Psychological state detection method and system for electroencephalogram signals based on evidence theory fusion

Through a decision-making-level fusion method based on evidence theory, combined with the output of multiple psychological state classifiers, the problem of insufficient accuracy of EEG signal detection methods in cross-data centers and multi-task analysis is solved, and the accuracy and generalization of detection are improved. to adapt to complex psychological state detection needs.

CN119026018APending Publication Date: 2024-11-26LANZHOU UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411102665.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-12
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

The existing psychological state detection methods for EEG signals have problems with insufficient accuracy and generalization ability in generalization ability in cross-data centers and multi-task analysis.

Method used

Using a decision-making-level fusion method based on evidence theory, the output probability values ​​of multiple psychological state classifiers are fused through the Demster-Schafer evidence theory to generate the final psychological state detection results, improving the accuracy and generalization of the detection ability.

Benefits of technology

By fusing decision results from multiple classifiers, the accuracy and robustness of psychological state detection are improved, the adaptability to different data sets and experimental conditions is enhanced, and the risk of overfitting and uncertainty of a single classifier is reduced. Difficulty in sexual handling.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119026018A_ABST
    Figure CN119026018A_ABST
Patent Text Reader

Abstract

The invention provides a psychological state detection method and system for electroencephalogram signals based on evidence theory fusion, and the method can extract electroencephalogram features of the electroencephalogram signals after the electroencephalogram signals of a target object are obtained. Extracting a target electroencephalogram feature with discriminability from the electroencephalogram features on the basis of a feature selection method, inputting the target electroencephalogram feature into a plurality of different psychological state classifiers to obtain classification probability values output by the plurality of psychological state classifiers, fusing the plurality of classification probability values on the basis of an evidence theory to obtain a fusion probability value, and outputting the fusion probability value; and finally obtaining a psychological state detection result of the electroencephalogram signal according to the fusion probability value. The method is based on an evidence theory decision-making level fusion method and a multi-task analysis method, and the accuracy and generalization ability of psychological state detection are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of deep learning technology, and in particular to a method and system for detecting a psychological state of an electroencephalogram signal based on evidence theory fusion. Background Art

[0002] Mental state refers to the complete characteristics of mental activities within a certain period of time. According to the characteristics of mental state, multiple mental states can be distinguished, such as attention, fatigue, tension, relaxation, sadness, joy, etc. Different mental states can produce different performances. Therefore, the user's mental state can be detected and classified based on the user's performance, and then the user's mental state can be detected. For example, users with depression have specific manifestations of mental state, including low emotions, slow thinking, reduced speech and movement, reduced willpower, lack of concentration, memory decline, slowness, etc.

[0003] Mental state detection can be performed by combining EEG signal technology with machine learning algorithms. The use of EEG technology for mental state detection mainly focuses on extracting the features of EEG signals and classifying them through machine learning methods. However, the generalization ability across data centers and multi-task analysis processing will still affect the accuracy and generalization ability of mental state detection, and improve the accuracy of mental state detection. Summary of the invention

[0004] The present application provides a method and system for detecting a psychological state of an electroencephalogram signal based on evidence theory fusion, so as to solve the problem of low accuracy of psychological state detection.

[0005] In a first aspect, the present application provides a method for detecting a psychological state of an EEG signal based on evidence theory fusion, comprising:

[0006] Acquire the EEG signal of the target object, wherein the EEG signal includes a resting EEG signal and a task-state EEG signal, wherein the resting EEG signal is an EEG signal collected when the target object is in a resting state, and the task-state EEG signal is an EEG signal collected when the target object performs a point detection task;

[0007] Extracting electroencephalogram (EEG) features of the EEG signal, wherein the EEG features include a linear feature matrix and a nonlinear feature matrix, wherein the linear feature matrix includes multiple linear features of the EEG signal, and the nonlinear feature matrix includes multiple nonlinear features of the EEG signal;

[0008] Extracting a target EEG feature from the EEG features, wherein the target EEG feature is a discriminative EEG feature;

[0009] Inputting the target EEG feature into a plurality of different psychological state classifiers to obtain classification probability values ​​output by the plurality of psychological state classifiers, wherein the psychological state classifiers are classifiers trained based on EEG signal sample data labeled with different psychological states;

[0010] Based on the evidence theory, a plurality of the classification probability values ​​are fused to obtain a fusion probability value;

[0011] The psychological state detection result of the EEG signal is output according to the fusion probability value.

[0012] Optionally, the step of preprocessing the EEG signal includes:

[0013] Filtering out high-frequency signals with frequencies higher than a first threshold and low-frequency signals with frequencies lower than a second threshold in the EEG signal;

[0014] Eliminating signal artifacts in the EEG signal, wherein the signal artifacts include electrooculographic artifacts and electromyographic artifacts;

[0015] The electroencephalogram signal is divided into electroencephalogram signals of multiple frequency bands.

[0016] Optionally, the step of extracting target EEG features from the EEG features includes:

[0017] Obtaining the guessed nearest neighbor and the wrongly guessed nearest neighbor of each EEG feature, wherein the guessed nearest neighbor is the nearest neighbor of the same category as the EEG feature, and the wrongly guessed nearest neighbor is the nearest neighbor of a different category from the EEG feature;

[0018] Calculating a first distance between the EEG feature and the guessed nearest neighbor, and calculating a second distance between the EEG feature and the guessed nearest neighbor;

[0019] assigning a weight value to the EEG feature according to the first distance and the second distance;

[0020] The EEG features whose weight values ​​are greater than a preset weight threshold are extracted to obtain target EEG features.

[0021] Optionally, the step of extracting target EEG features from the EEG features includes:

[0022] Calculating a first correlation between the EEG feature and the category to which the EEG feature belongs;

[0023] Calculating a second correlation between the EEG features;

[0024] Calculating an evaluation function value according to the first correlation and the second correlation based on an evaluation function;

[0025] Traversing the EEG features;

[0026] The EEG feature set when the evaluation function value is maximum is extracted to obtain the target EEG feature.

[0027] Optionally, the step of extracting target EEG features from the EEG features includes:

[0028] Mapping the EEG feature to a latent feature in a latent feature space, wherein the latent feature is a hidden or deep feature representation of the EEG feature;

[0029] Calculating the importance of the potential feature to the category to which the EEG feature belongs;

[0030] Calculating a score value of the EEG feature based on the importance;

[0031] The EEG features whose score values ​​are greater than a preset evaluation threshold are extracted to obtain target EEG features.

[0032] Optionally, the method further includes:

[0033] Inputting the target EEG feature into a nearest neighbor classifier to obtain a first classification probability value output by the nearest neighbor classifier;

[0034] Inputting the target EEG feature into a decision tree classifier to obtain a second classification probability value output by the decision tree classifier;

[0035] The target EEG feature is input into a support vector machine classifier to obtain a third classification probability value output by the support vector machine classifier.

[0036] Optionally, the method further includes:

[0037] Converting the first classification probability value, the second classification probability value, and the third classification probability value into basic probability distribution values ​​in evidence theory;

[0038] The basic probability distribution values ​​are fused to obtain a fused probability value.

[0039] Optionally, the basic probability distribution values ​​are fused according to the following formula:

[0040]

[0041] Among them, m 123(A) is the fusion probability value, A is the set of psychological state categories of the EEG signal, A1 is the set of psychological state categories output by the nearest neighbor classifier, m1(A1) is the first classification probability value output by the nearest neighbor classifier, A2 is the set of psychological state categories output by the decision tree classifier, m2(A2) is the second classification probability value output by the decision tree classifier, A3 is the set of psychological state categories output by the support vector machine classifier, and m3(A3) is the third classification probability value output by the support vector machine classifier.

[0042] Optionally, the method further includes:

[0043] Obtaining the fusion probability value of the psychological state category of the EEG signal;

[0044] The psychological state category with the largest fusion probability value is selected to obtain the psychological state detection result.

[0045] In a second aspect, the present application provides a psychological state detection system based on EEG signals fused with evidence theory, comprising:

[0046] A signal acquisition module is used to acquire the EEG signal of the target object; the EEG signal includes a resting EEG signal and a task-state EEG signal, the resting EEG signal is the EEG signal collected when the target object is in a resting state, and the task-state EEG signal is the EEG signal collected when the target object performs a point detection task;

[0047] A feature extraction module, used to extract EEG features of the EEG signal, wherein the EEG features include a linear feature matrix and a nonlinear feature matrix, wherein the linear feature matrix includes multiple linear features of the EEG signal, and the nonlinear feature matrix includes multiple nonlinear features of the EEG signal;

[0048] A feature selection module, used to extract a target EEG feature from the EEG features, wherein the target EEG feature is a discriminative EEG feature;

[0049] An evidence theory fusion module, used for inputting the target EEG feature into a plurality of different psychological state classifiers to obtain classification probability values ​​output by the plurality of psychological state classifiers, wherein the psychological state classifiers are classifiers trained based on EEG signal sample data labeled with different psychological states;

[0050] Based on the evidence theory, a plurality of the classification probability values ​​are fused to obtain a fusion probability value;

[0051] The psychological state detection module is used to output the psychological state detection result of the EEG signal according to the fusion probability value.

[0052] It can be seen from the above technical solutions that the present application provides a method and system for detecting the psychological state of EEG signals based on the fusion of evidence theory. The method can extract the EEG features of the EEG signals after acquiring the EEG signals of the target object. The target EEG features with discriminability are extracted from the EEG features based on the feature selection method, and the target EEG features are input into multiple different psychological state classifiers to obtain the classification probability values ​​output by multiple psychological state classifiers, and then the multiple classification probability values ​​are fused based on the evidence theory to obtain the fusion probability value, and finally the psychological state detection result of the EEG signal is obtained according to the fusion probability value. The method is based on the decision-level fusion method of evidence theory and the multi-task analysis method to improve the accuracy and generalization ability of psychological state detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0054] Figure 1 A flowchart of a method for detecting a psychological state of an EEG signal based on evidence theory fusion provided in an embodiment of the present application;

[0055] Figure 2 A schematic diagram of a point detection task provided in an embodiment of the present application;

[0056] Figure 3 A schematic diagram of the preprocessing process provided in the embodiment of the present application;

[0057] Figure 4 A schematic diagram of performing mental state detection based on three mental state classifiers provided in an embodiment of the present application;

[0058] Figure 5 A schematic diagram of a trust interval provided in an embodiment of the present application. DETAILED DESCRIPTION

[0059] The following embodiments are described in detail, and examples thereof are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following embodiments do not represent all implementations consistent with the present application. They are only examples of systems and methods consistent with some aspects of the present application as detailed in the claims.

[0060] In the embodiments of the present application, the mental state refers to the complete characteristics of mental activities within a certain period of time. According to the characteristics of the mental state, a variety of mental states can be distinguished, such as attention, fatigue, tension, relaxation, sadness, joy, etc. Different mental states can produce different performances. Therefore, the user's mental state can be detected and classified based on the user's performance, and then the user's mental state can be detected. For example, a user with depression has specific manifestations of mental state, including low emotions, slow thinking, reduced speech and movement, loss of willpower, lack of concentration, memory decline, slowness, etc.

[0061] In some embodiments, the psychological state can be determined by interviewing and observing the user's behavior. However, this subjective assessment is easily affected by individual experience and opinions, resulting in inconsistency in the detection results.

[0062] In some embodiments, the psychological state can be detected based on a scale, which is a standardized measurement tool used to assess certain psychological traits, behaviors or symptoms of an individual. However, this method is susceptible to individual subjectivity and possible memory bias, which reduces objectivity and the accuracy of psychological state detection. For example, when using a scale to assist in detecting whether a user suffers from depression, the user can fill out a self-report scale, such as the Self-rating Depression Scale (SDS), the Hamilton Depression Rating Scale (HAMD), the Beck Depression Inventory (BDI), etc., and determine whether the user suffers from depression through self-report scale analysis.

[0063] In some embodiments, in order to improve the accuracy of psychological state detection, biological signals can be combined for analysis and detection. Electroencephalogram (EEG) as an objective biological indicator can provide characteristics of brain activity. However, the detection method based on EEG still has the problems of high data processing complexity and low detection accuracy.

[0064] In some embodiments, a machine learning algorithm can be used to analyze EEG signals to establish a more objective and quantitative psychological state detection model. The collected EEG signals can be pre-selected for features, and linear features, nonlinear features, and power spectrum features in the EEG signals can be selected. After the extracted features are normalized, different EEG channels are selected for random forest model construction to obtain a trained psychological state classification model. The detection result of the target EEG signal is output according to the psychological state classification model.

[0065] This method is relatively basic in feature extraction and selection, mainly extracting linear features, nonlinear features and power spectrum features, and does not fully consider the importance of feature selection. Feature selection is crucial to improving the generalization ability of the model and reducing the risk of overfitting. The lack of this step may cause the model to be overly sensitive to noise and irrelevant information, affecting the overall performance of the system.

[0066] Moreover, this method uses a single random forest classifier, and the use of a single classifier will limit the model's ability to adapt to data complexity. When dealing with complex and diverse mental state detection, a single classifier may not be able to capture all potential patterns and features, thereby limiting the performance of the system. At the same time, a single classifier may overfit the training data during the training process, resulting in poor generalization performance on new data. In addition, a single classifier lacks the ability to model and process uncertainty, and uncertainty is a common and important problem in mental state detection. A single classifier cannot effectively handle the detection and classification related to mental states.

[0067] Furthermore, if only part of the dataset or experimental paradigm is used, the applicability of the system under different experimental conditions is not fully considered. A single dataset and experimental paradigm may not fully cover the diverse manifestations of mental states. If the model is trained mainly on data from a specific condition or group, its ability to generalize to other conditions or groups may be limited, reducing the practicality of the system. Mental states are affected by a variety of factors, including individual differences, environmental changes, etc. A single dataset and experimental paradigm may fail to capture these potential factors of variation, causing the model to perform poorly in the face of the diversity of the real world. A single dataset and experimental paradigm may fail to provide sufficiently diverse EEG signal patterns, thereby limiting the model's comprehensive capture of mental state-related features. This may lead to the model's insufficient recognition of a wider range of feature patterns in practical applications. The manifestation of mental states may vary under different experimental conditions, and the use of a single experimental paradigm may not fully account for these differences. Fluctuations in the performance of the model under different experimental conditions may affect its reliability in the real world.

[0068] In some embodiments, mental state detection can be performed based on various classification techniques and feature selection methods of EEG signals. For example, power and asymmetry features in EEG signals can be extracted, and mental state detection can be achieved through multi-cluster feature selection and support vector machine (SVM) classifier. Linear and nonlinear features can be extracted from EEG signals, and mental state detection can be achieved using the parcel subset evaluation feature selection method. However, the above methods focus on improving the model accuracy on specific data sets, ignoring the stability across different data centers and paradigms.

[0069] In some embodiments, it is possible to focus on identifying reliable biomarkers and developing models with robust performance across multiple centers. For example, EEG data of multiple participants in resting states are collected, and the classification performance of four common EEG features is tested. In order to overcome the data variability caused by the increase in data scale and multi-site recording, a conformal kernel (CK) transformation can be introduced. The results based on the above method show that the best feature subset and CK-SVM combination can achieve an accuracy of 91.07% on the training set, while it drops by 7% on the independent test data set. Methods based on linear discriminant analysis and nearest neighbor (K-Nearest Neighbor, KNN) drop by more than 20%. The poor performance on the independent test set may be due to the heterogeneity of psychological state, differences in data sources, and factors such as data acquisition equipment, experimental conditions, and the professional level of data collectors. These factors hinder the generalization of the model, challenge the acquisition of reliable biomarkers and the construction of a psychological state detection model with strong cross-dataset generalization capabilities.

[0070] However, most multi-center analyses focus on single-task analysis, while multi-task analysis can obtain more comprehensive information, covering multiple indicators of emotional state, cognitive function, and attention bias. In addition, multi-task analysis is able to learn shared features and patterns across tasks, thereby mining specific characteristics of psychological states. For example, when assisting in the detection of depression, individuals with depression tend to show higher attention processing to negative stimuli compared to positive events, which is one of the core cognitive risk factors for depression. Therefore, the abnormal attention bias of participants with depression can be analyzed based on the dot detection task, and the classification accuracy can be significantly improved by extracting features based on event-related potentials and adopting correlation feature selection.

[0071] In some embodiments, mental state detection is performed by combining EEG signal technology with machine learning algorithms. In terms of feature selection, feature selection methods such as filters, wrappers, and embeddings can be used. In order to improve the accuracy of mental state detection, it is necessary to explore different categories of features for identification. However, the research on feature selection methods focuses more on the accuracy of model classification, but lacks consideration on the stability and reproducibility of the quantified selected features.

[0072] At the same time, machine learning such as SVM, KNN and deep learning are used in the classification of psychological states. However, a single binary classifier will inevitably ignore the soft information between features and cannot handle cognitive uncertainty. At the same time, multiple classifiers use calculated probabilities to determine the possibility that new data belongs to different categories. Different classifiers often have different classification results for the same sample, making it impossible to combine multiple classifiers for decision making.

[0073] Based on the above implementation methods, it can be seen that the use of EEG technology for psychological state detection mainly focuses on extracting the features of EEG signals and classifying them through machine learning methods, but the generalization ability across data centers and multi-task analysis processing will still affect the accuracy and generalization ability of psychological state detection.

[0074] In order to improve the accuracy of psychological state detection, the embodiment of the present application provides a psychological state detection method of EEG signals based on evidence theory fusion, a decision-level fusion method based on Dempster-Shafer evidence theory (DS) and a multi-task analysis method to improve the accuracy and generalization ability of psychological state detection.

[0075] It can be understood that Demster-Schafer evidence theory is a decision-level fusion method, also known as evidence theory or DS theory. Among various fusion strategies, decision-level fusion has unique advantages. Decision-level fusion can integrate the outputs from different classifiers to produce effective results. DS theory is a framework that can combine all available evidence about a specific event from different sources. DS theory can not only deal with selective uncertainty, but also characterize the narrowable uncertainty caused by lack of knowledge, that is, cognitive uncertainty. In the decision of classification results, this embodiment proposes a new method based on the probability value of DS theory and the combination of embedded uncertainty. The idea is to represent the membership probability of each sample (the probability of belonging to each class), which is provided in the form of a quality function based on multiple binary classifiers in the context of DS theory. This method optimizes the fusion of these quality values ​​by eliminating the uncertainty of the results, thereby improving the accuracy of the classification rate.

[0076] like Figure 1 As shown, Figure 1 It is a flow chart of the psychological state detection method of EEG signals based on evidence theory fusion in the embodiment of the present application, including the steps of signal acquisition, preprocessing, feature extraction, feature selection, evidence theory fusion and psychological state detection.

[0077] In terms of signal acquisition, the EEG signals of the target object can be collected. Among them, the EEG signals include resting EEG signals and task-state EEG signals. Resting EEG signals are EEG signals collected when the target object is in a resting state. Task-state EEG signals are EEG signals collected when the target object performs a point detection task.

[0078] In some embodiments, both resting state EEG signals and task state EEG signals can be collected using a 128-electrode EEG signal sensor, such as HydroCel Geodesic Sensor Net, Net Station, etc. In order to improve the quality of the collected signal, the sampling frequency can be 250 Hz, the reference electrode is the central parietal area (Cz), and the electrode impedance is kept below 50 kΩ.

[0079] Among them, the dot probe task is an experimental paradigm used in psychology and cognitive neuroscience, which aims to detect the target object's attention bias or spatial attention allocation to specific stimuli such as pictures, words or other forms of visual materials.

[0080] For example, Figure 2 As shown in Figure 2, the dot probe task uses three types of emotional face stimulus pairs, such as happy + neutral, sad + neutral, and fearful + neutral, and a dot as a target. Each trial starts with a fixed white cross appearing in the center of the screen for 300-600 milliseconds. Then, the cross is displayed in the center throughout the experiment. A pair of emotional face stimulus pairs with different emotional values ​​are presented on the screen as cues for 500 milliseconds. Each pair of emotional face stimulus pairs is presented separately in a pseudo-random order. After a short interval (100 milliseconds to 300 milliseconds), the dot probe is randomly presented on the left or right side of the screen as a target for 150 milliseconds. At the same time, the target subject is required to identify and respond to the spatial location of the dot as quickly as possible. And press the "1" key on the configured response box to indicate that it is on the left or press the "4" key to indicate that it is on the right. If the dot appears on the left side of the fixed cross, the target subject should press the "1" key; if the dot appears on the right side, the target subject should press the "4" key on the response box. An automatic interval of 2000 milliseconds is set to receive the target subject's response. Otherwise, the target subject proceeded to the next trial, followed by a black screen presented for 600 ms. This process was repeated until a block was completed. Each block was run in a cyclic manner until the entire session was completed, which consisted of three blocks corresponding to the three types of emotional face stimulus pairs. Each block contained 160 trials, for a total of 480 trials.

[0081] In this embodiment, using data sets under two different experimental paradigms helps to more comprehensively consider the diversity and complexity of psychological states. Different experimental conditions may cause changes in the EEG signal patterns of the target object. By verifying under different experimental conditions, the changes in EEG signals can be better captured. And using different data sets can better evaluate the generalization performance of the model. It helps to ensure that the model remains robust and accurate when faced with new, unseen data. The diversity of verification also helps to reduce dependence on specific experimental conditions, making the evaluation of the model more comprehensive and reliable.

[0082] In terms of preprocessing, the Automagic toolbox can be used to perform standardized EEG signal preprocessing on resting state EEG signals and task state EEG signals. Figure 3 As shown in FIG. 1 , preprocessing includes steps such as re-referencing, filtering, interpolating bad leads, removing bad segments, and segmenting. Re-referencing is the process of adjusting the potential value of each electrode point in the EEG signal relative to one or more reference electrodes to reduce interference introduced by improper reference electrode position or noise of the reference electrode itself, and improve the reliability and consistency of the data. For example, methods such as whole-brain average reference and zero reference can be used.

[0083] Filtering is to filter out high-frequency signals with frequencies higher than the first threshold and low-frequency signals with frequencies lower than the second threshold in the EEG signal to reduce the interference of noise and artifacts on the data. For example, a finite impulse response filter (FIR) can be used to filter low-frequency signals below 1 Hz and high-frequency signals above 40 Hz, retaining EEG signals between 1-40 Hz.

[0084] Interpolation of bad conductors means that when the data of some electrode points cannot be used due to electrode detachment, high impedance, etc., the data of these bad conductors are filled by the data of surrounding electrode points. For example, the bad electrodes can be interpolated based on the average value of the surrounding electrode points, weighted average value, and other interpolation algorithms.

[0085] Bad segment removal refers to eliminating signal artifacts in EEG signals, including electrooculographic artifacts and electromyographic artifacts. For example, the Independent Component Analysis (ICA) algorithm can be used to analyze EEG signals to remove electrooculographic artifacts and electromyographic artifacts. The ICA algorithm can separate interfering signals such as electrooculographic artifacts and electromyographic artifacts, retaining pure EEG signals.

[0086] Segmentation refers to cutting the continuous EEG signal into multiple independent signal segments. For example, for resting EEG signals, the EEG signal of the target object for 120 seconds can be intercepted and segmented into 4-second intervals to obtain 30 signal segments. For task-state EEG signals, use a data analysis toolbox such as ERPLAB to extract the EEG signals of 200ms and 600ms before and after the emotional face stimulus pair appears in the point detection task. 160 signal segments can be obtained for each stimulus type, and a total of 480 signal segments can be obtained.

[0087] The EEG signal can also be divided into EEG signals of multiple frequency bands to form multiple EEG signal segments. For example, when assisting in detecting whether the target object suffers from depression, in terms of spectral characteristics, the EEG signal includes different frequency bands, such as Delta waves at 1-4Hz, Theta waves at 4-8Hz, Alpha waves at 8-13Hz, and Beta waves at 13-30Hz. The target object with depression has abnormal performance in the power of low frequency bands such as delta waves and theta waves and high frequency bands such as beta waves. Because the brain is in a lower arousal state, the power of the delta wave and theta wave frequency bands of the target object with depression is higher, while the power of the beta wave frequency band is lower, which can reflect the weakening of its brain activity and the decline of cognitive function. Therefore, the EEG signal can be divided into EEG signals of these four frequency bands by a Hanning filter.

[0088] In terms of feature extraction, the EEG features of EEG signals can be extracted. The EEG features include linear feature matrices and nonlinear feature matrices. The linear feature matrix includes multiple linear features of the EEG signals, and the nonlinear feature matrix includes multiple nonlinear features of the EEG signals. Among them, linear features include frequency domain features, time domain features, and time-frequency domain features, which can be obtained through Fourier transform, wavelet transform and other methods. Nonlinear features refer to the features of EEG signals obtained through nonlinear dynamic system analysis, such as complexity, approximate entropy, etc.

[0089] In some embodiments, linear features can reveal the basic statistical characteristics of EEG signals, and 8 linear features are extracted for EEG signals of each frequency band, including: the mean value of the peak-to-peak value (PPmean), mean square error (meanSquare), variance (variance), activity of EEG signals, mobility of EEG signals, complexity of EEG signals, maximum power spectral density (maxs) and total power (sumPower). Among them, the activity of EEG signals, the mobility of EEG signals and the complexity of EEG signals can be calculated based on Hjorth parameters, which are a set of statistics for quantifying the complexity of time series signals, including activity, mobility and complexity. The maximum power spectral density (maxs) and total power (sumPower) can be calculated based on an adaptive autoregressive model. The adaptive autoregressive model is a statistical model used for time series analysis.

[0090] Therefore, the feature dimension of the linear feature matrix is ​​1024, that is, 128 electrodes × 8 linear features.

[0091] In some embodiments, the human brain is a dynamic system with complexity, self-organization and nonlinear characteristics. Nonlinear dynamic parameters are suitable for describing the non-stationary characteristics of the brain. Three nonlinear features are extracted for the EEG signal of each frequency band, including: Lempel-Ziv complexity (LZC), singular entropy (Singular) and approximate entropy (ApEn). Among them, Lempel-Ziv complexity is a nonlinear dynamic parameter used to quantify the complexity of data sequences. Singular entropy is an average entropy parameter that measures the complexity of a one-dimensional time series, and approximate entropy is used to measure the irregularity and complexity of a signal sequence. The more complex the time series, the larger the corresponding approximate entropy, and the more regular the time series, the smaller the corresponding approximate entropy.

[0092] Therefore, the feature dimension of the nonlinear feature matrix is ​​384, that is, 128 electrodes × 3 nonlinear features.

[0093] In terms of feature selection, as a dimensionality reduction technique, feature selection aims to select a set of relevant target EEG features, i.e., discriminative EEG features, from the original EEG signal features by removing irrelevant, redundant, or noisy features. Feature selection can achieve better learning performance, higher learning accuracy, lower computational cost, and better model interpretability.

[0094] In some embodiments, the feature selection method of the Relieff algorithm (RF) can be used to extract the most discriminative EEG features from the EEG features. The Relieff algorithm assigns different weights to features according to the correlation between each feature and the category, and features with weights less than a certain threshold will be removed. The correlation between features and categories in the Relief algorithm is based on the ability of the features to distinguish close samples.

[0095] That is, the near-hit and near-miss of each EEG feature can be obtained. The near-hit is the nearest neighbor of the same category as the EEG feature, and the near-miss is the nearest neighbor of a different category from the EEG feature. The first distance between the EEG feature and the near-hit is calculated, and the second distance between the EEG feature and the near-miss is calculated. The EEG feature is assigned a weight value according to the first distance and the second distance. The EEG features whose weight values ​​are greater than the preset weight threshold are extracted to obtain the target EEG features.

[0096] It is understandable that the Relief algorithm uses a "correlation statistic" to measure the importance of EEG features. This statistic is a vector, in which each component represents the importance of the corresponding EEG feature. Therefore, the appropriate EEG feature subset can be selected based on the size of each component in the correlation statistic.

[0097] For each EEG feature xi in the EEG signal, first find the nearest neighbor of the same category and the nearest neighbor of a different category as the EEG feature xi, which are called the near-hit and the near-miss respectively. Then, each component of the relevant statistics can be calculated according to the following formula:

[0098]

[0099] Among them, δ j is the component that represents the relevant statistic corresponding to attribute j. is the value of the EEG feature at the jth attribute. is the value of the guessed neighbor of the EEG feature on the jth attribute, is the value of the wrongly guessed neighbor of the EEG feature on the jth attribute, express and The difference in the jth attribute, express and The difference on the j-th attribute.

[0100] It can be understood that for the guessed nearest neighbor, the smaller the distance between the two on attribute j, the larger the component value. For the wrongly guessed nearest neighbor, the larger the distance between the two on attribute j, the larger the component value. Among them, if attribute j is a discrete attribute, diff takes the Hamming distance, that is, if the two are the same, it takes 0, and if the two are different, it takes 1. If attribute j is a continuous attribute, diff takes the Manhattan distance, that is, the absolute value of the difference between the two is taken. Calculate each component separately, and finally take the average to get the entire related statistic as the weight value of the EEG feature, so as to select the weight value based on the weight value.

[0101] In some embodiments, the feature selection method of correlation-based feature selection (CFS) can be used to extract the most discriminative EEG features from the EEG features. CFS is a feature selection method based on feature correlation, which improves model performance by selecting feature subsets that are highly correlated with the target variable but have low correlation with each other. The goal of CFS is to find the optimal feature subset so that the features in the subset have the greatest information gain for the target variable. The core of CFS is to evaluate the value of feature subsets in an heuristic way, and to optimize the feature selection process by evaluating the overall quality of feature subsets rather than the quality of individual features.

[0102] That is, the first correlation between the EEG feature and the category to which the EEG feature belongs can be calculated, and the second correlation between the EEG features can be calculated. The evaluation function value is calculated based on the first correlation and the second correlation. The EEG features are traversed, and the EEG feature set when the evaluation function value is the largest is extracted to obtain the target EEG feature.

[0103] That is to say, first calculate the correlation between each EEG feature and its category, as well as the correlation between EEG features. Define an evaluation function, namely the Merit function, which is used to calculate the value of the EEG feature set. The evaluation function value can be calculated according to the following formula:

[0104]

[0105] Among them, Merit s is the evaluation function value of the EEG feature set s. k is the number of features in the EEG feature set s. It is the average correlation between each EEG feature and the category in the EEG feature set. is the average correlation between EEG features in the EEG feature set.

[0106] By traversing all EEG features, the EEG feature set that maximizes the evaluation function value is found as the final selected EEG feature.

[0107] In some embodiments, the most discriminative EEG features can be extracted from EEG features using the feature selection method of infinite latent feature selection (ILFS). ILFS uses infinite projection of the latent feature space to evaluate the importance of features. The most discriminative features are selected by considering all possible feature subsets and calculating their contribution to the target variable.

[0108] That is, the EEG features are mapped to latent features in the latent feature space, which are hidden or deep feature representations of the EEG features. It can be understood that the latent feature space is to learn and extract the intrinsic, hidden or deep feature representations of the data. In this space, the features are converted into a set of low-dimensional, highly abstract feature vectors that can capture the essential properties and laws of the features.

[0109] The ILFS algorithm represents the original EEG features as one or more latent features in the latent feature space. Then the importance of the latent features to the category to which the EEG features belong is calculated, that is, the weight of each latent feature is calculated. Then the score of the EEG feature is calculated based on the importance, where the score represents the importance of the EEG feature. The importance of the original EEG feature can be evaluated by the weight and variance of the latent feature. The score of the EEG feature can be calculated according to the following formula:

[0110]

[0111] Among them, ILFS(X) is the score value of EEG feature X, λ i is the weight of the ith potential feature of the EEG feature X. is the variance of the ith potential feature of the EEG feature X. After obtaining the score value of each EEG feature, the EEG feature with a score value greater than a preset evaluation threshold can be selected as the target EEG feature.

[0112] In this embodiment, by introducing multiple feature selection methods such as Relief, CFS, ILFS, etc., the features in the EEG signals are screened to select the most discriminative and relevant features, thereby improving the robustness of the system and the accuracy of psychological state detection, and making the model better adapt to the complexity of real data.

[0113] In terms of evidence theory fusion, the target EEG features can be input into multiple different mental state classifiers to obtain classification probability values ​​output by multiple mental state classifiers, and multiple classification probability values ​​can be fused based on evidence theory to obtain a fusion probability value. Among them, the mental state classifier is a classifier trained based on EEG signal sample data labeled with different mental states.

[0114] In some embodiments, Figure 4 As shown, three mental state classifiers can be used, namely, the nearest neighbor (K-Nearest Neighbor, KNN) classifier, the decision tree (Decision Tree, DT) classifier and the support vector machine (Support Vector Machine, SVM) classifier.

[0115] Among them, the nearest neighbor classifier can select K nearest neighbors by calculating the distance between the new data point and the neighbors in the training data, and perform majority voting to determine the probability of the category. The probability estimate is obtained by counting the proportion of samples belonging to a certain category among the K neighbors. In the decision tree classifier, the decision tree is a tree structure in which each internal node represents a test on an attribute, each branch represents a test output, and each leaf node represents a category. The probability estimate is achieved by counting the proportion of training samples belonging to different categories on each leaf node. The new data point reaches the leaf node through the path of the decision tree, and then the probability of each category is calculated based on the proportion of training samples of each category on the node. The support vector machine classifier can output a raw score as a measure of whether the data point belongs to a category. In order to obtain the probability value of the classification, Platt scaling can be used. Platt scaling learns a Logistic function through the training data, maps the raw score to the range of [0, 1], and thus obtains the probability estimate of the category.

[0116] Therefore, after feature selection, the most discriminative target EEG features are respectively input into the nearest neighbor classifier, decision tree classifier and support vector machine classifier to obtain the first classification probability value output by the nearest neighbor classifier, the second classification probability value output by the decision tree classifier and the third classification probability value output by the support vector machine classifier.

[0117] In this embodiment, by introducing multiple classifiers and integrating their decision results, the system can more comprehensively capture complex patterns related to psychological states. Multi-classifier fusion not only improves the robustness of the model, but also enhances its adaptability to different types of data, thereby improving the overall performance of the system.

[0118] The DS theory is explained below.

[0119] In DS theory, there is a non-empty set V as the identification framework. The non-empty set V contains n mutually exclusive events, that is, V = {v1, v2,..., vn}. The non-empty set V is a finite and exhaustive set of answers to a hypothesis or question (identification framework), that is, the n mutually exclusive events in V can all be answers to the question, and the question always has only one correct answer.

[0120] The Basic Probability Assignment (BPA) in the recognition framework is a mapping 2 V A function from 0 to [0, 1] is called a mass function and satisfies the following rules:

[0121] m (φ) =0

[0122]

[0123] 0≤m (A) ≤1,A∈2 V

[0124] Among them, m(A) represents the score of all evidential beliefs assigned to A, that is, the degree of support for A. A is one or more events in the recognition framework, and A with m(A)>0 is called a focal element.

[0125] The above formula represents the basic probability assignment and combination rules in the identification framework, which is used to integrate evidence from different sources of evidence. In DS theory, there are two functions: Belief Function (Bel) and Plausibility Function (P1), which are used to model uncertainty and can be derived from the basic probability assignment (BPA). The two functions can solve the limitations of classical probability theory in dealing with epistemic uncertainty. The trust function Bel represents the lower bound of BPA, expressing the degree of trust in the correctness of a statement. In contrast, the credibility function P1 represents the upper bound of BPA, indicating the maximum probability assigned to A, or the range of lack of confidence in the negation of A. The trust function Bel and the credibility function Pl can be defined as follows:

[0126]

[0127] In DS theory, for a hypothesis A in the identification framework V, the trust function Bel(A) and the likelihood function Pl(A) of the hypothesis are calculated according to the basic probability distribution BPA to form a trust interval [Bel(A), Pl(A)], which is used to indicate the degree of confirmation of a hypothesis.

[0128] For example, Figure 5 As shown in FIG. 1 , the figure illustrates an example in which the trust function value and credibility function value of an event are 0.6 and 0.8 respectively, and the probability is between these two values. In this case, the uncertainty is calculated as 0.8-0.6=0.2.

[0129] In summary, DS theory provides a formal framework for combining and synthesizing independent evidence. Dempster's combination rule is used to combine the BPAs from two independent information sources to obtain a combined BPA. To elaborate on Dempster's combination rule, consider two mess functions associated with elements of the same recognition frame, m1 and m2, whose focal elements are A and B respectively. If these mess functions are independent, then Dempster's combination rule can be applied to combine their information. The resulting combination is called the joint mass, denoted as m12, and is calculated as follows:

[0130]

[0131] Where B and C are subsets from two different information sources. m1(B) represents the basic probability assignment (BPA) of subset B from information source 1, and m2(C) represents the basic probability assignment (BPA) of subset C from information source 2. K is the conflict coefficient, which represents the total conflict degree between all subsets.

[0132] Similarly, the Dempster combination rule for n mess functions is:

[0133]

[0134] Among them, A1, A2...A n are subsets from different information sources. m1(A1), m2(A2)...m n (A n ) represents the basic probability assignment (BPA) of different subsets. K is the conflict coefficient, which represents the total conflict degree between all subsets.

[0135] Based on the above combination formula, information from different sources can be integrated to obtain a new basic probability assignment (BPA), thereby drawing more reliable conclusions in uncertain environments. By fusing the output results of different mental state classifiers, each mental state classifier has its own unique preferences, algorithms, and characteristics, thus avoiding the limitations of a single decision source and improving accuracy and robustness. DS theory can convert the predictions of mental state classifiers into evidence and use combination rules to deal with uncertainty and conflict. At the same time, it also provides interpretability by calculating confidence and uncertainty, thereby providing better decision support.

[0136] The probability values ​​of different mental state classifiers can be converted into the basic probability assignment value BPA in the evidence theory based on the probability conversion function. For example, the proportional function y=x can be used for conversion, or other linear functions can be used. After the conversion, the converted BPA can be combined according to the combination rules of the DS theory to calculate the trust function value (Bel) and the credibility function value (Pl) of each mental state category, so as to obtain the final fusion probability value predicted for different mental state categories, thereby obtaining the final mental state category of the EEG signal.

[0137] In some embodiments, Figure 4 As shown, when the above three psychological state classifiers are used, the first classification probability value output by the nearest neighbor classifier, the second classification probability value output by the decision tree classifier, and the third classification probability value output by the support vector machine classifier can be transformed according to the proportional function y=x to the basic probability distribution value BPA in the evidence theory, that is, the mess function value. Then, the transformed basic probability distribution value is fused according to the following formula:

[0138]

[0139] Among them, m 123(A) is the fusion probability value, A is the set of psychological state categories of the EEG signal, A1 is the set of psychological state categories output by the nearest neighbor classifier, m1(A1) is the first classification probability value output by the nearest neighbor classifier, A2 is the set of psychological state categories output by the decision tree classifier, m2(A2) is the second classification probability value output by the decision tree classifier, A3 is the set of psychological state categories output by the support vector machine classifier, and m3(A3) is the third classification probability value output by the support vector machine classifier.

[0140] Thus, after obtaining the fusion probability values ​​of the psychological state categories of the EEG signal, the psychological state category with the largest fusion probability value can be selected as the psychological state detection result.

[0141] For example, when assisting in detecting whether a target object suffers from depression, the classification probability values ​​of the three mental state classifiers for binary data labels (the first classification indicating suffering from depression and the first classification indicating not suffering from depression) can be extracted, and the classification probability values ​​obtained by the three mental state classifiers can be converted into mass function values ​​in DS theory through the proportional function y=x, and then fused into a unified mass function value according to the combination rules of evidence theory to calculate the final predicted probability value for the binary label.

[0142] The method of fusing multiple classifiers based on DS theory can effectively deal with uncertainty and help better deal with the complex and multi-source uncertainty that may exist in the field of mental state detection. Secondly, through information fusion, the system can make full use of the advantages of multiple classifiers and improve the overall performance. The error propagation control ability of DS theory can slow down the propagation of wrong decisions of a single classifier, avoid the risk of overfitting of a single classifier, and increase the robustness of the system. In addition, DS theory can adapt to the conflict and consistency between different sources of evidence, provide a flexible processing method for the multi-classifier system, and enable the system to better adapt to complex situations.

[0143] In this embodiment, a resting state data set and a point detection task state data set are used, and the universality is improved by a multi-center and multi-task data set. In terms of data preprocessing, basic operations such as filtering and removing artifacts are performed on the EEG signal to improve the signal quality. The EEG signal is divided into four frequency bands, and based on the different experimental paradigms of the resting state data and the task state data itself, the data segmentation processing is carried out in a targeted manner. In terms of feature extraction, 8 linear features and 3 nonlinear features are extracted to better characterize the nature of the EEG signal. In terms of feature selection, three feature selection methods are used to further sort and screen the extracted features to improve the operating efficiency and accuracy of the system. In terms of the final DS theory fusion and psychological state detection, the predicted probability values ​​for the psychological state categories are first obtained based on the nearest neighbor, decision tree, and support vector machine classifiers, and then the three probability values ​​are converted into the basic probability distribution values ​​of the DS theory through a linear conversion function, and then the probability fusion is performed to obtain the final prediction result.

[0144] Based on the above-mentioned psychological state detection method of EEG signals based on evidence theory fusion, an embodiment of the present application also provides a psychological state detection system of EEG signals based on evidence theory fusion, the system comprising a signal acquisition module, a feature extraction module, a feature selection module, an evidence theory fusion module and a psychological state detection module.

[0145] The signal acquisition module is used to acquire the EEG signal of the target object.

[0146] Among them, the EEG signals include resting-state EEG signals and task-state EEG signals. The resting-state EEG signals are EEG signals collected when the target object is in a resting state, and the task-state EEG signals are EEG signals collected when the target object performs a point detection task.

[0147] The feature extraction module is used to extract the EEG features of the EEG signal.

[0148] The EEG features include a linear feature matrix and a nonlinear feature matrix. The linear feature matrix includes multiple linear features of the EEG signal, and the nonlinear feature matrix includes multiple nonlinear features of the EEG signal.

[0149] The feature selection module is used to extract target EEG features from EEG features, where the target EEG features are discriminative EEG features.

[0150] The evidence theory fusion module is used to input the target EEG features into multiple different psychological state classifiers to obtain classification probability values ​​output by multiple psychological state classifiers, and fuse the multiple classification probability values ​​based on the evidence theory to obtain a fusion probability value. The psychological state classifier is a classifier trained based on EEG signal sample data marked with different psychological states.

[0151] The psychological state detection module is used to output the psychological state detection result of the EEG signal according to the fusion probability value.

[0152] It can be seen from the above technical solutions that the present application provides a method and system for detecting the psychological state of EEG signals based on the fusion of evidence theory. The method can extract the EEG features of the EEG signals after acquiring the EEG signals of the target object. The target EEG features with discriminability are extracted from the EEG features based on the feature selection method, and the target EEG features are input into multiple different psychological state classifiers to obtain the classification probability values ​​output by multiple psychological state classifiers, and then the multiple classification probability values ​​are fused based on the evidence theory to obtain the fusion probability value, and finally the psychological state detection result of the EEG signal is obtained according to the fusion probability value. The method is based on the decision-level fusion method of evidence theory and the multi-task analysis method to improve the accuracy and generalization ability of psychological state detection.

[0153] Similar parts between the embodiments provided in this application can be referenced to each other. The specific implementation methods provided above are only a few examples under the general concept of this application and do not constitute a limitation on the protection scope of this application. For those skilled in the art, any other implementation methods expanded based on the scheme of this application without creative work belong to the protection scope of this application.

Claims

1. A psychological state detection method based on EEG signals fusion with evidence theory, It is characterized in that include: Acquire the EEG signal of the target object, wherein the EEG signal includes a resting EEG signal and a task-state EEG signal, wherein the resting EEG signal is an EEG signal collected when the target object is in a resting state, and the task-state EEG signal is an EEG signal collected when the target object performs a point detection task; Extracting electroencephalogram (EEG) features of the EEG signal, wherein the EEG features include a linear feature matrix and a nonlinear feature matrix, wherein the linear feature matrix includes multiple linear features of the EEG signal, and the nonlinear feature matrix includes multiple nonlinear features of the EEG signal; Extracting a target EEG feature from the EEG features, wherein the target EEG feature is a discriminative EEG feature; Inputting the target EEG feature into a plurality of different psychological state classifiers to obtain classification probability values ​​output by the plurality of psychological state classifiers, wherein the psychological state classifiers are classifiers trained based on EEG signal sample data labeled with different psychological states; Based on the evidence theory, a plurality of the classification probability values ​​are fused to obtain a fusion probability value; The psychological state detection result of the EEG signal is output according to the fusion probability value.

2. The method for detecting the psychological state of EEG signals based on evidence theory fusion according to claim 1, It is characterized in that The step of preprocessing the EEG signal comprises: Filtering out high-frequency signals with frequencies higher than a first threshold and low-frequency signals with frequencies lower than a second threshold in the EEG signal; Eliminating signal artifacts in the EEG signal, wherein the signal artifacts include electrooculographic artifacts and electromyographic artifacts; The electroencephalogram signal is divided into electroencephalogram signals of multiple frequency bands.

3. The method for detecting the psychological state of EEG signals based on evidence theory fusion according to claim 1, It is characterized in that The step of extracting target EEG features from the EEG features comprises: Obtaining the guessed nearest neighbor and the wrongly guessed nearest neighbor of each EEG feature, wherein the guessed nearest neighbor is the nearest neighbor of the same category as the EEG feature, and the wrongly guessed nearest neighbor is the nearest neighbor of a different category from the EEG feature; Calculating a first distance between the EEG feature and the guessed nearest neighbor, and calculating a second distance between the EEG feature and the guessed nearest neighbor; assigning a weight value to the EEG feature according to the first distance and the second distance; The EEG features whose weight values ​​are greater than a preset weight threshold are extracted to obtain target EEG features.

4. The method for detecting the psychological state of EEG signals based on evidence theory fusion according to claim 1, It is characterized in that The step of extracting target EEG features from the EEG features comprises: Calculating a first correlation between the EEG feature and the category to which the EEG feature belongs; Calculating a second correlation between the EEG features; Calculating an evaluation function value according to the first correlation and the second correlation based on an evaluation function; Traversing the EEG features; The EEG feature set when the evaluation function value is maximum is extracted to obtain the target EEG feature.

5. The method for detecting the psychological state of EEG signals based on evidence theory fusion according to claim 1, It is characterized in that The step of extracting target EEG features from the EEG features comprises: Mapping the EEG feature to a latent feature in a latent feature space, wherein the latent feature is a hidden or deep feature representation of the EEG feature; Calculating the importance of the potential feature to the category to which the EEG feature belongs; Calculating a score value of the EEG feature based on the importance; The EEG features whose score values ​​are greater than a preset evaluation threshold are extracted to obtain target EEG features.

6. The method for detecting the psychological state of EEG signals based on evidence theory fusion according to claim 1, It is characterized in that The method further comprises: Inputting the target EEG feature into a nearest neighbor classifier to obtain a first classification probability value output by the nearest neighbor classifier; Inputting the target EEG feature into a decision tree classifier to obtain a second classification probability value output by the decision tree classifier; The target EEG feature is input into a support vector machine classifier to obtain a third classification probability value output by the support vector machine classifier.

7. The method for detecting mental state of EEG signals based on evidence theory fusion according to claim 6, It is characterized in that The method further comprises: Converting the first classification probability value, the second classification probability value, and the third classification probability value into basic probability distribution values ​​in evidence theory; The basic probability distribution values ​​are fused to obtain a fused probability value.

8. The method for detecting mental state of EEG signals based on evidence theory fusion according to claim 7, It is characterized in that The basic probability distribution values ​​are combined as follows: Among them, m 123 (A) is the fusion probability value, A is the psychological state category set of EEG signals, and A 1 is the set of mental state categories output by the nearest neighbor classifier, m 1 (A 1 ) is the first classification probability value output by the nearest neighbor classifier, A 2 is the set of psychological state categories output by the decision tree classifier, m 2 (A 2 ) is the second classification probability value output by the decision tree classifier, A3 is the set of psychological state categories output by the support vector machine classifier, m 3 (A 3 ) is the third classification probability value output by the support vector machine classifier.

9. The method for detecting mental state of EEG signals based on evidence theory fusion according to claim 8, It is characterized in that The method further comprises: Obtaining the fusion probability value of the psychological state category of the EEG signal; The psychological state category with the largest fusion probability value is selected to obtain the psychological state detection result.

10. A psychological state detection system based on EEG signals fused with evidence theory, It is characterized in that include: A signal acquisition module, used to acquire the EEG signal of the target object; The EEG signals include resting EEG signals and task-state EEG signals, wherein the resting EEG signals are EEG signals collected when the target object is in a resting state, and the task-state EEG signals are EEG signals collected when the target object performs a point detection task; A feature extraction module, used to extract EEG features of the EEG signal, wherein the EEG features include a linear feature matrix and a nonlinear feature matrix, wherein the linear feature matrix includes multiple linear features of the EEG signal, and the nonlinear feature matrix includes multiple nonlinear features of the EEG signal; A feature selection module, used to extract a target EEG feature from the EEG features, wherein the target EEG feature is a discriminative EEG feature; An evidence theory fusion module, used for inputting the target EEG feature into a plurality of different psychological state classifiers to obtain classification probability values ​​output by the plurality of psychological state classifiers, wherein the psychological state classifiers are classifiers trained based on EEG signal sample data labeled with different psychological states; Based on the evidence theory, a plurality of the classification probability values ​​are fused to obtain a fusion probability value; The psychological state detection module is used to output the psychological state detection result of the EEG signal according to the fusion probability value.