Cognitive overload online detection method and system based on electroencephalogram feature selection

Abstract

The invention discloses a cognitive overload online detection method and system based on electroencephalogram feature selection, and the method comprises the steps: 1, building a cognitive overload state task model, designing a cognitive overload task according to the cognitive overload state task model, and collecting electroencephalogram signals of participants in the process of completing the cognitive overload task; 2, calculating an intrinsic mode function of the electroencephalogram signal according to a variational mode decomposition method, processing the decomposed electroencephalogram signal according to short-time Fourier transform, and extracting a frequency spectrum, a power spectrum density and a differential entropy feature of each channel to construct a feature candidate set; and step 3, screening features in the feature candidate set according to a fusion feature selection method of mutual information and neighborhood component analysis to obtain screened features. The invention provides a fusion feature selection method of mutual information and neighborhood component analysis to screen signal features, and the accuracy of cognitive load overload detection can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of electroencephalogram (EEG) signal analysis, and in particular to an online detection method and system for cognitive load overload based on EEG feature selection. Background Art

[0002] Cognitive workload reflects the effort the brain puts into meeting task requirements, and in human factors research, the inverted U model has been used to describe the relationship between workload and performance. The model assumes that as demands increase, humans maintain their performance by investing more energy in tasks and mobilizing more resources. However, when task demands become excessive, it can lead to a sharp decline in task performance, which is known as cognitive overload. As an overloaded state of cognitive workload, cognitive overload makes it difficult for individuals to effectively process and understand information, and greatly increases the risk of human error. Therefore, cognitive overload assessment can identify the risk of individuals experiencing cognitive overload and optimize task design through feedback to avoid human errors.

[0003] At present, the common methods of cognitive load assessment can be divided into subjective measurement, task performance and psychophysiological measurement. Although the subjective measurement and task performance methods are easy to implement, they cannot produce real-time results. In psychophysiological measurement, EEG is the behavior of electrical signals caused by brain activity, which can reflect the electrophysiological activity of the brain in real time. Due to its high temporal resolution, convenience, non-invasiveness and low cost, it has become one of the effective methods for detecting brain conditions. Many studies have confirmed that EEG signals can effectively distinguish different levels of cognitive load in specific tasks. However, these studies did not separately analyze the cognitive overload state that may occur in cognitive load.

[0004] Recently, some studies have attempted to explore changes in neural activity before and after cognitive overload by analyzing EEG signals. Kosachenko et al. used a digit span task to induce a memory overload state. Through joint recording and analysis of electroencephalogram (EEG) and pupillometry, they found that theta wave activity and pupil size decreased as the memory overload state was reached. Morton et al. used a tangram puzzle task and a working memory task to induce cognitive overload and found that the alpha power of parietal electrodes was lower under the overload condition compared with the high-load and low-load conditions. Hu et al. used a picture selection task to induce a choice overload state in individuals. Event-related potential (ERP) results confirmed that choice overload impaired early processing by reducing the amplitude of P1 (exogenous component) and P2 (exogenous component). In summary, these studies have confirmed that there are differences in EEG signals before and after cognitive overload, but no study has attempted to apply these differences to early warning of cognitive overload, and no method has been developed to achieve online detection of cognitive overload. Summary of the invention

[0005] In order to solve the problem of low accuracy of cognitive load overload detection based on differences in EEG signals, the present invention provides an online detection method and system for cognitive load overload based on EEG feature selection, which decomposes EEG signals through a variational mode decomposition (VMD) method, then reduces the feature dimension through the fusion of mutual information (MI) and neighborhood component analysis (NCA), and inputs the screened features into a machine learning method to complete online detection of cognitive load overload.

[0006] In order to achieve the above object, the technical solution of the present invention is:

[0007] The first aspect of the present invention proposes an online detection method for cognitive load overload based on EEG feature selection, comprising:

[0008] Step 1: Build a cognitive overload state task model, design a cognitive overload task based on the cognitive overload state task model, and collect the EEG signals of participants during the cognitive overload task to facilitate cognitive overload detection;

[0009] Step 2: Calculate the intrinsic mode function of the EEG signal according to the variational mode decomposition method, and process the decomposed EEG signal according to the short-time Fourier transform, extract the spectrum, power spectrum density and differential entropy features of each channel to construct a feature candidate set, so as to facilitate cognitive overload detection based on the feature candidate set;

[0010] Step 3: Screen the features in the feature candidate set according to the fusion feature selection method of mutual information and neighborhood component analysis to obtain the screened features and avoid the interference of useless information;

[0011] Step 4: Input the filtered features into the machine learning algorithm model to complete cognitive load overload detection.

[0012] Furthermore, in step 1, designing a cognitive overload task according to the cognitive overload state task model specifically includes:

[0013] Constructing original object images and object view images for cognitive overload tasks;

[0014] Participants were presented with original object images;

[0015] The object view image is presented to the participant, so that the participant can judge the original object image corresponding to the object view image based on the object view image, so as to facilitate the collection of the participant's electroencephalogram signal.

[0016] Further, the original object image includes different numbers of geometric objects pseudo-randomly arranged in space;

[0017] The object view image is a part of the original object image.

[0018] Furthermore, in the step 2, a preprocessing is performed before calculating the intrinsic mode function of the EEG signal, and the preprocessing includes:

[0019] Amplify the frequency of EEG signals to facilitate the regulation of EEG signals;

[0020] Independent component analysis is used to remove eye artifacts from EEG signals to avoid interference;

[0021] The EEG signals were re-referenced using the average reference method, and abnormal EEG signal channels were visually detected and interpolated;

[0022] The EEG signals 200ms before the cognitive overload task were used as the baseline to improve the detection accuracy.

[0023] Furthermore, the step three specifically includes:

[0024] The mutual dependence weights of the features and the labels corresponding to the features in the feature candidate set are calculated through mutual information;

[0025] The weights of the features in the feature candidate set are obtained through neighborhood component analysis;

[0026] Normalize the interdependence weights and the weights of the features in the feature candidate set to eliminate the dimension effect;

[0027] Add the normalized interdependence weights and the weights of the features in the feature candidate set to obtain the weighted feature weights;

[0028] The features in the feature candidate set are sorted according to the weighted feature weights, and the top ten percent of the features in the sorting are selected as the screened features to improve the accuracy of detection.

[0029] Furthermore, the interdependence weight is expressed by the following formula:

[0030] I(x i ,y)=∑∑p(x i ,y)*log(p(x i ,y) / (p(x i )*p(y)))

[0031] Among them, x i is the i-th feature in the feature candidate set, and y is x i The corresponding label, p(x i ,y) is x i The joint probability distribution of and y, p(x i ) is x i The boundary probability distribution of p(y) is the marginal probability distribution of y, I(x i ,y) are the interdependence weights.

[0032] The second aspect of the present invention proposes an online detection system for cognitive load overload based on EEG feature selection, comprising:

[0033] The acquisition module is used to build a cognitive overload state task model, design cognitive overload tasks according to the cognitive overload state task model, and collect EEG signals of participants during the cognitive overload task to facilitate cognitive overload detection;

[0034] An extraction module is used to calculate the intrinsic mode function of the EEG signal according to the variational mode decomposition method, and process the decomposed EEG signal according to the short-time Fourier transform, extract the spectrum, power spectrum density and differential entropy features of each channel to construct a feature candidate set, so as to facilitate cognitive overload detection based on the feature candidate set;

[0035] The dimension reduction module is used to screen the features in the feature candidate set according to the fusion feature selection method of mutual information and neighborhood component analysis to obtain the screened features and avoid the interference of useless information;

[0036] The detection module is used to input the filtered features into the machine learning algorithm model to complete cognitive load overload detection.

[0037] The third aspect of the present invention proposes a terminal device including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements an online detection method for cognitive load overload based on EEG feature selection as described in the first aspect above.

[0038] The fourth aspect of the present invention proposes a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the storage medium is located is controlled to execute an online detection method for cognitive load overload based on EEG feature selection as described in the first aspect above.

[0039] Beneficial effects of the present invention:

[0040] (1) The present invention proposes an online detection method for cognitive load overload based on EEG feature selection. For the first time, the variational mode decomposition method is used for the decomposition of EEG signals, and then a feature candidate set is constructed. The features in the feature candidate set are screened using a fusion feature selection method of mutual information and neighborhood component analysis. The screened features are input into a machine learning algorithm and a classifier to complete cognitive load overload detection. This method facilitates the effective detection of cognitive load overload based on EEG signals and improves the accuracy of detection.

[0041] (2) The present invention constructs the original object image and the object view image, and enables the participants to match the original object image and the object view image to complete the cognitive overload task. This cognitive overload task can quickly put the participants into a cognitive overload state, making it easier to collect the participants' EEG signals for subsequent testing.

[0042] (3) The present invention completes feature selection through a fusion feature selection method of mutual information and neighborhood component analysis. The correlation between features is calculated by mutual information to select relevant features, and feature weights are obtained through neighborhood component analysis. The combination of the two can effectively improve the accuracy of cognitive load overload detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 The present invention provides a method for online detection of cognitive load overload based on EEG feature selection, which is a flow chart of one embodiment of the present invention.

[0044] Figure 2 The present invention provides a flowchart of a method for online detection of cognitive load overload based on EEG feature selection according to an embodiment of the present invention.

[0045] Figure 3 A schematic diagram of an original object image and an object view image provided by an embodiment of the present invention.

[0046] Figure 4 A schematic diagram of an experimental paradigm of a cognitive overload task provided in an embodiment of the present invention.

[0047] Figure 5 A schematic diagram of a fusion feature selection method for information and neighborhood component analysis provided in an embodiment of the present invention.

[0048] Figure 6 An online detection schematic diagram of an online detection method for cognitive load overload based on EEG feature selection provided by an embodiment of the present invention.

[0049] Figure 7 A schematic diagram of participant behavior data analysis of an online cognitive load overload detection method based on EEG feature selection provided in an embodiment of the present invention.

[0050] Figure 8 A schematic diagram of a subband signal of an EEG signal obtained by a variational mode decomposition (VMD) method provided in an embodiment of the present invention.

[0051] Fig. 9 A schematic diagram of the classification accuracy of different feature selection methods based on the 5-fold cross-validation method provided in an embodiment of the present invention.

[0052] Fig.10A schematic diagram of the classification accuracy of the top 10% ranked features of different feature selection methods based on a 5-fold cross validation method provided in an embodiment of the present invention.

[0053] Fig.11 A schematic diagram of cross-subject classification accuracy of different feature selection method combinations based on the leave-one-out method provided in an embodiment of the present invention.

[0054] Fig.12 A schematic diagram of the cross-subject classification accuracy of the top 10% features in different feature selection methods based on the leave-one-out method provided in an embodiment of the present invention.

[0055] Fig.13 A schematic diagram of online evaluation results of cognitive overload status of all subjects provided in an embodiment of the present invention.

[0056] Fig.14 An architecture diagram of an online cognitive load overload detection system based on EEG feature selection provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution in the embodiment of the present invention will be clearly described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] Example 1

[0059] like Figure 1 As shown, an online detection method for cognitive load overload based on EEG feature selection includes the following steps:

[0060] S101: Build a cognitive overload state task model, design cognitive overload tasks based on the cognitive overload state task model, and collect EEG signals of participants during the cognitive overload task.

[0061] S102: calculating the intrinsic mode function of the EEG signal according to the variational mode decomposition method, processing the decomposed EEG signal according to the short-time Fourier transform, extracting the spectrum, power spectrum density and differential entropy features of each channel to construct a feature candidate set;

[0062] Specifically, the variational mode decomposition (VMD) method was applied to the EEG data of each channel to obtain the corresponding decomposition signal, and the mean absolute percentage error was used to determine the number of decomposition layers. The decomposition level was cycled from 2 to 10, and the MAPE value calculated after each decomposition was recorded. When the value of MAPE was less than the critical value of 1%, the current decomposition level was used as the final number. The final number of decomposition layers obtained by this method was 4.

[0063] The decomposed EEG signal is converted into a frequency domain signal using short-time Fourier transform (STFT). Short-time Fourier transform can effectively process non-stationary EEG signals through window functions. Then, a feature candidate set (including spectrum, power spectrum density and differential entropy features) of all channels in different frequency bands is constructed.

[0064] S103: Screening features in the feature candidate set according to a fusion feature selection method of mutual information and neighborhood component analysis to obtain screened features;

[0065] S104: Input the filtered features into the machine learning algorithm model to complete cognitive load overload detection.

[0066] The present invention designs a cognitive overload task to induce individual cognitive overload. Due to the nonlinearity and non-stationarity of EEG signals, the variational mode decomposition method (VMD) is used for the first time as an adaptive time-frequency analysis method to decompose EEG signals. Then the spectrum, power spectral density (PSD) and differential entropy (DE) of each intrinsic mode function (IMF) are calculated to construct a feature candidate set. Next, a fusion feature selection method of mutual information (MI) and neighborhood component analysis (NCA) is constructed to select effective cognitive overload features, and the screened features are used to perform cognitive load overload detection. The present invention can take advantage of the advantages of different feature selection methods to realize online detection of cognitive overload status and improve the accuracy of detection.

[0067] Example 2

[0068] Based on the above embodiments, Figure 2 As shown, the embodiment of the present invention provides a specific process of an online detection method for cognitive load overload based on EEG feature selection, which specifically includes:

[0069] S201: Build a cognitive overload state task model, design a cognitive overload task based on the cognitive overload state task model, and collect EEG signals of participants during the process of completing the cognitive overload task.

[0070] Specifically, a spatial configuration task with four geometric objects of varying numbers was designed to induce a state of cognitive overload in participants. This task is intended to assess an individual’s ability to generate and use knowledge of the environment configuration. Figure 3As shown, a cognitive overload state task model was established on a computer to design cognitive overload tasks. Each task consisted of two types of images (original object images and images with different object views), and the original stimulus materials were generated by the built-in 3D drawing software (3D viewer) in Windows 10. The original object images consisted of different numbers of geometric objects and were pseudo-randomly arranged in a space-like environment. The object view images were obtained based on the object information presented in the views of different objects in the original environment. Figure 4 As shown, each participant needs to complete a practice experiment and four formal experimental blocks. The practice experiment is to familiarize the participants with the experimental process of the entire task. In order to avoid the influence of the task order, each participant was asked to complete the spatial configuration task in a pseudo-random order of six objects, four objects, seven objects, and five objects, with a rest time of more than 15 minutes between each task. In the experiment, the original object images arranged in a pseudo-random order were first presented to the participants to learn the spatial position and layout of the objects. Then, the participants were asked to judge which original object image the current object view image came from based on the object information presented in the object view image. At the same time, the participants needed to select the original object image through the button provided at the bottom of the object view image. The electroencephalogram (EEG) signal of each participant was collected throughout the experiment, and after each block, each participant was asked to complete the NASA-TLX test scale to obtain the subjective cognitive load score.

[0071] S202: Calculate the intrinsic mode function of the EEG signal according to the variational mode decomposition method, process the decomposed EEG signal according to the short-time Fourier transform, extract the spectrum, power spectrum density and differential entropy features of each channel to construct a feature candidate set.

[0072] Specifically, in order to ensure the consistency of sample duration in the dataset, the EEG signal 3 seconds before the judgment button is selected as the sample data, and the number of objects in the task is used as the sample label. All original sample data are 40×4×50 (number of participants×number of blocks×number of object view images in a single block).

[0073] All data preprocessing was performed using MATLAB. The EEG signals were filtered from 0.1 Hz to 60 Hz, and independent component analysis (ICA) was used to remove eye artifacts from the signals. The signals were then re-referenced using the average reference method, and abnormal EEG channels were visually detected and interpolated. Finally, the EEG signals 200 ms before stimulus presentation were used as the baseline to reduce baseline differences caused by data instability.

[0074] The non-stationary characteristics of EEG signals determine that the adaptive decomposition method may have more advantages than the commonly used Fourier-based method. The present invention first uses the variational mode decomposition (VMD) method to calculate the intrinsic mode function of the preprocessed EEG signal, and then extracts the spectrum, power spectral density (PSD) and differential entropy (DE) features of each channel to construct a feature candidate set.

[0075] S203: Screening features in the feature candidate set according to a fusion feature selection method of mutual information and neighborhood component analysis to obtain screened features.

[0076] Specifically, due to the randomness and weakness of EEG signals, the extracted EEG features often contain a lot of redundant information, and some redundant or even irrelevant features often reduce the performance of EEG classification. Therefore, in the field of EEG, feature selection methods are usually used to find effective subsets from high-dimensional EEG features, perform feature extraction, and improve the representativeness of the selected features. The purpose of feature extraction is to filter out significant features, reduce feature dimensionality, increase model training speed, and improve object classification performance. Therefore, in EEG classification, feature selection is a key part of selecting the best feature set from all available features.

[0077] like Figure 5 As shown, in order to improve the effectiveness of the selected EEG features, the present invention proposes a fusion feature selection method based on mutual information (MI) and neighborhood component analysis (NCA) to reduce the feature dimension. The main idea is to calculate the degree of interdependence between the features and the labels corresponding to the features through MI as the interdependence weight, and the interdependence weight is expressed by the following formula:

[0078] I(x i ,y)=∑∑p(x i ,y)*log(p(x i ,y) / (p(x i )*p(y)))

[0079] Among them, x i is the i-th feature in the feature candidate set, and y is x i The corresponding label, p(x i ,y) is x i The joint probability distribution of and y, p(x i ) is x i The boundary probability distribution of p(y) is the marginal probability distribution of y, I(x i ,y) are the interdependence weights.

[0080] At the same time, the distribution probability obtained by NCA is used as the weight of the feature in the feature candidate set to consider the local structure of the feature to avoid selecting redundant features. NCA is a distance-based method derived from the k-nearest neighbor algorithm (k-NN algorithm). In the calculation process of NCA, it first selects a feature from the feature candidate set as the data point s i Select another feature reference point s j , the distance between two points is measured by Mahalanobis distance, which is expressed by the following formula:

[0081]

[0082] Where N is the length of the feature vector, w n For i The corresponding feature weight, s i is the i-th feature in the feature candidate set, s j is the reference feature in the feature candidate set, d w (s i ,s j ) is the distance between two features in the feature candidate set, s i,n For i The length of the eigenvector, s j,n For j The length of the feature vector.

[0083] Then calculate the feature s j Become a reference feature i The probability of getting the neighbor probability is expressed as follows: j Become a reference feature i Probability:

[0084]

[0085] Among them, p ij For the characteristics j Become a reference feature i The probability of k is s j The kth nearest neighbor feature of .

[0086] Finally, an optimization function is constructed, and the weights of the features in the feature candidate set are obtained according to the optimization function. The optimization function is expressed by the following formula:

[0087]

[0088] Among them, f(w n ) is the weight of the feature in the feature candidate set, C i is the probability parameter, C i ={j|c i =c j}.

[0089] Then, the interdependence weights and the weights of the features in the feature candidate set are normalized using the minimum-maximum normalization method to balance the scale influence of the feature weights calculated by the two feature selection methods, mutual information and neighborhood component analysis. Then, the weighted feature weights are obtained by directly adding the interdependence weights and the weights of the features in the feature candidate set. Finally, all features are sorted in turn according to the size of the weighted feature weights to obtain the sorted features, and the top ten percent of the sorted features are selected as the screened features.

[0090] S204: Input the filtered features into the machine learning algorithm model to complete cognitive load overload detection.

[0091] Specifically, after obtaining the filtered features, the filtered features are further passed to traditional machine learning methods such as RF, SVM, DAC and Bayes, and the classification performance is evaluated to find a suitable classifier to complete the detection of cognitive load overload.

[0092] Example 3

[0093] On the basis of the above-mentioned embodiment, the embodiment of the present invention provides an effect verification process of an online detection method for cognitive load overload based on EEG feature selection, which specifically includes:

[0094] The present invention uses two validation methods, 5-fold cross validation and leave-one-out validation, to validate the results within and between participants. The accuracy and robustness of the proposed method are verified on different data sets.

[0095] like Figure 6 As shown in the figure, the proposed method was tested online in the spatial configuration task. Each participant first became familiar with the task operation in the practice phase and then completed four formal experimental blocks. In the EEG data acquisition phase, EEG signals were recorded from a multi-channel amplifier (g.HIamp) EEG acquisition system at a sampling rate of 250Hz, and data buffering was used to save the EEG signals of each block. The buffered data was processed in the rest phase. The online processing part of the EEG signal was implemented based on Matlab, which mainly includes sample processing, sample segmentation and sample classification.

[0096] For preprocessing, a bandpass filter from 0.1 Hz to 60 Hz was used, and the signal was re-referenced using the average reference method, and any bad channels were linearly interpolated based on the three nearest neighbors. The EEG signals during the practice phase were then used as a baseline to reduce baseline differences caused by data instability. In the sample segmentation phase, each block-based EEG data was segmented according to a window size of 6 s to obtain N segments for each block. An online detection method for cognitive load overload based on EEG feature selection was used to assess the state of cognitive overload.

[0097] (1) Analysis of cognitive overload induction

[0098] The behavioral data results of all participants at the four task difficulty levels are as follows Figure 7 As the difficulty of the task increased, the average reaction time of the participants gradually increased, and the scores on the scale became higher and higher, which showed that the cognitive workload caused by the task gradually increased with the increase in the number of objects.

[0099] In order to obtain more accurate EEG signals before and after individual cognitive overload, the present invention uses whether the average accuracy of individual single tasks is lower than the blind guessing probability as the criterion for judging cognitive workload overload. In this case, it is found that Sub4, Sub6, Sub9, Sub10, Sub12, Sub31 and Sub34 are excluded because they do not meet the condition of being lower than the blind guessing probability, even if the accuracy drops sharply, and Sub38 is excluded because the accuracy is lower than the blind guessing probability from the beginning of the task. Sub18 and Sub39 show cognitive overload in the task involving 5 objects. Sub1, Sub2, Sub8, Sub14, Sub15, Sub19, Sub25, Sub26, Sub30, Sub33, Sub35, Sub36 and Sub37 show cognitive overload in the task involving 6 objects. Sub3, Sub5, Sub7, Sub11, Sub13, Sub16, Sub17, Sub20, Sub21, Sub22, Sub23, Sub24, Sub27, Sub28, Sub29, Sub32, and Sub40 showed cognitive overload in the task containing 7 subjects. Therefore, by summarizing the EEG signals of participants who experienced cognitive workload overload, an EEG dataset including two states was constructed: cognitive overload and non-overload. When an individual first showed a cognitive overload state, the EEG signal of the current task block could be used as a sample of the cognitive overload state, while the EEG signal of the previous task block could be regarded as a sample of the cognitive non-overload state. The total sample data was 32 × 2 × 50 (number of participants × cognitive load state × trial).

[0100] (2) Subband analysis:

[0101] In order to ensure the consistency of sample duration, the present invention selects the EEG signal 3 seconds before the button is pressed in the local data set as the sample data for each trial. The VMD method is applied to the EEG data of each channel to obtain the corresponding decomposition signal, and the mean absolute percentage error (MAPE) is used to determine the number of decomposition layers. The decomposition level is cycled from 2 to 10, and the MAPE value calculated after each decomposition is recorded. When the value of MAPE is less than the critical value of 1%, the current decomposition level is used as the final number. Figure 8 As shown in the figure, the decomposition of the EEG data extracted from the FCz channel of Sub3 in the local dataset, where subband SB1 represents the subband signal with the lowest frequency and SB4 represents the subband signal with the highest frequency. After extracting the spectrum, PSD and DE features of all subbands on different channels, a total of 62×4×3 features are obtained (channel×subband×feature type).

[0102] (3) Within-subject analysis

[0103] In order to verify the effectiveness of the proposed method, the features are sorted according to the weighted feature weights based on the fusion feature selection method of mutual information and neighborhood component analysis to obtain the screened features, and the performance of different classifiers is verified. The incremental evaluation method is used to calculate the incremental evaluation of the sorted feature subsets in different feature selection methods. Fig. 9 The figure shows the classification accuracy of different feature selection methods tested using a 5-fold cross validation method in a local dataset. The results show that when using feature selection methods to assess cognitive overload, stable classification accuracy can be obtained with a small number of features, and the accuracy of the support vector machine (SVM) classifier is much higher than that of other classifiers. This verifies that this method can effectively reduce the feature dimension in cognitive overload assessment.

[0104] In order to more intuitively analyze the performance of the proposed method in different classifiers, the classification performance of the top 10% ranked features from different feature selection methods is compared with the classification performance when all features are used. Fig.10 As shown, the present invention achieves the highest classification accuracy of 88.25% on the SVM classifier, which is 4.64% higher than the MI feature selection method alone. Surprisingly, the classification accuracy of the fusion feature selection method of mutual information and neighborhood component analysis is 4.02% higher than that of using all features.

[0105] (4) Analysis of cross-subject results

[0106] In order to evaluate the robustness of the proposed method in cross-subject cognitive overload assessment, the cross-subject performance of different feature selection methods was verified using the leave-one-out method on different classifiers, and the incremental evaluation method was used to calculate the incremental evaluation of different feature selection methods on the sorted feature subsets. Fig.11 The performance of different feature selection methods for cognitive overload assessment across subjects in a local dataset is shown. It is found that the fusion feature selection method of MI and NCA on SVM shows significantly superior classification accuracy compared to other methods, thus confirming the robustness of the proposed method in cognitive overload assessment in this study.

[0107] In order to intuitively analyze the robustness of different methods, the cross-subject classification accuracy of the top 10% features selected by different feature selection methods was compared with the classification accuracy using all features. Fig.12 As shown in Figure 2, the fusion feature selection method of MI and NCA achieved the highest cross-subject classification accuracy of 83.44%, which is 1.55% higher than using MI alone and 1.52 higher than using all features. This indicates that not all feature attributes are conducive to cross-subject cognitive overload assessment. At the same time, the cross-subject accuracy of the SVM classifier is higher than that of other classifiers. In order to verify the performance of feature extraction in the network,

[0108] (5) Ablation experiment

[0109] Given the excellent performance of the support vector machine on the local dataset, the cross-subject classification performance of various feature selection methods on the support vector machine classifier was verified on different datasets. In addition, four indicators were used to evaluate the performance of various combined feature selection methods: mean accuracy, standard deviation, sensitivity (TPR), and specificity (TNR). Sensitivity and specificity are provided for each cognitive load level in the classification results of the COG-BCI dataset and the local dataset. The cross-subject classification results using various feature selection methods or all features on different datasets are shown in Table 1:

[0110] Table 1: Classification results of subjects

[0111]

[0112] The results show that the fusion of MI and NCA achieves the highest cross-subject classification accuracy of 83.44% and 78.24% in local and public datasets, respectively. The results show that the fusion feature selection method of MI and NCA and the SVM classifier can be used to build an effective EEG-based cognitive overload assessment system.

[0113] (6) Method comparison

[0114] The results of the comparison between the present invention and the methods used in recent literature are shown in Table 2.

[0115] Table 2: Comparison results between the present invention and methods used in recent literature

[0116]

[0117]

[0118] It is clear that the proposed method significantly outperforms other methods in classification performance on both datasets. Compared with other methods, the proposed method improves the cross-subject cognitive load assessment performance by 15.81% and 9.63% on the local dataset and open dataset, respectively. The results show that the proposed method can significantly improve the robustness and effectiveness of cognitive overload assessment. Since the performance achieved by the proposed method is promising, it can be considered as a reliable algorithm for online assessment of cognitive overload.

[0119] (7) Online Results

[0120] Based on the analysis of the above results, it can be found that the online detection method for cognitive overload based on EEG feature selection proposed in the present invention has the best performance. Therefore, using this method to conduct online assessment of cognitive overload, a total of 8 participants underwent online assessment test tasks for cognitive overload. The online cognitive overload assessment results of all participants in the spatial perception task are shown in Figure 2. Fig.13 The results show that the average accuracy of the method in online cognitive overload assessment is 79.90±2.53%, and the highest accuracy of Sub5 is 84.01%.

[0121] Example 4

[0122] Corresponding to the above method, such as Fig.14 As shown, the present invention proposes an online detection system for cognitive load overload based on EEG feature selection, comprising:

[0123] The acquisition module is used to build a cognitive overload state task model, design cognitive overload tasks based on the cognitive overload state task model, and collect the EEG signals of participants during the completion of the cognitive overload task.

[0124] The extraction module is used to calculate the intrinsic mode function of the EEG signal according to the variational mode decomposition method, and process the decomposed EEG signal according to the short-time Fourier transform, extract the spectrum, power spectral density and differential entropy features of each channel to construct a feature candidate set.

[0125] The dimension reduction module is used to screen features in the feature candidate set according to the fusion feature selection method of mutual information and neighborhood component analysis to obtain the screened features.

[0126] The detection module is used to input the filtered features into the machine learning algorithm model to complete cognitive load overload detection.

[0127] It should be noted that the online detection system for cognitive load overload based on EEG feature selection provided in an embodiment of the present invention is for realizing the above-mentioned online detection method for cognitive load overload based on EEG feature selection. Its specific functions can be referred to the above-mentioned method embodiments, which will not be repeated here.

[0128] Example 5

[0129] Based on the above embodiments, an embodiment of the present invention further provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, an online detection method for cognitive load overload based on EEG feature selection in the above embodiments is implemented.

[0130] Preferably, the present invention also provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the storage medium is located is controlled to execute an online detection method for cognitive load overload based on EEG feature selection in the above-mentioned embodiment.

[0131] In summary, the present invention proposes an online detection method for cognitive load overload based on EEG feature selection, which firstly uses the variational mode decomposition method for the decomposition of EEG signals, then constructs a feature candidate set, uses the fusion feature selection method of mutual information and neighborhood component analysis to screen the features in the feature candidate set, and inputs the screened features into the machine learning algorithm and classifier to complete the cognitive load overload detection, which is convenient for effectively realizing the detection of cognitive load overload according to EEG signals and improves the accuracy of detection. The present invention completes the cognitive overload task by constructing the original object image and the object view image, so that the participants match the original object image and the object view image. This cognitive overload task can quickly put the participants in a cognitive load overload state, and facilitates the collection of the participants' EEG signals for subsequent detection. The present invention completes the feature selection by the fusion feature selection method of mutual information and neighborhood component analysis, selects the relevant features by calculating the correlation between the features through mutual information, and obtains the feature weights through neighborhood component analysis. The combination of the two can effectively improve the accuracy of cognitive load overload detection.

[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An online detection method for cognitive load overload based on EEG feature selection, characterized in that: include: Step 1: Build a cognitive overload state task model, design a cognitive overload task based on the cognitive overload state task model, and collect the EEG signals of participants during the cognitive overload task. Step 2: Calculate the intrinsic mode function of the EEG signal according to the variational mode decomposition method, and process the decomposed EEG signal according to the short-time Fourier transform, extract the spectrum, power spectrum density and differential entropy features of each channel to construct a feature candidate set; Step 3: Screen the features in the feature candidate set according to the fusion feature selection method of mutual information and neighborhood component analysis to obtain the screened features; Step 4: Input the filtered features into the machine learning algorithm model to complete cognitive load overload detection.

2. The online detection method for cognitive load overload based on EEG feature selection according to claim 1 is characterized in that: In the step 1, a cognitive overload task is designed according to the cognitive overload state task model, specifically including: constructing original object images and object view images; Participants were presented with original object images; The object view image is presented to the participant, so that the participant can judge the original object image corresponding to the object view image based on the object view image.

3. The online detection method for cognitive load overload based on EEG feature selection according to claim 2 is characterized in that: The original object image includes different numbers of geometric objects pseudo-randomly arranged in space; The object view images are views of different geometric objects in the original object image.

4. The online detection method for cognitive load overload based on EEG feature selection according to claim 1 is characterized in that: Before calculating the intrinsic mode function of the EEG signal in step 2, preprocessing is required, and the preprocessing includes: Amplify the frequency of EEG signals; Independent component analysis was used to remove eye artifacts from EEG signals; The EEG signals were re-referenced using the average reference method, and abnormal EEG signal channels were visually detected and interpolated; The EEG signals 200ms before the cognitive overload task were used as the baseline.

5. The online detection method for cognitive load overload based on EEG feature selection according to claim 1 is characterized in that: The step three specifically includes: The mutual dependence weights of the features and the labels corresponding to the features in the feature candidate set are calculated through mutual information; The weights of the features in the feature candidate set are obtained through neighborhood component analysis; Normalize the interdependence weights and the weights of the features in the feature candidate set respectively; Add the normalized interdependence weights and the weights of the features in the feature candidate set to obtain the weighted feature weights; The features in the feature candidate set are sorted according to the weighted feature weights, and the top ten percent of the features in the sorting are selected as the screened features.

6. The online detection method for cognitive load overload based on EEG feature selection according to claim 5 is characterized in that: The interdependence weight is expressed as follows: I(x i ,y)=∑∑p(x i ,y)*log(p(x i ,y) / (p(x i )*p(y))) Among them, x i is the i-th feature in the feature candidate set, and y is x i The corresponding label, p(x i ,y) is x i The joint probability distribution of and y, p(x i ) is x i The boundary probability distribution of p(y) is the marginal probability distribution of y, I(x i ,y) are the interdependence weights.

7. An online detection system for cognitive load overload based on EEG feature selection, characterized in that: include: The acquisition module is used to build a cognitive overload state task model, design cognitive overload tasks according to the cognitive overload state task model, and collect EEG signals of participants during the process of completing cognitive overload tasks; An extraction module is used to calculate the intrinsic mode function of the EEG signal according to the variational mode decomposition method, and process the decomposed EEG signal according to the short-time Fourier transform, extract the spectrum, power spectrum density and differential entropy features of each channel to construct a feature candidate set; The dimension reduction module is used to screen the features in the feature candidate set according to the fusion feature selection method of mutual information and neighborhood component analysis to obtain the screened features; The detection module is used to input the filtered features into the machine learning algorithm model to complete cognitive load overload detection.

8. A terminal device, characterized in that: It comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, it implements an online detection method for cognitive load overload based on EEG feature selection as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The storage medium includes a stored computer program, wherein when the computer program is running, the device where the storage medium is located is controlled to execute the online detection method for cognitive load overload based on EEG feature selection as described in any one of claims 1 to 6.

Images