CN-121980521-A - Cross-test cognitive load identification method based on depth-to-antigen domain self-adaption

Abstract

The invention discloses a depth-based adaptive cross-test cognitive load recognition method based on a contrast domain, which comprises the steps of collecting electroencephalogram (EEG) signals under different cognitive load levels, preprocessing and slicing the collected EEG signals, extracting features of preprocessed EEG time sequence data, fusing the extracted features and converting the features into new features of fused space-time information, and sending the fused features under different cognitive load levels to a trained depth-based adaptive network for training to finish recognition of different cognitive load levels. By analyzing EEG signals, the characteristics related to time, space and frequency are extracted, and the characteristic information of the three characteristics is fused. The depth domain is used for resisting the self-adaptive network, and the MMD characteristic distribution loss function is assisted, so that the network model is forced to learn characteristic information irrelevant to a tested object, and the recognition performance of the cognitive load recognition level across the tested object is improved.

Inventors

• FU XIANGYU
• XU CHAO
• LI XING

Assignees

• 天津大学

Dates

Publication Date

20260505

Application Date

20260320

Claims (4)

1. 1. The invention provides a depth-domain-based self-adaptive cross-test cognitive load identification method, which is characterized by comprising the following steps of: step S100, acquiring EEG signals under different cognitive load levels; step 200, preprocessing and slicing the acquired EEG signals to obtain preprocessed EEG time sequence data; Step 300, extracting features of the preprocessed EEG time sequence data, and calculating features such as power spectral density on EEG frequency bands closely related to cognitive load level; step 400, performing feature fusion and conversion on the extracted features, and converting the features into new features fusing space-time information by combining time and electroencephalogram cap electrode space arrangement features; Step S500, the fusion characteristics of different tested under different cognitive load levels are sent into a training depth-to-antigen domain self-adaptive network for training, and recognition of the different cognitive load levels is completed; The step S200 specifically includes: Step S210 using 0.1 The 40Hz band-pass filter carries out filtering denoising treatment on the original brain electrical signal to remove high-frequency noise, power frequency interference and other irrelevant noise; Step S220, dividing the electroencephalogram signals after filtering treatment according to a preset time window, preferably dividing the electroencephalogram signals into time sequences with the duration of 2S; step S220, decomposing the residual electroencephalogram signal fragments by adopting an independent component analysis method, and identifying and removing the electrooculogram and other artifacts to obtain preprocessed electroencephalogram signals; The step S300 specifically includes: Step S310, converting the time domain signal into a frequency domain signal by adopting Fourier transformation, wherein the time domain signal is specifically expressed as: Wherein, the Indicating the angular frequency of the brain electrical signal A lower frequency domain representation of the frequency domain, Is an imaginary unit; Step S320, calculating the power spectrum density under the corresponding frequency according to the frequency domain signal, wherein the calculation formula of the power spectrum density is as follows: Wherein, the , And Representing the sampling period, the number of samples in a time period and the sampling frequency respectively, Representing the power spectral density of the corresponding frequency band; s330, integrating or averaging the power spectral density in different frequency bands to obtain power spectral density characteristic values corresponding to each preset frequency band, wherein the frequency bands comprise delta frequency bands, theta frequency bands, alpha frequency bands and beta wave frequency bands, so as to form a power spectral density characteristic vector; In the step S400, the power spectrum density characteristic and the electrode space coordinate information are fused to form a multi-channel two-dimensional characteristic image as a fusion characteristic, which specifically includes: Step S410, according to the type of the EEG cap used in EEG data acquisition, acquiring the three-dimensional space distribution coordinates of the upper electrode of the EEG cap Converting the three-dimensional coordinates into spherical coordinates by taking the reference point as a projection center ; Step S420, mapping the spherical coordinates of the electrodes to a two-dimensional plane space by using equidistant azimuth projection, wherein the projection relation satisfies the formula: Wherein, the For the plane coordinates after projection, Is the radius of the sphere, and the radius of the sphere, As a scale factor, the number of the elements is, For the coordinates of the projection points, Is the projection center; step S430, according to the projection coordinates of each electrode Generating a single-channel two-dimensional image on each frequency band, and obtaining a pixel value through interpolation calculation of electrode coordinates and power spectrum density characteristics of the frequency band; Step S440, performing projection and interpolation on each time segment, and calculating to obtain a four-channel two-dimensional image sequence as training data; the step S500 specifically includes: Step S510, organizing the shape format of the training data set data array; Step S520, randomly selecting one tested data as a test set and the rest data as a training set according to different tested numbers, randomly selecting one tested data as target domain data again in the training set and the rest training set data as source domain data; Step S530, selecting proper data batch size, mixing the data of the source domain and the target domain in one batch, supplementing domain label information, and sending batch data to a frame attention depth feature extractor part of a model to obtain the depth feature of each sample; Step S540, based on the depth characteristics of the sample, classifying by using a domain label discriminator and a task label discriminator of the model respectively, calculating classification loss according to classification results, classification label information and domain label information, and iteratively updating model parameters through a back propagation algorithm to complete model training; and step S550, a frame attention depth feature extractor and a task tag discriminator of the model are reserved, test set data are sent into the model, a classification result is calculated, and recognition of the recognition load level is completed.
2. 2. The depth-to-domain adaptive network according to claim 1, wherein the model comprises three parts, namely a frame attention depth feature extractor, a domain label discriminator and a task label discriminator, wherein the frame attention depth feature extractor firstly calculates depth features for each time slice of input data, then calculates attention weights for each time slice by using a self-attention mechanism, then performs weighted fusion to obtain fusion features, and further performs attention weight calculation to ensure that global information shared by different time slices is fully considered for the calculation of the weights, firstly performs weighted average on the feature vectors of each time slice by using the attention weights, calculates a feature vector representing global information, and then splices the global feature vector onto the feature vector of each time slice, so that the global information is fused, and then calculates weights again for the feature vectors spliced by each time slice, and then calculates weighted average features as final depth features.
3. 3. The depth-directed domain adaptive network joint loss function of claim 1, wherein the joint loss function formula is: Wherein, the 、 And Network parameters respectively representing a feature extractor, a task classifier and a domain arbiter; for the domain to counter the loss weight coefficient, The weighting coefficients are aligned for the feature distribution, The loss function is classified for the task, For the domain discrimination loss function, The loss function is distributed for MMD features.
4. 4. The frame attention depth feature extractor according to claim 2, wherein the model comprises three parts, namely a frame attention depth feature extractor, a domain label discriminator and a task label discriminator, the frame attention depth feature extractor calculates a depth feature for each time slice of input data, calculates an attention weight for each time slice by using a self-attention mechanism, performs weighted fusion to obtain a fusion feature, and performs weighted fusion to ensure that global information shared by different time slices is fully considered for the calculation of the weight in the calculation process of the attention weight, firstly performs weighted average on the feature vectors of each time slice by using the attention weight to calculate a feature vector representing global information, then splices the global feature vector to the feature vector of each time slice, so that the time global information is fused, and after global information is obtained, calculates the weight again for the feature vector spliced by each time slice, and then calculates a weighted average feature as a final depth feature.

Description

Cross-test cognitive load identification method based on depth-to-antigen domain self-adaption Technical Field The invention belongs to the field of cognitive neuroscience, and particularly relates to a cognitive load identification method based on an EEG signal and depth opposite domain adaptive technology. Background Cognitive load describes the ratio of available resources to the brain to resources required for a task. Cognitive load recognition is an important research topic in the fields of human-computer interaction and brain-computer interfaces. In a real environment, excessive cognitive load levels or underruns may affect the brain activity state of an operator, reduce work efficiency, and even cause damage to the brain. Therefore, the recognition of the cognitive load is significant in reducing work errors and guaranteeing personnel safety. The cognitive load theory is mainly based on a brain cognitive architecture consisting of working memory and long-term memory. The working memory is limited in capacity and time, and is mainly used for processing new information accepted by the brain in a short term, while the capacity of the long-term memory is almost infinite. The brain may have a bottleneck in processing new information due to the limitation of working memory capacity. In more complex scenarios, the brain is required to not only store information in working memory, but also to handle complex logical relationships between information elements. The capacity of the working memory is further reduced in this case. The cognitive load can be regarded as the occupation condition of limited cognitive resources in the brain working memory under the comprehensive influence of cognitive tasks and the characteristics of individuals. The evaluation of cognitive load mainly comprises a subjective method and an objective method. In the subjective method, most of researches adopt subjective score scales to evaluate the cognitive load, the time interval of the method is too large, instantaneous load fluctuation which changes along with time cannot be inspected, and the actual and objective measurement of the tested cognitive load level cannot be ensured. In the method for objectively evaluating the cognitive load based on the physiological signals, certain physiological indexes are weakly related to the cognitive load, such as blink rate and blink duration, and other methods have insufficient time resolution for real-time classification, such as hormone level, heart rate variability and the like. In addition, the methods such as positron emission tomography, functional magnetic resonance imaging and the like are sensitive to cognitive load response and can measure real-time change, but due to the invasive characteristic, the measuring process is complex, and the deployment difficulty is high. Electroencephalogram (EEG) is an effective tool for continuously measuring cognitive load. The EEG signals have stable and repeatable correspondence with the individual cognitive load levels in the time domain, frequency domain and time-frequency domain characteristics. For example, when cognitive load levels increase, θ and β band power generally exhibit an increasing trend, while α band power is relatively decreasing. The above-mentioned change in neuroelectric activity reflects the difference of brain in terms of attention resource allocation, work memory occupation and information processing intensity. Therefore, by collecting, analyzing and modeling the EEG signal characteristics, the cognitive load level of the individual can be objectively and real-time represented, so that the effective and quantifiable association relationship between the EEG and the cognitive load of the individual can be proved. Current cognitive load recognition methods based on EEG generally face the problem of significant decline across the effects tested. The EEG signals have obvious individual differences, and different individuals have differences in scalp structures, electrode wearing positions, brain cortex morphology, nerve function organization modes and the like, so that the EEG signals acquired under the same cognitive load condition show inconsistency in frequency spectrum characteristics and space topological structures. In addition, differences in neural response strategies, attention resource allocation patterns, and learning experience among individuals can further amplify the non-uniformity between EEG features and cognitive load levels. In order to improve the recognition performance of the cross-test, the invention provides a recognition method of the cross-test cognitive load based on EEG signals and depth-to-antigen domain self-adaption. Disclosure of Invention Aiming at the technical problems, the invention provides a cross-tested cognitive load identification method based on EEG signal and depth-to-antigen domain self-adaption. The technical scheme adopted for solving the technical problems is as follows: a cross-test co