both categories and subjects are perfectly balanced. Specifically, for each subject, K alert samples and K fatigue samples are uniformly retained. When a certain category has more than K samples, a stricter alertness or fatigue classification threshold is used to exclude the excess samples. The value of K is set between 50 and 150.

[0119] The invention ultimately constructed an EEG signal dataset containing 100 positive and negative sample pairs from each of 12 subjects (a total of 2400 samples), which can be used to train and test a machine learning classifier capable of detecting mental fatigue.

[0120] Processing and feature extraction based on electroencephalogram (EEG) signals:

[0121] Scalp EEG signals are relatively weak electrophysiological signals, easily interfered with by the subject's own electrical activity and that of other electrical activities in the surrounding environment. The subject's own interference mainly stems from the physiological electrical activity of organs such as the eyes, muscles, and heart, while external interference primarily originates from the electrodes recording the signals and the power frequency of the driving equipment. Therefore, to obtain pure neural activity information from the acquired EEG signals, a series of processing steps must be performed on the raw data. Specific operations include deleting useless segments, deleting useless channels, configuring channel localization, low-pass filtering, high-pass filtering, removing power frequency interference, channel rereference, reducing the sampling rate, eliminating bad channels, artifact subspace reconstruction, missing channel reconstruction, independent component decomposition, and independent component selection.

[0122] The samples in the dataset constructed in this invention are EEG segments of a fixed length preceding the stimulus presentation time in the corresponding trials. Therefore, the selection of the segment length can be used as a feature of the sample. Since the minimum value of the random stimulus interval is 2 seconds and the neural activity in the short period before the stimulus is more strongly correlated with the response performance after the stimulus, one of the benchmark feature extraction methods proposed in this invention is to set the length of the EEG segment corresponding to the sample to 1000ms or 2000ms.

[0123] To reflect different types of brain activity and state characteristics, EEG signals can be divided into five main frequency bands: δ (1–4 Hz), θ (4–8 Hz), α (8–14 Hz), β (14–31 Hz), and γ (31–50 Hz). Therefore, this invention selects two frequency band features for extracting EEG segments: logarithmic power, which belongs to the frequency domain, and differential entropy, which belongs to the nonlinear domain, as another benchmark feature extraction method for constructing the dataset.

[0124] In the method of extracting the logarithmic power of each frequency band as a feature, a series of bandpass filtering operations need to be performed on the denoised (80 minutes) complete EEG signal to obtain 5 complete EEG signals of different frequency bands. Then, a certain number of segments containing classification labels can be extracted from each EEG signal according to the method described above, and the natural logarithmic power of these signal segments can be calculated:

[0125]

[0126] Since the actual acquired EEG signals are discrete signals with a certain time resolution, each time series segment x can be calculated using this formula. j The natural logarithmic power value P j In this formula, the subscript j represents the sequence number of the signal segment, M represents the number of sampling points contained in the segment, n represents the sequence number of the sampling point, and x... j [n] represents the signal amplitude at the nth sampling point in the j-th signal segment. Compared to the absolute value of EEG power, its logarithmic value usually shows a stronger linear correlation with cognitive and behavioral performance; the logarithmic power feature can also weaken the influence of extreme samples and reduce individual differences among different subjects.

[0127] In this invention, the noise-reduced EEG data contains a total of 60 channels. By extracting all fatigue and alertness samples and calculating the logarithmic power of each sample in each channel and frequency band, a graph can be plotted as follows. Figure 6 and Figure 7The brain topography map shown. In the mean difference brain topography map, the value of each point is obtained by subtracting the corresponding value of all alert samples from the logarithmic power average of all fatigue samples in that channel and frequency band. The significance test: Each point on the brain topography map represents the significance test result, i.e., the p-value, of the logarithmic power characteristics of fatigue and alert samples in that channel and frequency band. Observation Figure 6 It can be found that the log power of the two types of samples differs significantly in the δ, θ and γ frequency bands, and the channels with large differences in log power in each frequency band are mainly concentrated in the occipital lobe region at the back of the brain. Figure 7 The results can be largely related to Figure 6 The results correspond to the fact that points with larger differences in average values ​​generally have smaller p-values, and the corresponding log power characteristics show significant differences between the two types of samples.

[0128] Differential entropy is a type of entropy used to measure the complexity of continuous random variables. Its calculation formula can be expressed as:

[0129] h(X)=-∫ X f(x)log(f(x))dx;

[0130] In this formula, X represents any random variable, h(X) is its entropy, and f(x) is the probability density function of X. If X is a time series and follows a normal distribution N(μ,σ)... 2 If ), then its differential entropy h(X) can be specifically defined as:

[0131]

[0132] μ and σ in the formula 2 Let x and y be the mean, variance, and x-th element of the time series X, respectively. Although not all sub-band EEG segments obtained through bandpass filtering conform to a normal distribution, the vast majority of them pass the normality test. Therefore, this invention chooses to use the following calculation method to approximately estimate the differential entropy of each sample:

[0133]

[0134] H in the formula j That is, the j-th EEG fragment x j The differential entropy, and It is the j-th EEG fragment x j The variance of the signal. Since differential entropy reflects the complexity of the signal, this characteristic can, to some extent, balance the ability of high-energy low-frequency signals and low-energy high-frequency signals to distinguish neural fatigue patterns.

[0135] Referring to the method described earlier for plotting brain topography maps based on the differences in the mean values ​​of logarithmic power features and significance tests, a set of corresponding brain topography maps can be plotted for the differential entropy features, namely... Figure 8 and Figure 9 .analyze Figure 8 and Figure 9 It can be obtained with Figure 6 and Figure 7 Very similar conclusions, but Figure 9 The results indicate that there are significantly fewer points where the differential entropy differs significantly between the two types of samples than points where the log power differs significantly. In particular, the differential entropy in the α and β bands may not be able to help the model effectively distinguish between fatigued samples and alert samples.

[0136] Based on retaining 2400 labeled samples, this invention proposes four EEG feature extraction methods that can be used as benchmarks, namely, combinations of two fragment lengths and two feature types.

[0137] Verification experiment:

[0138] In the validation experiments, this invention chose XGBoost as the primary benchmark machine learning model because it not only has high computational efficiency and excellent classification performance, but also allows for the evaluation of the importance of each input feature based on the branch gain in tree algorithms, helping users understand which channel or frequency band features play a more significant role in mental fatigue detection. Furthermore, this invention also employs classic RF and SVM methods as benchmark comparison models in some experiments, using their performance to verify the applicability of the dataset across various models.

[0139] As mentioned above, the dataset construction method proposed in this invention can ensure a balanced distribution of samples from different categories, thus the model's performance can be accurately reflected by most classification performance evaluation metrics. Specifically, the classification performance evaluation metrics used in the validation experiments of this invention are accuracy, precision, recall, and F1 score.

[0140] Based on different training and test set partitioning methods, two testing scenarios can be defined for the mental fatigue detection model: within-subject and between-subject. In the within-subject scenario, each subject's samples need to be split into training and test sets. The model is trained using only its own training set and evaluated using only its own test set. For the between-subject scenario, Leave-One-Subject-Out-Cross Validation (LOSO-CV) is typically used. This involves selecting a specific subject's samples as the test set and using samples from all other subjects as the training set. To verify the model's generalization ability across different individuals, both within-subject and between-subject tests need to evaluate the model's mental fatigue detection performance across all subjects. Due to significant individual differences in neural activity among subjects in alert and fatigue states, the model often exhibits poor classification performance in the between-subject scenario because it struggles to learn common neural patterns among individuals. However, training samples with weak signal-noise ratios, effective annotation methods, and appropriate feature extraction can significantly enhance the model's performance in cross-subject mental fatigue detection tasks. Therefore, this invention chooses to utilize the between-subject testing method to verify the quality of the dataset and the model's performance.

[0141] Validation of data annotations:

[0142] This invention proposes a method for labeling samples by combining local reaction time and global reaction time. To verify whether this method is more effective than labeling methods that rely solely on either local or global reaction time, i.e., whether it can improve the model's performance in detecting mental fatigue, this invention uses the differential entropy of a 1000ms EEG signal segment as the input feature of the XGBoost model and conducts comparative experiments on different sample labeling methods based on this condition. The specific results are shown in Table 1.

[0143] Table 1 Summary of mental fatigue detection results using different sample labeling methods

[0144]

[0145]

[0146] Based on the information in Table 1, it can be determined that the sample labeling method proposed in this invention is relatively more effective.

[0147] Validation of the differences in feature extraction:

[0148] In terms of feature extraction, this invention proposes four benchmark feature extraction methods composed of two lengths and two feature combinations of EEG fragments. To compare their performance in the mental fatigue detection task, this invention uses the XGBoost model for relevant tests, and the specific results are shown in Table 2.

[0149] Table 2 Summary of mental fatigue detection results using different feature extraction methods

[0150]

[0151] Analyzing the information in Table 2, it can be inferred that the model learns more neural patterns related to mental fatigue from relatively short time segments (1000ms) rather than relatively long time segments (2000ms). Furthermore, differential entropy features are more suitable as neural markers of mental fatigue than logarithmic power features.

[0152] Applicability verification of multiple models:

[0153] To verify whether the dataset construction method proposed in this invention can be applied to various different machine learning models, this invention uses the differential entropy extracted from a 1000ms EEG signal segment as a baseline feature to test three models—XGBoost, RF, and SVM—with the goal of evaluating their performance in detecting mental fatigue. The specific results are shown in Table 3.

[0154] Table 3 Summary of mental fatigue detection results using different machine learning methods

[0155]

[0156] Table 3 shows that all three models demonstrate good classification ability across subjects' mental fatigue states, validating the general applicability of the dataset constructed in this invention across different models. The XGBoost model exhibits the best performance among the three, achieving a leading advantage in almost all evaluation metrics.

[0157] A brief description of the basic hardware for implementing this invention is provided. A computer system operated by the subject is used to run the PVT experimental program, presenting visual stimuli to the subject via a monitor and receiving their response signals via a keyboard. A specialized EEG signal acquisition device is then required to acquire the EEG signal data generated by the subject during the experiment. Furthermore, the data recorded by the EEG signal acquisition device needs to be stored and processed by a computer system operated by the experimenter. Finally, communication and synchronization between the various main hardware components rely on several high-speed signal cables of different specifications connecting the hardware components.

[0158] This invention, based on the cognitive behavioral principles of PVT and the need to construct a mental fatigue detection dataset, improves some design details and implementation procedures in the standard PVT experimental paradigm, and simultaneously collects the subjects' behavioral performance and EEG signal data during the experiment. By analyzing the task performance data obtained from the PVT experiment, a novel derived task performance index with a stronger correlation to mental fatigue status is proposed, building upon existing reaction time task performance indices. Using these indices, more effective data annotation and cleaning methods are developed, and a certain number of fatigue and alertness samples are selected to construct the dataset. By extracting features from EEG fragments, the annotated EEG fragments are transformed into an input feature matrix form more suitable for training and testing machine learning models. This invention also experimentally verifies the effectiveness of the proposed data annotation method, the differences in feature extraction methods, and the applicability of various models to the constructed dataset, further demonstrating the advanced nature of the proposed mental fatigue detection dataset construction method. Furthermore, this invention provides data support for subsequent cross-dataset mental fatigue detection inventions based on EEG signals.

[0159] The embodiments described above are only used to illustrate the technical ideas and features of the present invention. Their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The patent scope of the present invention should not be limited by these embodiments. That is, any equivalent changes or modifications made in accordance with the spirit disclosed in the present invention still fall within the patent scope of the present invention.

Claims

1. A method for constructing a mental fatigue detection dataset based on a PVT experiment paradigm, characterized in that, The method comprises the following steps: Step 1, inducing mental fatigue of a subject using a PVT experimental paradigm, and synchronously collecting behavioral performance and corresponding electroencephalogram (EEG) data of the subject; Step 2, performing noise reduction preprocessing on the collected EEG data; Step 3, cutting out fixed-length segments corresponding to each trial from the EEG data; Step 4, formulating a plurality of derived performance indicators based on the reaction time performance indicators of the PVT, and using the performance indicators and combinations thereof to represent the mental fatigue state; Step 5, setting classification thresholds, classification criteria and cleaning methods based on the performance indicators and the mental fatigue state represented thereby, and performing classification labeling and cleaning on the EEG segments; Step 6, extracting features of the EEG segments with classification labels remaining after cleaning, compiling an input feature matrix and a label vector, and thereby constructing a data set for training and testing a machine learning model; Step 7, verifying the constructed input feature matrix and label vector data set in one or a combination of the following manners: verifying the effectiveness of data labeling, verifying the difference of feature extraction, and verifying the applicability of multiple models; In step 4, the actual reaction time of each trial is defined as the local reaction time thereof, the local reaction time of each trial and a weighted sum of a plurality of sequentially adjacent trials adjacent to the trial are defined as the global reaction time thereof, and the weight sum of the trials is 1; the local reaction time and / or the global reaction time is used to represent the degree of mental fatigue of the subject, and is used as a performance indicator reflecting the mental fatigue state; let i be the trial number, l1 represent the number of adjacent trials before the i-th trial included in the weighted average, l2 represent the number of adjacent trials after the i-th trial included in the weighted average, and the sum of the number of sequentially adjacent trials be l1+l2+1; then the calculation formula of the global reaction time is as follows: In the formula: w k weight for the kth trial; LRT k LRTk is the local response time for the kth trial. GRT i GRT is the global reaction time for the i-th trial. In step 5, the potential reaction time, the alert reaction time and the fatigue reaction time are set as the classification thresholds of the mental fatigue state; let each subject experience N trials in the experiment, i.e. N local reaction times are generated, and all the local reaction times are sorted from small to large, then the local reaction time with a sorting position of 5%×N is defined as the potential reaction time; the alert reaction time is set to 1.20-1.30 times the potential reaction time; the fatigue reaction time is set to 1.50-2.00 times the potential reaction time; the ratio of the alert reaction time to the fatigue reaction time is fixed when constructing the same data set; for each sample, if the local reaction time and the global reaction time thereof are both less than the alert reaction time thereof, the sample is labeled as an alert sample; if the local reaction time and the global reaction time thereof are both greater than the fatigue reaction time thereof, the sample is labeled as a fatigue sample; samples that neither belong to the alert samples nor belong to the fatigue samples are removed.

2. The PVT experiment paradigm based mental fatigue detection dataset construction method according to claim 1, characterized in that, Step 1 includes the following method steps: during the execution of the PVT task, the subject always gazes at the gaze point pattern displayed on the display, and presses the key immediately after the gaze point pattern is converted into the stimulus pattern; after receiving the key information, the time difference between the stimulus presentation and the key reaction is calculated, which is recorded as the reaction time of the current trial; the reaction end is defined as the stimulus pattern switching back to the gaze point pattern, and the interval time from the reaction end to the next stimulus presentation is randomly selected by the experimental program in the range of 2-10 seconds; in a complete experiment, each subject needs to perform 3-5 PVT tasks, and needs to sit and rest for 5-15 minutes after completing each task; the duration of the first PVT task is set to 20-40 minutes, and the duration of the other PVT tasks is set to 10-20 minutes; the experimental program generates digital markers at the stimulus presentation time and the key reaction time of each trial, and sends them to the electroencephalogram signal acquisition device in time; the time resolution of the acquired electroencephalogram signal and the generated behavior markers is 1000 Hz. 3.The PVT experiment paradigm based mental fatigue detection dataset construction method according to claim 1, wherein, In step 2, the noise reduction preprocessing method includes one or a combination of the following methods: deleting useless segments, deleting useless channels, configuring channel positioning, low-pass filtering, high-pass filtering, removing power frequency interference, channel re-reference, reducing sampling rate, eliminating bad channels, pseudo-trace subspace reconstruction, missing channel reconstruction, independent component decomposition, independent component selection, and band-pass filtering.

4. The PVT experiment paradigm based mental fatigue detection dataset construction method according to claim 1, characterized in that, In step 3, a fixed length of electroencephalogram signal segment before the stimulus presentation time in each trial is extracted.

5. The PVT experiment paradigm based mental fatigue detection dataset construction method according to claim 1, characterized in that, If the proportion of fatigue samples and vigilance samples in the labeled sample set from the same subject is greater than 5 or less than 0.2, it is judged that the sample of the subject has a class imbalance problem, and all samples of the subject are removed and not included in the construction of the data set.

6. The PVT experiment paradigm based mental fatigue detection dataset construction method according to claim 1, characterized in that, In step 6, the method for extracting features of the electroencephalogram signal segment with a classification label includes one or a combination of the following: setting the length of the electroencephalogram signal segment corresponding to the sample to 1000-2000 milliseconds; selecting one or a combination of the following features of the electroencephalogram signal segment: time domain feature, frequency domain feature, space domain feature, nonlinear feature, network feature, or fusion feature.

7. The PVT experiment paradigm based mental fatigue detection dataset construction method according to claim 1, characterized in that, In step 7, XGBoost is used as a benchmark machine learning model, and the differential entropy of each frequency band and each channel extracted from the electroencephalogram signal segment is used as the input feature of the benchmark model, and the effectiveness of the data labeling is verified through the model spirit fatigue detection performance comparison experiment of different sample labeling methods; The benchmark features are used to test the XGBoost, RF and SVM machine learning models to evaluate the spirit fatigue detection performance, and the cross-subject spirit fatigue state classification ability of the three models is used to verify the generality of the constructed data set between different models. 8.The PVT experiment paradigm based mental fatigue detection dataset construction method according to claim 1, characterized in that, In step 7, the difference verification method of feature extraction includes the following methods and steps: taking out two kinds of labeled samples, calculating the electroencephalogram signal features of each channel and each frequency band of each sample, and drawing the average value difference brain topographic map and the significance test brain topographic map. In the average difference brain topography, the value of each point is equal to the average value of the EEG signal feature of the labeled sample of one class in the channel and the frequency band minus the corresponding average value of the labeled sample of the other class; each point on the significance test brain topography represents the significance test result, that is, the p value, of the EEG signal feature value of the labeled sample of the two classes in the channel and the frequency band; XGBoost is used as a benchmark machine learning model, and the performance of the model for mental fatigue detection is compared by taking different EEG signal features as input features of the benchmark model, so as to verify the difference of feature extraction.

Images