Method, device and equipment for detecting cognitive state of attention glasses in real time

Abstract

The invention relates to the technical field of brain-computer interfaces, in particular to a method, a device and equipment for detecting the cognitive state of attention glasses in real time. The method comprises the following steps: acquiring multi-channel electroencephalogram data of attention glasses, acquiring multi-dimensional features corresponding to the multi-channel electroencephalogram data, and generating a detection sequence; performing time sequence analysis on the detection sequence, extracting change characteristics to construct a cognitive sequence, and generating state description according to the cognitive sequence; performing trend analysis on the state description, performing feature extraction on an analysis result corresponding to the trend analysis, constructing a conversion rule according to the extracted features, and generating a prediction sequence according to the conversion rule; performing threshold division on the prediction sequence, marking state intervals, constructing a criterion sequence, and generating an evaluation result according to the criterion sequence for performing signal detection to obtain detection information; and performing data conversion on the detection information, constructing a display sequence to mark a state type, and generating cognitive state information. And data continuity and signal integrity in a feature extraction process are realized.

Description

Technical Field

[0001] The present application relates to the field of brain-computer interface technology, and in particular to a method, device and equipment for real-time detection of cognitive status of attention glasses. Background Art

[0002] As a portable cognitive state monitoring device, attention glasses have important application value in the fields of education, learning, and work efficiency improvement. Traditional cognitive state detection methods mainly rely on questionnaire evaluation or behavioral observation. These methods not only cannot achieve real-time and objective state evaluation, but are also easily affected by subjective factors. Although EEG signals contain rich cognitive state information, existing EEG analysis methods have obvious deficiencies in key links such as signal processing, feature extraction, and state recognition. In the signal acquisition and preprocessing stages, the data quality of portable devices is unstable, and traditional methods lack effective noise suppression and quality control mechanisms. In the feature analysis stage, existing methods often use simple time domain or frequency domain features, which cannot fully explore the multidimensional features reflecting cognitive states in the signal, and it is also difficult to capture the dynamic characteristics of state changes.

[0003] Current cognitive state detection technologies face challenges in terms of real-time performance, accuracy, and reliability. The EEG data collected by portable devices needs to be processed immediately, but existing methods have low processing efficiency in data cache management, feature extraction, and state recognition, making it difficult to meet the needs of real-time monitoring. The cognitive state itself is highly dynamic and has individual differences, and fixed detection models are difficult to adapt to such changes, resulting in insufficient accuracy of the detection results. At the same time, there are complex interactions between the multidimensional features of cognitive states. Existing methods lack in-depth analysis of such correlations, and have not established systematic state assessment standards, making it difficult to ensure the reliability of the detection results.

[0004] Therefore, a method is urgently needed to solve at least one of the above problems. Summary of the invention

[0005] The embodiments of the present application provide a method, device and equipment for real-time detection of cognitive states of attention glasses, and the method aims to solve the challenges of real-time, accuracy and reliability faced by current cognitive state detection technologies. The EEG data collected by portable devices need to be processed immediately, but the processing efficiency of existing methods in data cache management, feature extraction and state recognition is low, and it is difficult to meet the needs of real-time monitoring. The cognitive state itself is highly dynamic and individual, and fixed detection models are difficult to adapt to such changes, resulting in insufficient accuracy of the detection results. At the same time, there are complex interactions between the multi-dimensional characteristics of cognitive states. Existing methods lack in-depth analysis of such correlations, and have not established systematic state evaluation standards, making it difficult to ensure the reliability of detection results.

[0006] In a first aspect, an embodiment of the present application provides a method for real-time detection of cognitive status of attention glasses, comprising:

[0007] Acquire multi-channel EEG data of the attention glasses, and acquire multi-dimensional features corresponding to the multi-channel EEG data; perform time domain transformation according to the multi-dimensional features, acquire signal components, and acquire state representations corresponding to the signal components; perform hierarchical processing on the state representations to determine feature levels; label cognitive attributes based on the feature levels; construct a state map according to the cognitive attributes, and generate a detection sequence according to the state map;

[0008] Performing time series analysis on the detection sequence to extract change features, marking state nodes based on the change features, using the state nodes to construct a cognitive sequence, and generating a state description based on the cognitive sequence; performing trend analysis on the state description, performing feature extraction on the analysis result corresponding to the trend analysis, so as to construct a conversion rule based on the extracted features, and generating a prediction sequence based on the conversion rule;

[0009] Performing threshold division on the prediction sequence, marking state intervals, performing feature aggregation according to the state intervals, constructing a criterion sequence using aggregation results corresponding to the feature aggregation, and generating an evaluation result according to the criterion sequence; performing signal detection on the evaluation result to obtain detection information;

[0010] The detection information is converted into data, a display sequence is constructed, a state organization is performed based on the display sequence, and a state type is marked according to the organization result of the state organization; and cognitive state information is generated according to the state type.

[0011] In a second aspect, the present application also provides a real-time detection device for cognitive status, comprising:

[0012] A data acquisition unit is used to acquire multi-channel EEG data of the attention glasses, acquire multi-dimensional features corresponding to the multi-channel EEG data; perform time domain transformation according to the multi-dimensional features, acquire signal components, and acquire state representations corresponding to the signal components; perform hierarchical processing on the state representations to determine feature levels; mark cognitive attributes based on the feature levels; construct a state map according to the cognitive attributes, and generate a detection sequence according to the state map;

[0013] A timing analysis unit, configured to perform timing analysis on the detection sequence, extract change features, mark state nodes based on the change features, construct a cognitive sequence using the state nodes, and generate a state description based on the cognitive sequence; perform trend analysis on the state description, perform feature extraction on the analysis result corresponding to the trend analysis, so as to construct a conversion rule based on the extracted features, and generate a prediction sequence based on the conversion rule;

[0014] A threshold division unit is used to perform threshold division on the prediction sequence, mark state intervals, perform feature aggregation according to the state intervals, construct a criterion sequence using the aggregation results corresponding to the feature aggregation, generate an evaluation result according to the criterion sequence; perform signal detection on the evaluation result to obtain detection information;

[0015] A data conversion unit is used to perform data conversion on the detection information, construct a display sequence, organize the state based on the display sequence, mark the state type according to the organization result of the state organization; and generate cognitive state information according to the state type.

[0016] In a third aspect, the present application further provides a computer device, comprising a processor and a memory, wherein the memory is used to store a computer program, and when the computer program is executed by the processor, the real-time detection method of the cognitive state of the attention glasses as described in the first aspect is implemented.

[0017] This method realizes real-time monitoring and classification of user cognitive state through dynamic feature extraction and state sequence modeling of multi-channel electroencephalogram (EEG) signals. EEG data is collected through the multi-channel sensors of attention glasses to extract multidimensional features (such as time domain amplitude, frequency band energy, and time-frequency joint features). The EEG data is transformed in the time domain (such as filtering, segmentation, and Fourier transform) to generate signal components (such as α wave and β wave components). State representations (such as frequency band energy distribution matrices) are generated based on signal components, and feature levels (such as high-frequency / low-frequency feature priorities) are determined through hierarchical processing (such as clustering and principal component analysis). Cognitive attributes (such as concentration and fatigue) are marked, state mappings (such as mapping rules from features to cognitive states) are constructed, and detection sequences (such as state sequences within time windows) are generated. Time series analysis (such as sliding window statistics and trend fitting) is performed on the detection sequence to extract change features (such as energy fluctuations and frequency offsets). State nodes (such as slope change points and abnormal fluctuation points) are marked, cognitive sequences (such as state transition chains) are constructed, and state descriptions (such as the transition path of "concentration→fatigue") are generated. Generate a prediction sequence (such as the probability of cognitive state at a future time point) through trend analysis (such as regression prediction), and mark the state interval (such as high concentration interval) based on the threshold. Aggregate features to generate a judgment sequence (such as classification threshold rules) and output evaluation results (such as "current concentration: 80%"). Convert the detection results into a display sequence (such as a visual chart), mark the state type (such as "efficient work" or "mild distraction") through state organization (such as clustering or rule matching), and finally generate cognitive state information (such as real-time reminders or data reports).

[0018] Through multi-dimensional feature extraction and time series analysis, the reliance of traditional methods on single features is overcome, and the accuracy of cognitive state classification is improved (e.g., the accuracy of concentration detection is increased by more than 20%). Based on dynamic modeling of state nodes and prediction sequences, robust detection under environmental interference (such as noise) is achieved, and the false positive rate is reduced (e.g., the false alarm rate in a fluctuating environment is reduced by 30%). Combined with the local computing power of the attention glasses (such as edge processing chips), real-time signal processing (delay <100ms) is completed to avoid cloud transmission bottlenecks. Through visual feedback (such as state type labeling) and personalized criteria (such as threshold adaptation), intuitive cognitive state monitoring is provided, which is suitable for multiple scenarios such as education, medical care, and office. Support multi-channel data fusion (such as the combination of EEG and eye movement data), providing the underlying technical framework for future brain-computer interface (BCI) applications.

[0019] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A schematic flow chart of a method for real-time detection of cognitive status of attention glasses according to an embodiment of the present application;

[0021] Figure 2 This is a schematic diagram of the structure of a real-time detection device for cognitive status according to an embodiment of the present application;

[0022] Figure 3 A schematic diagram of the structure of a computer device shown in an embodiment of the present application. DETAILED DESCRIPTION

[0023] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0024] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or combinations thereof.

[0025] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0026] As used in the specification and appended claims of this application, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [described condition or event] is detected" or "in response to detecting [described condition or event]", depending on the context.

[0027] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0028] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0029] The technical solution of the embodiment of the present application is introduced below.

[0030] As a portable cognitive state monitoring device, attention glasses have important application value in the fields of education, learning, and work efficiency improvement. Traditional cognitive state detection methods mainly rely on questionnaire evaluation or behavioral observation. These methods not only cannot achieve real-time and objective state evaluation, but are also easily affected by subjective factors. Although EEG signals contain rich cognitive state information, existing EEG analysis methods have obvious deficiencies in key links such as signal processing, feature extraction, and state recognition. In the signal acquisition and preprocessing stages, the data quality of portable devices is unstable, and traditional methods lack effective noise suppression and quality control mechanisms. In the feature analysis stage, existing methods often use simple time domain or frequency domain features, which cannot fully explore the multidimensional features reflecting cognitive states in the signal, and it is also difficult to capture the dynamic characteristics of state changes.

[0031] Current cognitive state detection technologies face challenges in terms of real-time performance, accuracy, and reliability. The EEG data collected by portable devices needs to be processed immediately, but existing methods have low processing efficiency in data cache management, feature extraction, and state recognition, making it difficult to meet the needs of real-time monitoring. The cognitive state itself is highly dynamic and has individual differences, and fixed detection models are difficult to adapt to such changes, resulting in insufficient accuracy of the detection results. At the same time, there are complex interactions between the multidimensional features of cognitive states. Existing methods lack in-depth analysis of such correlations, and have not established systematic state assessment standards, making it difficult to ensure the reliability of the detection results.

[0032] Please refer to Figure 1 , Figure 1 The flowchart of a method for real-time detection of cognitive state of attention glasses provided in an embodiment of the present application is shown in FIG. The method for real-time detection of cognitive state of attention glasses provided in an embodiment of the present application can be applied to computer devices, including but not limited to smart phones, laptops, tablet computers, desktop computers, physical servers, cloud servers and other devices. Figure 1 As shown, the real-time detection method of cognitive state of attention glasses in this embodiment includes steps S101 to S104, which are described in detail as follows:

[0033] Step S101, obtain multi-channel EEG data of the attention glasses, obtain multi-dimensional features corresponding to the multi-channel EEG data; perform time domain transformation according to the multi-dimensional features, obtain signal components, and obtain state representations corresponding to the signal components; perform hierarchical processing on the state representation to determine the feature level; label cognitive attributes based on the feature level; construct a state mapping according to the cognitive attributes, and generate a detection sequence according to the state mapping.

[0034] Specifically, when the attention glasses collect multi-channel EEG data and establish data cache, the dry electrode sensor built into the glasses collects signals from the frontal lobe area. The dry electrode fits tightly to the forehead skin of the human body, the sampling frequency is set to 256Hz, and the signal amplitude range is ±100μV. To ensure signal quality, the system monitors the electrode impedance in real time. When the electrode impedance exceeds 50kΩ, the channel data is automatically marked as invalid data.

[0035] In some embodiments, the generation of multidimensional features corresponding to the multi-channel EEG data includes: establishing a data cache for the multi-channel EEG data; performing window division based on the data cache, and generating a real-time sequence according to the division result of the window division; performing feature marking on the real-time sequence and extracting state indicators; constructing a feature set based on the state indicators, and generating multidimensional features according to the feature set.

[0036] The collected raw signal data is transmitted to the data processing unit in real time via the Bluetooth 5.0 protocol, and 128-bit encryption is used during the transmission process to ensure data security.

[0037] Exemplarily, the establishing of a data cache for the multi-channel EEG data includes: adopting a double buffering mechanism for cache management to calculate the kurtosis and skewness of the divided multi-channel EEG data to evaluate data quality.

[0038] After receiving the signal, the data processing unit first performs a data integrity check, and then stores the data in a circular data cache after the check passes. The data cache is managed in a first-in-first-out (FIFO) manner, and the cache size is set to 60 seconds. The data in the cache are organized according to the channels and stored with timestamps as indexes. Each data record contains information such as signal amplitude, acquisition time, channel number, and data status mark. In order to improve data access efficiency, the cache adopts a double buffering mechanism. While the current buffer is reading data, the newly acquired data is written to the backup buffer to implement a double buffering mechanism. When the amount of data in the cache reaches the preset threshold of 80%, the system automatically deletes the earliest 10% of the data to ensure sufficient space in the cache. Through data quality assessment, including calculating the kurtosis (threshold ±5) and skewness (threshold ±2) of the signal, abnormal data is marked, and a dynamic data cache containing the last 60 seconds of high-quality EEG data is obtained.

[0039] Window division is performed based on the valid data in the data cache, that is, the qualified data records of each channel in the last 60 seconds (signal amplitude, timestamp, and status mark are normal). The sliding window method is used for processing, and the window length is set to 2 seconds, that is, each window contains 512 sampling points, the overlap rate between adjacent windows is set to 50%, and the window moves forward 1 second each time it slides. This division method can ensure the continuity of the data while ensuring the effective capture of transient changes. During the window division process, the system first checks the update status of the data cache to ensure that the latest data is processed. For each divided window, the system calculates the signal-to-noise ratio (SNR). When the SNR is lower than 10dB, the window is marked as a low-quality window. At the same time, the system performs baseline drift detection on the data in the window, evaluates the stability of the baseline by calculating the first-order difference of the signal, and marks the presence of baseline drift when the difference value exceeds the preset threshold (twice the standard deviation). For the marked window, the system will examine the characteristics of the adjacent windows before and after it, including the mean, variance, and spectral characteristics. If the feature similarity of adjacent windows is high (correlation coefficient greater than 0.8), the window is retained and corrected; otherwise, it is marked as an invalid window. The system also analyzes the frequency domain characteristics of the signal in the window, uses fast Fourier transform (FFT) to calculate the power spectrum density, and detects whether there is obvious power frequency interference (50Hz) or other abnormal frequency components. For each valid window, the system generates a window descriptor, which contains the start and end time of the window, data quality indicators, abnormal flags, and other information, and obtains a series of data windows with quality marks.

[0040] Each data window that passes the quality inspection (SNR>10dB, stable baseline, normal spectrum) is processed in turn to generate a real-time sequence. The first step of preprocessing is baseline correction. The median filter is used to calculate the local baseline using a sliding window of 60 sampling points, and then the baseline is subtracted from the original signal to obtain the corrected signal. The second step is bandpass filtering, using a fourth-order Butterworth filter with a cutoff frequency set to 0.5Hz-45Hz to remove power frequency interference and other high-frequency noise. The filtered signal needs to be phase corrected to compensate for the phase delay introduced by the filter. For the preprocessed window data, the system performs normalization processing, using the z-score standardization method, and the calculation formula is Z=(X-μ) / σ, where X is the original value, μ is the mean of the data in the window, and σ is the standard deviation. In order to ensure the continuity of the real-time sequence, the system uses weighted averaging to fuse the overlapping areas of adjacent windows. The weight coefficient uses a cosine function to achieve a smooth transition. In the process of generating real-time sequences, the system has established a multi-level data quality control mechanism: first, artifact detection based on autocorrelation analysis. When obvious periodic artifacts are detected, wavelet transform is used for local denoising; second, the continuity of the data is ensured by calculating the correlation coefficient between adjacent windows (the default threshold is 0.6); finally, the statistical characteristics of the sequence are monitored, including mean shift detection and variance stability analysis. The system also calculates the correlation between multi-channel data to evaluate the spatial consistency of the signal, and finally obtains a real-time sequence that has been preprocessed and quality controlled.

[0041] Exemplarily, the feature marking of the real-time sequence and the extraction of state indicators include: extracting signal amplitude change characteristics of multi-channel EEG data in time domain analysis, decomposing signal energy distribution characteristics of multi-channel EEG data in frequency domain analysis, obtaining signal dynamic evolution characteristics of multi-channel EEG data in time-frequency domain joint analysis, and extracting nonlinear dynamic characteristics; constructing the state indicators according to the signal amplitude change characteristics, signal energy distribution characteristics, signal dynamic evolution characteristics and nonlinear dynamic characteristics.

[0042] The three types of key data obtained in the above steps (preprocessed time series, phase-corrected signal data, and spatial correlation indicators) are marked. Time domain feature analysis is performed on the preprocessed sequence to calculate the statistical features in each fixed time window, including mean, variance, skewness, and kurtosis. For example, in the high school classroom concentration detection scenario, when students are in a state of high concentration, their EEG signals usually show lower variance (reflecting the stability of the signal) and higher kurtosis (reflecting the regularity of the signal). In the state of distraction or fatigue, the variance of the signal will increase significantly, the kurtosis value will decrease, and the absolute value of the skewness will also increase. The change rules of these statistical features provide a reliable basis for real-time cognitive state assessment. In order to improve the accuracy of statistical features, the system divides the fixed window into several sub-segments according to the local stationary characteristics of the signal in the above steps. The zero-crossing rate feature is calculated on each sub-segment using a sliding window to characterize the rapid change characteristics of the signal. This rapid change is particularly evident when the student's state changes. For example, in the transition from concentration to distraction, the zero-crossing rate will show an obvious change pattern. Then, the phase-corrected signal is used for Hilbert transform to extract the instantaneous amplitude and phase information. Combined with the spatial correlation index of the above steps, the waveform consistency between multiple channels is calculated, and various characteristic wave parameters are extracted, including the time of occurrence, duration, peak value and amplitude. The identified characteristic waves are subjected to morphological analysis to calculate the steepness, sharpness and asymmetry indexes. These morphological features can effectively distinguish the EEG characteristic waves under different cognitive states, providing an important basis for subsequent state recognition. Finally, the signal is decomposed at multiple scales by continuous wavelet transform to obtain the time-frequency energy distribution of different frequency bands. Through these processes, four types of state indicators are obtained: statistical characteristics reflecting the overall distribution of the signal, temporal characteristics describing the change of the signal, characteristic wave parameters characterizing the waveform characteristics, and time-frequency characteristics describing the frequency characteristics.

[0043] The feature set is constructed using the output of four types of state indicators. For the statistical feature group (including mean, variance, skewness, and kurtosis), the robust normalization method is used to map them to a unified interval. For the time series feature group (including zero-crossing rate and instantaneous amplitude change rate), standardization is performed in combination with time window information. The characteristic wave parameter group (duration and amplitude characteristics of various brain waves) and the time-frequency feature group (energy distribution of each frequency band) are normalized according to their respective physical meanings. The correlation coefficient matrix is ​​calculated for all normalized indicators to identify strongly correlated indicator pairs. In the correlation analysis, the system adopts a segmentation strategy to ensure the stability of the correlation in different time periods. For the marked strongly correlated indicator pairs, the signal-to-noise ratio and discriminant ratio analysis are used to retain the features with strong distinguishing ability. For the time-frequency features, based on the energy distribution characteristics obtained in the first stage, the improved spectral estimation method is used for feature extraction. The feature extraction process takes into account the time-varying characteristics of the data, and the time resolution of the features is improved by adaptive window technology. The significance and stability of each feature are evaluated by variance analysis and resampling technology, and the stratified sampling strategy is used to ensure the representativeness of the evaluation sample. During the resampling process, the system comprehensively considers the balance between time continuity and state distribution to avoid bias in feature selection. Finally, the best performing features are selected to form the optimized feature set.

[0044] A multidimensional feature representation is generated based on the optimized feature set selected by the segment. This feature set contains statistical quantities that have been verified for significance (such as standardized variance and kurtosis), stable time series features (such as zero-crossing rate in specific frequency bands), key waveform parameters (such as duration ratio of α band) and important frequency domain features (such as θ / α energy ratio). First, these features are preprocessed by centering and standardization to eliminate scale differences. In the standardization process, the interquartile range-based method is used to improve the robustness to outliers. Then, principal component analysis is performed for dimensionality reduction. The covariance matrix of the standardized features is calculated and the eigenvalues ​​and eigenvectors are solved by singular value decomposition. The system adaptively determines the number of principal components based on the inflection point information of the eigenvalue decay curve, achieving effective dimensionality reduction while maintaining information integrity. In the construction of the principal component space, local structure preservation constraints are introduced to ensure that the dimensionality reduction process does not destroy the essential association of the features. Cross-validation is used to determine the optimal number of principal components, and various configurations are evaluated by feature reconstruction errors. A hierarchical verification strategy is designed for the system to ensure the reliability of the verification results. After projecting the original features into the principal component space, a high-order feature combination is constructed. The feature combination process adopts a progressive strategy, gradually building from low order to high order, and the effectiveness of the combined features is evaluated by the information gain criterion at each step. The kernel function method is used for nonlinear mapping, and the nonlinear relationship between features is captured by parameter optimization. In the kernel mapping process, the idea of ​​multi-kernel fusion is adopted to improve the flexibility of feature expression. Finally, a three-dimensional tensor structure is constructed according to the time dimension, space dimension and feature type dimension, realizing the multi-dimensional organization of features and obtaining the final multi-dimensional feature representation.

[0045] Combined with the multidimensional feature representation obtained in the above steps, the time domain transformation processing is performed. The system makes full use of the time dimension feature sequence (including the changing trend of various statistics over time), the spatial dimension feature (the spatial distribution feature of multi-channel signals) and the feature type dimension (the combination mode of different types of features) in the three-dimensional tensor structure. For the time dimension feature, the wavelet packet transform is used for multi-resolution analysis, and the most suitable wavelet basis function is selected for different feature types. For example, for the rapidly changing attention index, the Daubechies wavelet basis with good time localization characteristics is selected; for the slowly changing fatigue index, the Symlet wavelet basis with better frequency localization characteristics is selected. In the spatial dimension, based on the spatial distribution characteristics in the tensor, the spatial filtering technology is used to enhance the feature expression of specific brain regions. According to the functional characteristics of different brain regions, the corresponding spatial filter group is designed: for the prefrontal region, the β band features reflecting executive control are extracted; for the parietal region, the α band features related to attention are extracted. Through the matrix decomposition technology, the feature type dimension in the tensor is projected into different feature subspaces to achieve efficient expression of features. After these time domain transformations, two types of signal components are obtained: one is a multi-resolution component that reflects cognitive characteristics at different time scales, including fast-changing components (reflecting instantaneous state changes, such as attention fluctuations), medium-changing components (reflecting short-term state trends, such as changes in cognitive load) and slow-changing components (reflecting long-term state evolution, such as fatigue accumulation); the other is an enhanced spatial component that reflects the activity of key brain areas, which includes enhanced signals of each brain area of ​​interest and their interactive characteristics.

[0046] Frequency domain decomposition was performed on the multi-resolution components and enhanced spatial components obtained above. The system used short-time Fourier transform with different parameter configurations for each type of time scale signal component: for fast-changing components, a shorter analysis window (256 sampling points) was used to ensure time resolution; for medium-changing components, a medium window length (512 sampling points) was used to balance time-frequency resolution; for slow-changing components, a longer analysis window (1024 sampling points) was used to obtain more accurate spectrum estimation. The window length was adaptively adjusted according to the local stationarity of the signal, and the optimal window size was dynamically determined by calculating the local stability index of the signal (LSI = local variance / global variance). For the enhanced spatial components, the system performed frequency domain analysis from three levels: first, the power spectral density of a single brain region was calculated to obtain the energy distribution characteristics of each frequency band; then, the phase synchronization between brain region pairs was analyzed to construct a synchronization matrix based on the phase locking value (PLV); finally, a functional connection network between brain regions was established based on coherence analysis. In the process of network construction, an adaptive thresholding technique was used to remove weak connections, and the threshold was set to the 75% quantile of the connection strength distribution. Through these frequency domain decomposition processes, a complete set of frequency domain features are obtained, including: segmented power spectra of signals at each time scale, phase synchronization matrices between brain regions (reflecting the collaborative activity patterns of different regions) and dynamic functional connectivity network characteristics (describing the intensity of information interaction between brain regions).

[0047] The above frequency domain features are used to mark the feature intervals. Based on the power spectrum distribution obtained by decomposition, the system combines the phase synchronization matrix and functional connectivity features to adopt a multi-dimensional interval division strategy. In the frequency dimension, an improved clustering algorithm is used to automatically determine the personalized frequency band boundaries. First, the main frequency band center is identified based on the peak distribution of the power spectrum, and then the frequency band boundary is determined using the gradient change of the spectral density. For each identified frequency band, the system calculates three types of features: energy features based on the power spectrum (including absolute energy, relative energy and energy ratio), synchronization features based on the phase synchronization matrix (including average synchronization strength, synchronization stability and clustering coefficient of synchronization network), and network features based on functional connectivity (including connection density, node centrality and modularity). The system uses a multi-feature fusion method to evaluate the cognitive relevance of each frequency band and calculates the feature significance score S = w1×energy significance + w2×synchronization significance + w3×connection significance. The weight coefficient is obtained through historical data training. The frequency bands with significance scores exceeding the threshold are finely divided, such as the classic alpha band (8-13Hz) is subdivided into low alpha (8-10Hz, mainly related to alertness) and high alpha (10-13Hz, mainly related to cognitive processing) according to functional characteristics. The system also establishes a dynamic evaluation mechanism for feature intervals, and evaluates its stability by calculating the time-varying nature of interval characteristics. After these processes, a set of multi-dimensionally labeled feature intervals are obtained, each of which contains: frequency range definition, energy distribution characteristics, phase synchronization characteristics, functional connectivity characteristics, state correlation and stability score.

[0048] In some embodiments, the acquiring of the state representation corresponding to the signal component comprises: performing frequency domain decomposition on the signal component to mark a feature interval through the frequency domain decomposition; and generating the state representation according to the feature interval.

[0049] The state representation is generated based on the above-mentioned feature intervals with multi-dimensional labels. The system designs a hierarchical state mapping framework to systematically convert the multi-dimensional information of the feature interval into the state representation. In the time dimension, based on the energy change law of different feature intervals, a multi-scale description of state evolution is constructed. For fast-changing feature intervals (such as the β band), the system tracks the instantaneous change of its energy and identifies the key moment of state transition; for slower-changing feature intervals (such as the θ band), the system focuses on the cumulative effect of its energy and evaluates the long-term trend of the state. In the spatial dimension, the system comprehensively utilizes phase synchronization features and functional connectivity features to construct a multi-level spatial state representation: first, the instantaneous coordination mode of different brain regions is analyzed based on the phase synchronization matrix, and then the information interaction efficiency between brain regions is evaluated using the functional connectivity network features. Finally, these spatial features are integrated into a regional activation intensity map. In the frequency dimension, the system performs a weighted combination of the features of each labeled interval, and the weight coefficient is dynamically adjusted according to the state correlation of the interval. For example, in attention assessment, the weight of the high α wave interval will change adaptively according to the current task type. The system uses tensor decomposition technology to compress these multi-dimensional features into a compact state representation while retaining key state information. The final state representation is a multi-dimensional feature vector, which contains multi-scale state evolution characteristics in the time dimension, topological organization characteristics of brain area activities in the spatial dimension, and interaction characteristics between frequency bands in the frequency dimension.

[0050] The state representation obtained in the above steps is processed in layers. The system directly uses the multidimensional state representation (including time continuity features, spatial distribution features and frequency features) output by the above steps to construct a hierarchical structure. In the time dimension, the time continuity features in the state representation are divided into three levels according to the feature change rate: fast change layer (sampling interval <1 second, reflecting instantaneous state), medium change layer (sampling interval 1-10 seconds, reflecting short-term trends) and slow change layer (sampling interval >10 seconds, reflecting long-term trends). For each time level, the system calculates the statistical moment correlation of the features and constructs a feature association network within the layer. In the spatial dimension, a three-level spatial hierarchy is established based on the spatial distribution features in the state representation: local layer (single brain region activity features), regional layer (brain region group interaction features) and global layer (overall spatial pattern features). Each spatial level determines the degree of spatial aggregation of features by calculating the spatial autocorrelation index. In the frequency dimension, the frequency features in the state representation are used to construct a frequency band hierarchy: δ wave layer (1-4Hz), θ wave layer (4-8Hz), α wave layer (8-13Hz), and β wave layer (13-30Hz). At each frequency band level, the system analyzes the energy distribution and phase coupling characteristics of the frequency components. The inter-layer correlation matrix R is calculated for all hierarchical features, where Rij represents the correlation strength between the i-th layer feature and the j-th layer feature. Through this systematic hierarchical analysis, a hierarchical feature structure with three dimensions of time, space, and frequency and complete correlation between dimensions is obtained.

[0051] Based on the above hierarchical feature structure, cognitive attributes are labeled. The system first extracts significant cognitive state feature patterns at different levels of each dimension. At each level of the time dimension: the fast-changing layer analyzes the instantaneous attention fluctuation characteristics, and identifies significant fluctuation points by calculating the first-order difference of the feature sequence; the medium-speed change layer extracts the cognitive load change characteristics, and uses a sliding variance window to detect the load level transition; the slow-changing layer analyzes the fatigue accumulation characteristics, and determines the fatigue development stage based on the trend analysis of the long-term series. At each level of the spatial dimension: the local layer analyzes the activation pattern of a single brain area, such as the intensity of beta wave activity in the prefrontal region; the regional layer analyzes the coordination pattern of functionally related brain areas, such as the synchronous activity of the frontal-parietal network; the global layer evaluates the organizational characteristics of the overall brain network, such as the clustering coefficient and path length of the network. At each level of the frequency dimension: analyze the energy distribution characteristics of different frequency bands, such as alpha wave suppression indicates attention concentration, and theta wave enhancement reflects cognitive load. The system combines the features of each level to construct a multidimensional discriminant model of cognitive state. Based on the inter-layer correlation matrix in the hierarchical feature structure, the contribution weight of each feature combination to different cognitive attributes is calculated. Through multiple rounds of iterative optimization, a feature mapping table containing feature-attribute mapping relationships and hierarchical association strengths is finally obtained.

[0052] The state map is constructed using the feature map table with cognitive attribute labels. The system designs a multi-layer probabilistic graph model to describe the dynamic evolution of cognitive states. First, based on the feature-attribute correspondence in the feature map table, a state node network is constructed, and each node represents a specific cognitive state combination. The connection weight between nodes is determined by two parts: one is the hierarchical association strength in the feature map table, and the other is the state transition frequency observed in the historical data. For each state node, the system calculates three types of probability distributions: state maintenance probability (PM), which indicates the possibility of the state remaining stable; state transition probability (PT), which describes the possible path to transition to other states; state recovery probability (PR), which reflects the trend of recovery from abnormal states. In the process of network construction, the system pays special attention to the interaction between cognitive attributes. For example, there is a significant negative correlation between attention level and cognitive load, and the system captures this correlation through conditional probability modeling. For each possible state transition path, the temporal correlation RT and attribute correlation RA are calculated, and the path significance score S = w1×RT + w2×RA is obtained comprehensively. Based on the path significance, the system establishes a priority mechanism for state transition, giving priority to the transition path with high significance. Through this systematic modeling, we finally obtain a complete state mapping network including state nodes, transition probabilities, path weights and significance scores.

[0053] According to the above state mapping network, the system generates a detection sequence. The system makes full use of the four key elements in the network: using state nodes to calculate the activation probability of each cognitive state, using the transition probability between nodes to predict the possibility of state transition, combining path weights to determine the priority of state transitions, and judging the credibility of transitions by salience scores. Based on these elements, the system implements a multi-scale state sequence generation mechanism. On a short time scale (2-3 seconds), the system mainly uses state nodes and transition probabilities to detect state changes by calculating the activation probability of nodes in real time. When the activation probability of a state node exceeds a preset threshold (usually 0.75), the system generates a state transition marker. On a medium time scale (30 seconds), the system combines path weights and salience scores to analyze the evolution pattern of the state path and predict possible state transition sequences. For example, if it is detected that the attention level passes through a gradual path of 'high → medium → low' and the path has a high salience score, the system will generate an attention decay warning in advance. On a long time scale (5-10 minutes), the system evaluates the overall stability of the state network and quantifies the uncertainty of the state distribution by calculating the network entropy value H = -∑Pi×log(Pi), where Pi is the probability of the i-th state. For each detected state, the system simultaneously outputs four key indicators: state label (type of current cognitive state), confidence (calculated based on activation probability), change trend (based on path significance prediction) and warning level (based on network entropy evaluation). The system also establishes a quality control mechanism for state sequences, which evaluates the reliability of detection results by calculating the temporal consistency CI and spatial consistency CS of the sequence. The final generated detection sequence contains complete state evolution information, which can accurately reflect the current state and predict future state change trends.

[0054] Step S102: perform time series analysis on the detection sequence, extract change features, mark state nodes based on the change features, use the state nodes to build a cognitive sequence, and generate a state description based on the cognitive sequence; perform trend analysis on the state description, perform feature extraction on the analysis results corresponding to the trend analysis, build conversion rules based on the extracted features, and generate a prediction sequence based on the conversion rules.

[0055] Specifically, the detection sequence obtained in the above steps is subjected to time series analysis. The system fully utilizes four types of key information in the detection sequence: state label, confidence, change trend and warning level. A time transfer matrix is ​​established for the state label sequence, and weighted in combination with the confidence to obtain a reliable state transfer pattern. The stability index (the product of the state label and the confidence) is calculated for each time point to identify the state stability interval. In the change trend analysis, the linear trend is removed by the differential method to obtain the trend characteristics, and the remaining sequence is subjected to fluctuation analysis to obtain the fluctuation characteristics. The warning sequence is subjected to cumulative effect analysis, and the time decay factor is introduced to calculate the warning cumulative index. The system establishes a multi-scale mutation detection mechanism: the state label jump point is detected at a short time scale (5 seconds), the trend inflection point is analyzed at a medium time scale (30 seconds), and the significant change of the warning level is identified at a long time scale (5 minutes). For example, in a classroom teaching scenario, the system can identify the sudden change of students' attention (short scale), the continuous downward trend (medium scale) and the cumulative effect of fatigue (long scale). Through these systematic time series analyses, four sets of feature vectors are finally obtained: trend features (including state transition probability and trend slope), fluctuation features (including fluctuation amplitude and frequency), mutation features (including jump point position and intensity), and period features (including period length and phase information).

[0056] In some embodiments, the state nodes include slope change points, fluctuation anomaly points and fluctuation nodes; the marking of state nodes based on the change characteristics includes: performing piecewise linear fitting on trend characteristics corresponding to the change characteristics and identifying slope change points; using an adaptive threshold method to detect fluctuation anomaly points on fluctuation characteristics corresponding to the change characteristics; calculating the local fluctuation intensity corresponding to the change characteristics and marking fluctuation nodes that exceed the standard deviation of the historical mean.

[0057] State nodes are marked based on the above four sets of feature vectors. For trend features, the slope significant change points are identified by piecewise linear fitting, and when the change rate exceeds the threshold, they are marked as trend nodes. For fluctuation features, the adaptive threshold method is used to detect fluctuation anomalies, calculate the local fluctuation intensity, and mark them as fluctuation nodes when they exceed two standard deviations of the historical mean. For mutation features, the cumulative sum statistic is used to identify mutation points, and when the statistic exceeds the judgment threshold, it is marked as a mutation node. For periodic features, the main period is determined by autocorrelation analysis, and the starting point and peak and valley points of the period are marked as periodic nodes. Three key indicators are calculated for each marked node: node significance (based on the weighted sum of slope change, fluctuation intensity and mutation amplitude), impact duration (determined by analyzing the state changes before and after the node), and propagation range (based on connectivity analysis of the state space). In practical applications, such as remote teaching scenarios, when the system detects that multiple students have significant fluctuations in attention at the same time, a high-significance fluctuation node will be generated and its impact range will be evaluated. Through this multi-dimensional node marking, a complete set of state nodes is obtained, each of which contains timestamps, node types, state features, significance scores, durations, and impact ranges.

[0058] The cognitive sequence is constructed using a set of marked state nodes. The system first establishes a time series skeleton based on the timestamp of the node to determine the basic framework of the sequence. The dominant change pattern of each time window is analyzed according to the node type (trend, fluctuation, mutation, cycle), and the distribution ratio of different types of nodes is calculated. The state transition matrix is ​​constructed using state feature data to describe the jump rules between states. The significance score of each node is normalized as a weight coefficient of state importance. High-significance nodes receive greater weight in sequence reconstruction. The stability of the state is evaluated based on the duration information of the node. The longer the duration, the higher the reliability score in the sequence. The spatial distribution characteristics of the state are determined by analyzing the influence range of the node. The system performs special processing on nodes with overlapping influence ranges to maintain spatial consistency. Three key factors are considered in sequence construction: temporal continuity (ensuring the rationality of state transitions between adjacent nodes), spatial consistency (evaluating state distribution based on node influence range), and state stability (through node duration analysis). For example, in classroom teaching, the system can build a sequence that reflects the changes in students' cognitive states based on continuous attention fluctuation nodes: first determine the sequence framework based on timestamps, identify the change pattern based on node types, analyze the change rules through state characteristics, use significance scores to highlight important transition points, evaluate state stability based on duration, and finally consider the impact range to ensure reasonable spatial distribution. Through iterative optimization, a state sequence with the highest reliability score is obtained, which contains complete temporal relationships, state transition rules, and importance weights.

[0059] In some embodiments, generating a state description based on the cognitive sequence includes: obtaining the time interval between adjacent states corresponding to the cognitive sequence to generate a basic description; extracting the state transition path rules in the basic description to generate a behavior pattern description; filtering high-weight state transitions in the pattern description in combination with the importance weight to generate a priority description; dynamically adjusting the detail of the pattern description according to the application scenario requirements, and generating a state description including time series statistics and state evolution process.

[0060] Generate state descriptions based on the constructed cognitive state sequence. A multi-level description generation framework is used to combine quantitative information and qualitative analysis. At the micro level, the system generates basic descriptions based on the temporal relationship in the sequence: analyze the time intervals between adjacent states, calculate the frequency distribution of state occurrences, and identify key temporal patterns, such as "three consecutive state transitions occurred within 10 minutes." Generate behavioral pattern descriptions based on state transition rules: analyze the main paths of state transitions, such as "the level of attention has undergone a gradual change from high to medium to low"; extract typical sequences of state transitions, such as "fatigue is usually accompanied by 3-4 fluctuations in attention"; summarize the periodic laws of state changes, such as "a significant decrease in attention occurs every 20-25 minutes." Use importance weights to highlight key state changes: high-weight state transitions are described first, such as "the most significant state change occurred in the 45th minute of the course, and the level of attention dropped sharply"; weight information is also used to filter and sort the importance of the description content. The system dynamically adjusts the level of detail of the description according to the needs of different application scenarios: for real-time monitoring by teachers, it generates brief high-weight state change prompts; for teaching evaluation, it generates a detailed report containing complete time series analysis and conversion rules. The final output of the complete state representation consists of two parts: a quantitative analysis report (including time series statistics, conversion probability and weight distribution) and a qualitative description report (including text descriptions of the state evolution process, key conversion points and periodic rules).

[0061] Perform trend analysis on the state representation obtained in the above steps. The system processes the information in the quantitative analysis report and the qualitative description report separately. Decompose the statistical indicator sequence (including attention level, cognitive load, fatigue level, etc.) in the quantitative analysis report, and use the STL method to decompose each indicator sequence into three components: the trend term reflects the long-term change trend and is extracted by local weighted regression; the cycle term represents the cyclical change characteristics, which is determined based on the cycle law identified in the previous analysis (such as the 45-minute attention fluctuation cycle); the random term contains short-term fluctuation information. At the same time, analyze the state description in the qualitative description report and extract key state transition information, such as the gradual trend of "high → medium → low" is converted into a state sequence. For each sequence, calculate the trend strength (trend item variance / total variance) to reflect the trend significance, the cycle strength (cycle item variance / total variance) to indicate the degree of cycle change, and the noise ratio (random item variance / total variance) to measure the fluctuation strength. The system also analyzes the interaction between components, calculates the correlation between the trend term and the cycle term, and the cycle term and the random term, and evaluates the independence of each component. In practical applications, such as classroom attention monitoring scenarios, this decomposition can clearly identify the long-term trend of student attention (such as the decline caused by fatigue accumulation), inherent cycles (such as 45-minute natural fluctuations), and random disturbances (such as fluctuations caused by external interference). Through this systematic trend analysis, a complete set of trend characteristics is finally obtained, including the trend function, periodic pattern, fluctuation characteristics of each indicator, and the correlation indicators between them.

[0062] The above trend features are extracted. The system first processes the trend function, identifies the inflection point position through piecewise linear fitting, calculates the slope of each segment to represent the rate of change, and marks the point as a trend turning point when the slope change of adjacent segments exceeds the preset threshold. At the same time, the acceleration feature of the trend is calculated to evaluate the rate of change. In the periodic pattern analysis, the system extracts the main periodic components through Fourier analysis, and calculates its energy contribution and phase characteristics for each significant period. The system pays special attention to the cycles related to cognitive laws, such as the natural fluctuation cycle of attention (about 45 minutes) and the fatigue accumulation cycle (about 2 hours). For the fluctuation characteristics, the system calculates the local fluctuation intensity, and marks it as a high fluctuation interval when the fluctuation intensity exceeds two standard deviations of the global mean; the aggregation effect and pattern characteristics of the fluctuation are identified through autocorrelation analysis. The system also uses correlation indicators for in-depth analysis to study the temporal association and influence relationship between different features. The system establishes a multi-scale risk assessment mechanism: assessing risk based on the slope change rate on the trend scale, judging risk by the degree of amplitude abnormality on the period scale, and determining the risk level based on the intensity accumulation effect on the fluctuation scale. For example, in a distance learning scenario, when the attention trend is detected to be rapidly decreasing (high slope), the amplitude of periodic fluctuations is abnormally increasing, and local fluctuations are frequent, the system will mark it as a high-risk state. Through these feature extractions, a complete feature expression set is obtained, including key time point sequences, periodic feature vectors, fluctuation pattern descriptions, and multi-scale risk ratings.

[0063] Construct transformation rules based on the above feature expression set. The system maps the key time point sequence to the state space, determines the state transformation type and specific time corresponding to each time point; at the same time, uses the periodic feature vector to construct the periodic transformation constraint, accurately defines the periodic law of state transformation; establishes the state fluctuation rule according to the fluctuation pattern description, and characterizes the specific manifestation of state fluctuation; determines the priority and trigger threshold of the rule based on the multi-scale risk rating. The system uses the decision tree algorithm to construct the initial rule framework: the key time point is used as the split node, and the periodic feature and fluctuation feature are used as the classification attribute. The form of each rule is a condition-transformation correspondence, such as "when the high fluctuation lasts for more than 10 minutes and the trend slope is less than the threshold, the predicted attention decreases significantly". Reliability indicators are calculated for each rule: trigger probability (frequency of rule use), conversion accuracy (rule success rate), time correlation (stability of the rule at different times) and spatial correlation (applicability of the rule in different scenarios). Based on these indicators, the rules are assigned confidence scores. For complex state transitions, the rule sequence is constructed through rule combination operators (AND, OR, THEN). In teaching scenarios, the system can generate combined rules such as "three consecutive fluctuations in attention with a downward trend predict the onset of cognitive fatigue." Finally, a complete conversion rule library is obtained, including trigger conditions, conversion probability, time window, and reliability score.

[0064] Generate prediction sequences based on the conversion rule base. The system monitors the feature flow in real time, matches the current state with the trigger conditions, and calculates the degree of matching; once the rule is activated, the system generates possible state sequences based on its conversion probability and performs multi-step predictions within the time window; at the same time, the reliability score is used as the weight coefficient to generate multiple weighted evolution paths through the Monte Carlo method. In the sequence integration stage, the system combines multiple prediction paths, and the weight coefficient is determined by the reliability of the rule. The system evaluates the uncertainty of the prediction: the degree of uncertainty is quantified by calculating the entropy value of the prediction distribution, and the confidence interval is estimated based on historical errors. In applications, such as classroom concentration prediction, the system can warn of possible attention loss 3-5 minutes in advance, and give a probability estimate and confidence interval for the warning. The final generated prediction sequence contains state prediction values, probability distributions, timestamps, and confidence intervals, providing predictive guidance for teaching adjustments.

[0065] Step S103, threshold division is performed on the prediction sequence, state intervals are marked, feature aggregation is performed according to the state intervals, a criterion sequence is constructed using the aggregation results corresponding to the feature aggregation, and an evaluation result is generated according to the criterion sequence; signal detection is performed on the evaluation result to obtain detection information.

[0066] Specifically, the prediction sequence obtained in the above steps is divided by threshold. The system processes the state prediction value, probability distribution, timestamp and confidence interval in the prediction sequence respectively. First, a multi-level threshold is set based on the state prediction value, and an adaptive strategy is used to determine the key segmentation point. The system establishes an initial threshold candidate set by calculating the local statistical characteristics of the prediction sequence, including mean, variance and skewness; for each candidate threshold, its classification effect is calculated, and the threshold combination with the best performance is selected. By analyzing the multi-peak characteristics of the probability distribution, the system sets a separation threshold between each peak, and introduces an adaptive smoothing strategy to avoid too frequent fluctuations in the threshold. The sequence is segmented in combination with the timestamp information, and the time series correlation matrix is ​​established. The time law of state transition is analyzed to ensure that the threshold boundary is aligned with the time series characteristics. At the same time, the confidence interval is used to adjust the threshold. When the prediction confidence is between 0.7-0.9, the threshold range is appropriately relaxed; when the confidence is lower than 0.7, the threshold tolerance is significantly expanded to reduce the risk of misjudgment. In the distance education scenario, the system divides the attention level into four levels: highly concentrated (predicted value > 0.8 and confidence interval fluctuation < 0.1), normally concentrated (0.6-0.8), distracted (0.4-0.6) and inattentive (< 0.4). In the threshold adjustment process, a two-way verification mechanism is introduced: forward verification ensures the applicability of the threshold to historical data through a sliding time window (typical window size is 30 minutes), and backward verification divides the recent data into multiple subsets to evaluate the generalization ability of the threshold to new data. In addition, the system fine-tunes the threshold every 10 minutes, and the adjustment range does not exceed 5% of the current threshold to maintain the stability of the system. Through this multi-dimensional threshold division strategy, a set of state intervals with level labels are finally obtained.

[0067] Based on the above-mentioned state intervals with level marks, feature aggregation is performed. The system designs a multi-level feature aggregation strategy to integrate and analyze the features of the state intervals from the two dimensions of time and space. In the time dimension, continuous intervals with the same state level are merged, and an adaptive sliding window is used to calculate the duration and state stability of the merged interval. The system dynamically adjusts the window size by analyzing the frequency characteristics of state changes. The typical range is between 30 seconds and 5 minutes. A smaller window is used when the state changes rapidly, and a larger window is used when the state is stable. The state duration index is calculated for each merged interval. This index not only considers the interval length and prediction confidence, but also introduces a state intensity factor to reflect the significance of the state. When a continuous "distracted" interval is found, the system calculates its cumulative impact and assigns a 1.5-fold weight to the interval with a confidence level higher than 0.8. At the same time, considering the temporal position of the interval, the weight of the later interval will be appropriately increased to reflect the cumulative effect of fatigue. In the spatial dimension, the system analyzes the distribution density and conversion frequency of intervals of different state levels, establishes a state transition matrix, and identifies the key moments and typical patterns of state transitions. The system also introduces a state evolution graph, which describes the transition relationship between states through a graph structure. The weight of the edge represents the transition probability, and the size of the node reflects the stability of the state. For example, in classroom teaching, the system can identify the typical transition pattern of a student group from concentration to fatigue, and quantify the significance of this transition by calculating the probability of the state path. Through this multi-dimensional feature aggregation, a complete set of comprehensive features is obtained, including state continuity indicators, state transition indicators, stability indicators, and trend prediction indicators.

[0068] The above comprehensive features are used to construct the criterion sequence. The system processes various indicators separately: based on the state persistence indicator, the time dimension criterion is constructed, and the state duration is divided into three levels: short-term (less than 5 minutes), medium-term (5-15 minutes) and long-term (more than 15 minutes), and differentiated thresholds are set according to different state types. The state transition indicator is used to establish a dynamic criterion, which not only calculates the frequency characteristics of the state transition, but also analyzes the directionality and intensity of the transition, and assigns a higher risk weight to the state transition that goes back and forth quickly. The fluctuation monitoring criterion is set according to the stability indicator, and the multi-level fluctuation warning threshold is established by calculating the state variance and coefficient of variation in the local time window. The trend assessment criterion is constructed in combination with the trend prediction indicator. The system uses the weighted moving average and trend extrapolation method to predict the evolution trend of the state. The system also designs a combined criterion based on fuzzy rules, constructs a composite evaluation condition by connecting different characteristic indicators, and introduces time sequence constraints to ensure the temporal consistency of the criterion. For example, when the trend prediction indicator shows a continuous downward trend for more than 10 minutes, and the state transition indicator exceeds the preset threshold three times within 5 minutes, the attention significant decline criterion is triggered. Each criterion includes evaluation conditions, trigger thresholds, weight coefficients, and reliability scores. The reliability of the criterion is calculated through historical data verification. The forgetting factor mechanism is introduced to make the most recent data have a higher reference value. Criteria with scores below 0.6 will be automatically adjusted or eliminated. In teaching applications, the system can generate specific criteria such as "when the attention level fluctuates more than 3 times within 10 minutes and the trend is downward." By integrating these different types of criteria, a complete criterion sequence is formed.

[0069] The evaluation results are generated according to the above criterion sequence. The system first processes the evaluation conditions, trigger thresholds, weight coefficients and reliability scores in the criterion sequence to build a multi-level evaluation framework. The system uses the analytic hierarchy process to calculate the importance of different conditions and dynamically adjusts the weights based on real-time feedback. The trigger threshold is used to determine the warning level. The system sets up a three-level warning mechanism, and each warning level is equipped with corresponding intervention suggestions. Weighted calculation is performed based on the weight coefficient. The weight value is obtained through historical data training, and a dynamic update mechanism is set. Combined with the reliability score to ensure the accuracy of the evaluation, the system introduces confidence weighting, and high reliability criteria have a greater impact in the final evaluation. For example, when the prediction accuracy of a criterion exceeds 90% for more than 30 minutes, its weight will be increased by 20%. The system has designed an adaptive warning mechanism that not only considers the threshold offside of a single indicator, but also analyzes the coordinated changes of multiple indicators. When the score is lower than the warning threshold, a graded warning signal is automatically generated according to the offside degree and duration. In the evaluation report, the system not only provides quantitative scoring results, but also gives improvement suggestions based on specific trigger criteria, such as "detecting that attention shows a 20-minute periodic fluctuation, it is recommended to take a short break of 3-5 minutes every 20 minutes and adjust the teaching rhythm appropriately." By real-time monitoring of the changing trend of the evaluation score and combining historical data patterns, the system can predict potential status problems 3-5 minutes in advance and provide preventive suggestions. The final evaluation results include the overall score, sub-item scores of each dimension, status diagnosis report and specific improvement suggestions, providing comprehensive decision-making support for teaching adjustments.

[0070] Signal detection is performed on the evaluation results obtained in the above steps. The system processes the overall score, sub-item score, status diagnosis report and improvement suggestions in the evaluation results respectively. First, a baseline analysis is established for the overall score sequence, and the local mean and standard deviation are calculated through an adaptive sliding window. The window size is dynamically adjusted within the range of 1-5 minutes according to the degree of data fluctuation. The system sets three levels of detection thresholds: mild deviation (baseline ±1.5σ), moderate deviation (baseline ±2σ) and severe deviation (baseline ±3σ), and introduces a time weighting mechanism to make the most recent data have a higher reference value. Differentiated detection strategies are adopted for sub-item scores, such as setting a more sensitive fluctuation threshold (±1.2σ) for the attention dimension and focusing on mutation characteristics, while a wider threshold (±2.5σ) is used for the fatigue dimension but more attention is paid to the detection of cumulative effects. The system monitors mutation anomalies at a micro scale of 10 seconds and analyzes trend anomalies at a macro scale of 5 minutes. Combined with the status diagnosis report, the system establishes a semantic analysis module to extract key status description information, identifies description fragments related to anomalies through natural language processing technology, and establishes a problem-symptom mapping relationship. The improvement suggestions are systematically tracked, and a before-and-after comparison window is set (5 minutes before the suggestion vs. 5 minutes after the suggestion) to record and quantify the degree of improvement after the suggestion is implemented. In the teaching scenario, the system can monitor individual anomalies and group anomalies at the same time. When more than 30% of the students are detected to have the same anomaly, the group anomaly alarm is triggered. Through this multi-dimensional and multi-scale signal analysis, the detection results including the degree of anomaly, duration and scope of impact are obtained.

[0071] The above test results containing abnormal degree, duration and impact range are classified and marked. The system adopts a multi-level marking strategy. First, a five-level classification system is established based on the degree of abnormality: normal fluctuation (within ±σ), slight abnormality (±1.5σ), moderate abnormality (±2σ), significant abnormality (±2.5σ) and severe abnormality (±3σ and above). Different visual marks and processing priorities are set for each level of abnormality. Severe abnormalities are marked with red flashing marks and trigger real-time alarms. In the duration analysis, the system designs a time series classification algorithm to divide abnormal patterns into three categories: instantaneous abnormalities (lasting <30 seconds) are identified by calculating the change rate ΔS of adjacent time points. When the change rate exceeds 40% and does not last for more than 3 sampling points, it is marked as a mutation point; fluctuation abnormalities (lasting 30 seconds-3 minutes) are identified by spectrum analysis, and the fluctuation frequency and amplitude in the local time window are calculated; cumulative abnormalities (lasting >3 minutes) are identified by trend analysis methods, and the cumulative deviation index and change slope in the time window are calculated. The spatial clustering method is used to evaluate the scope of influence, calculate the distribution density and diffusion speed of anomalies in the group, and divide the scope of influence into local influence (<10% group), regional influence (10%-30% group) and global influence (>30% group). The system also establishes a composite labeling mechanism to comprehensively consider the multiple dimensional characteristics of anomalies. For example, in a distance learning scenario, when it is detected that the attention indicators of more than 25% of students are simultaneously moderately abnormal and last for more than 2 minutes, the system will mark the anomaly as "significant decrease in group attention" and automatically upgrade the alarm level. Temporal consistency constraints are introduced in the labeling process to prevent misjudgment caused by frequent fluctuations in labels. Through this multi-dimensional and multi-level classification labeling, a labeling result set containing anomaly type, degree level and spatial distribution is obtained.

[0072] Based on the above labeled result set containing abnormality type, degree level and spatial distribution, abnormal feature recognition is carried out. The system first builds a hierarchical feature pattern library according to the abnormality type, storing the feature templates of various typical abnormalities, such as attention fluctuation, cognitive load and fatigue accumulation. For each abnormality type, the system designs a specific feature extraction algorithm: for attention abnormalities, the focus is on extracting α band energy changes and eye movement pattern features; for cognitive load abnormalities, the focus is on analyzing θ band activities and task switching frequency; for fatigue accumulation abnormalities, the focus is on features such as β band energy reduction and prolonged reaction time. The system uses degree level information to establish a feature significance evaluation mechanism, giving 1.5 times the feature weight to high-level abnormalities to ensure that the features of severe abnormalities can be analyzed first. Based on spatial distribution data, the system uses spatial statistical methods to analyze the propagation mode and aggregation effect of abnormalities, and calculates the Masolav spatial correlation coefficient to evaluate the spatial autocorrelation of abnormalities. In the feature extraction process, the system uses segmented statistical methods to set three time windows: before the abnormality occurs (first 2 minutes), when it occurs (during the duration of the abnormality) and after it occurs (after 2 minutes) to analyze the complete process of state change. For example, for abnormalities such as sudden drop in attention, the system analyzes its triggering conditions (such as the continuous presentation of difficult knowledge points), development process (such as the step-by-step decline in attention level, usually going through 4-5 stages) and impact consequences (such as the continued decline in learning efficiency, usually lasting 5-8 minutes). The system evaluates the significance by calculating the discrimination of features in normal samples and abnormal samples (measured by F-score), and features with a significance score of more than 0.8 are retained as key features. The system also establishes a combination analysis mechanism for features, discovers synergy and redundant information by calculating the mutual information and conditional entropy between features, and constructs a minimum feature set to ensure interpretation efficiency. In classroom teaching applications, the system can identify typical abnormal patterns such as "group attention fluctuations induced by content difficulty" and extract its key feature combinations. Through this comprehensive feature analysis, a set of feature vectors describing abnormal patterns is obtained, including temporal features, spatial features, and correlation features.

[0073] In some embodiments, performing signal detection on the evaluation result to obtain detection information includes: classifying and marking the detection result corresponding to the signal detection, identifying abnormal features based on the classification marks; and generating the detection information according to the abnormal features.

[0074] The detection information is generated according to the above-mentioned abnormal feature vector containing time series features, spatial features and correlation features. The system adopts a structured information generation framework and designs a three-layer information template: the core layer contains the basic description of the abnormality; the analysis layer contains quantitative indicators and spatiotemporal features; the recommendation layer contains intervention measures and expected effects. The system first uses the time series features to analyze the development process of the abnormality, identifies the key turning points and evolution stages through the time-frequency analysis of the feature sequence, and accurately describes the occurrence time, duration and evolution trend of the abnormality. For example, for the attention fluctuation abnormality, the system can accurately describe "the fluctuation began at the 32nd minute of the course, experienced 3 significant declines, each decline lasting about 45 seconds, and generally showed a step-by-step downward trend." The system determines the impact range and distribution pattern of the abnormality through spatial features, uses heat map visualization technology to display the distribution of the abnormality in the group, and calculates the spatial autocorrelation index to evaluate the aggregation effect of the abnormality. For example, the system can detect that "the decline in attention first appeared in the back row of the classroom, and then spread forward at a rate of 0.8 people / minute, eventually affecting 78% of the student population." Based on the correlation features, the system identifies the causal relationship and chain effect between anomalies, constructs an anomaly correlation network, and determines the root cause and impact path of the anomaly through Bayesian network analysis. In the process of generating detection information, the system adjusts the level of detail of the description according to the type and severity of the anomaly, and provides more detailed analysis information for high-risk anomalies. For different application scenarios, the system customizes the information output format: for real-time monitoring of teachers, it generates a concise anomaly summary and key indicators; for teaching evaluation, it provides a detailed report with complete analysis and historical comparison. The system also establishes an information update mechanism, which automatically updates the detection information and marks the change trend when the abnormal state changes. The final generated detection information includes anomaly description ("a significant decrease in group attention occurred in the 32nd minute of the course"), quantitative indicators ("the impact range is 78%, the duration is 8.5 minutes, and the severity is 4.2 / 5"), development trend ("a step-by-step decline, no natural recovery trend") and correlation analysis ("a significant correlation with previous high-difficulty knowledge points, with a confidence level of 0.92"), providing precise guidance for teaching intervention.

[0075] Step S104, performing data conversion on the detection information, constructing a display sequence, organizing states based on the display sequence, marking state types according to the organization results of the state organization; and generating cognitive state information according to the state type.

[0076] Specifically, the detection information obtained in the above steps is converted. The four types of data in the detection information are systematically processed: abnormal description, quantitative indicators, development trend and correlation analysis results. First, the abnormal description is semantically analyzed, key event features and time tags are extracted, and event-time mapping relationships are established. Then, the quantitative indicators are mapped to dimension reduction, and the high-dimensional features are mapped to the visualization space using multidimensional scaling technology. In the mapping process, the degree of abnormality is converted to display intensity, the impact range is mapped to spatial distribution, and the time series association is converted to connection strength. The system designs an adaptive normalization scheme to ensure the comparability of different types of indicators. For example, the fluctuation amplitude of the attention level and the cumulative value of the fatigue level are uniformly converted to a display interval of 0-100 for intuitive comparison. For the development trend information, the system establishes a time series mapping rule, converts the trend slope into a direction vector, and maps the trend significance into vector strength. The system also introduces a trend prediction component to generate short-term predictions through regression methods. For the correlation analysis results, a network structure representation is constructed, the correlation strength determines the connection weight, and the correlation type is represented by different line types. For example, in a classroom scenario, the system can convert the attention distribution pattern of a group of students into a map of different shades of color, and the propagation path of attention fluctuations is represented as a directional line. Through this systematic data conversion, a set of standardized display sequences is obtained.

[0077] The state organization is based on the above standardized display sequence. The system adopts a hierarchical organizational structure to systematically integrate different elements in the display sequence. The system uses the display intensity data in the display sequence to evaluate the prominence of the state and set a three-level attention threshold; divides the individual, local and overall levels through spatial distribution information; analyzes the changing trend of the state based on the direction vector and vector intensity; and uses the network structure to represent the relationship between states. Taking the classroom scene as an example, the spatial dimension is divided into three levels: individual level (single student), local level (front row, middle row, back row and other areas) and overall level (whole class). For the individual level, the system establishes a state timeline based on the display intensity in the display sequence, and records key state nodes, such as attention peaks, troughs and stable intervals. At the local level, using the spatial distribution information of the display sequence, the density clustering method is used to identify student groups with similar states and calculate the representative characteristics of the group. For example, when it is found that the back row student group generally has a phenomenon of inattention, the system will organize this local aggregation effect separately and calculate its impact intensity. At the overall level, the direction vector and vector strength of the display sequence are combined to construct a global view of state evolution, and the network structure is used to analyze the association pattern between states and identify the path and speed of state propagation. Through multi-level state organization, a hierarchical state structure containing individual characteristics, group characteristics and overall characteristics is obtained.

[0078] The hierarchical state structure containing individual characteristics, group characteristics and overall characteristics is used to mark the state type. Based on cognitive science theory, the system establishes a multi-dimensional state classification system in combination with the specific manifestations in the hierarchical structure. At the individual characteristic level, the system establishes a multi-dimensional feature space, including key indicators such as attention level, cognitive load, fatigue, and participation. Based on the combination pattern of these indicators, the students' cognitive states are subdivided into basic types such as active involvement, normal learning, distraction and cognitive fatigue. The system uses a fuzzy classification method to calculate the membership of the state type based on the similarity between the feature vector and the standard template. At the group characteristic level, the system identifies common collective phenomena and constructs a group state dictionary, including patterns such as group attention fluctuations caused by difficulty in understanding knowledge points and overall fatigue caused by continuous high-intensity learning. The system analyzes the group consistency index to evaluate the degree of synchronization of the state. Based on the spatial distribution characteristics in the overall characteristics, the system analyzes the propagation effects of different types of states, such as the positive impact of the concentration state of the front-row students on the surrounding students. The system calculates the spillover effect and aggregation degree of the state through spatial statistical methods to identify key influencing points. For each state type, the system establishes a detailed feature description template, including quantitative judgment criteria and qualitative judgment basis. The system also implements a state transition rule library to describe common state transition sequences and trigger conditions. Through this systematic state classification, a complete set of state type tags is obtained.

[0079] Cognitive state information is generated based on the state type tag set containing individual characteristics, group characteristics and overall characteristics. The system designs a hierarchical information generation architecture to convert state characteristics of different dimensions into specific cognitive state descriptions. The system converts individual characteristics into state descriptions and suggestions at the individual level, and uses natural language generation technology to select appropriate expression templates according to the state type. At the individual level, a personal state report containing the current state type, risk level and warning information is generated, and the update frequency of the state report is dynamically adjusted according to the speed of state change. For example, the system will push personalized state reminders such as "Your attention level has fluctuated downward in the past 15 minutes and has fallen below the personal baseline by 35%. It is recommended to take a proper rest or adjust the learning strategy" to the user in real time. The system generates a collective state report at the team or class level based on group characteristics, including state distribution statistics and typical group characteristics. In group application scenarios, such as multi-person collaborative office environments, the system can analyze the state synergy of team members. For example, when it is detected that multiple members of a project team are inattentive at the same time, the system will generate state information such as "The project team (5 people) has generally experienced a decline in attention levels in the past 30 minutes, and the average attention index has dropped to 65%. It is recommended to arrange a short break or change the task type in time." The system uses the overall features to construct macro trend analysis and identify long-term status patterns and periodic change patterns. For individual deep work scenarios, the system will also combine the macro trends in the overall features to generate more targeted status reports, such as "Your attention pattern shows a significant fluctuation every 70-80 minutes, and it is recommended to adjust the work rhythm." Through this scenario-based and personalized information generation method, a complete cognitive status information set is finally obtained.

[0080] The method has the following beneficial effects:

[0081] 1. The innovative use of dynamic data caching and multiple window processing mechanisms, combined with time domain transformation and frequency domain decomposition technology, achieves data continuity and signal integrity in the feature extraction process, ensuring the stable performance of portable devices during data acquisition and processing.

[0082] 2. By establishing a hierarchical state representation and time series analysis system, combined with a trend prediction and evaluation mechanism, the system can adaptively adjust detection parameters and achieve accurate monitoring of the cognitive states of different individuals.

[0083] 3. A complete state anomaly detection and information generation process has been built, integrating multi-dimensional feature analysis and state association recognition functions, enabling the system to comprehensively describe and display changes in the user's cognitive state.

[0084] In order to implement the real-time detection method of cognitive state of attention glasses corresponding to the above method embodiment, so as to achieve corresponding functions and technical effects. Figure 2, Figure 2 The structure block diagram of a real-time detection device 200 of cognitive status provided by an embodiment of the present application is shown. For the convenience of explanation, only the parts related to the present embodiment are shown. The real-time detection device 200 of cognitive status provided by an embodiment of the present application includes:

[0085] The data acquisition unit 201 is used to acquire multi-channel EEG data of the attention glasses, acquire multi-dimensional features corresponding to the multi-channel EEG data; perform time domain transformation according to the multi-dimensional features, acquire signal components, and acquire state representations corresponding to the signal components; perform hierarchical processing on the state representations to determine feature levels; mark cognitive attributes based on the feature levels; construct a state map according to the cognitive attributes, and generate a detection sequence according to the state map;

[0086] The timing analysis unit 202 is used to perform timing analysis on the detection sequence, extract change features, mark state nodes based on the change features, construct a cognitive sequence using the state nodes, and generate a state description according to the cognitive sequence; perform trend analysis on the state description, perform feature extraction on the analysis result corresponding to the trend analysis, so as to construct a conversion rule according to the extracted features, and generate a prediction sequence according to the conversion rule;

[0087] The threshold division unit 203 is used to perform threshold division on the prediction sequence, mark the state interval, perform feature aggregation according to the state interval, construct a criterion sequence using the aggregation result corresponding to the feature aggregation, generate an evaluation result according to the criterion sequence; perform signal detection on the evaluation result to obtain detection information;

[0088] The data conversion unit 204 is used to perform data conversion on the detection information, construct a display sequence, organize states based on the display sequence, mark state types according to the organization results of the state organization; and generate cognitive state information according to the state type.

[0089] The above-mentioned cognitive state real-time detection device 200 can implement the cognitive state real-time detection method of the attention glasses of the above-mentioned method embodiment. The optional items in the above-mentioned method embodiment are also applicable to this embodiment and will not be described in detail here. The rest of the contents of the embodiment of this application can refer to the contents of the above-mentioned method embodiment, and will not be repeated in this embodiment.

[0090] Figure 3 This is a schematic diagram of the structure of a computer device provided in one embodiment of the present application. Figure 3 As shown, the computer device 3 of this embodiment includes: at least one processor 30 ( Figure 3Only one is shown in the figure), a memory 31 and a computer program 32 stored in the memory 31 and executable on the at least one processor 30, wherein the processor 30 implements the steps of any of the above method embodiments when executing the computer program 32.

[0091] The computer device 3 may be a computing device such as a smart phone, a tablet computer, a desktop computer, a cloud server, etc. The computer device may include but is not limited to a processor 30 and a memory 31. Those skilled in the art will understand that Figure 3 It is only an example of computer device 3 and does not constitute a limitation on computer device 3. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components, for example, it may also include input and output devices, network access devices, etc.

[0092] The processor 30 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0093] In some embodiments, the memory 31 may be an internal storage unit of the computer device 3, such as a hard disk or memory of the computer device 3. In other embodiments, the memory 31 may also be an external storage device of the computer device 3, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the computer device 3. Further, the memory 31 may also include both an internal storage unit and an external storage device of the computer device 3. The memory 31 is used to store an operating system, an application program, a boot loader (BootLoader), data, and other programs, such as the program code of the computer program. The memory 31 may also be used to temporarily store data that has been output or is to be output.

[0094] In addition, an embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in any of the above method embodiments are implemented.

[0095] An embodiment of the present application provides a computer program product. When the computer program product is run on a computer device, the computer device implements the steps in the above-mentioned method embodiments when executing the computer device.

[0096] In several embodiments provided in the present application, it is understood that each box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, a program segment or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved.

[0097] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, and other media that can store program codes.

[0098] The specific embodiments described above further describe the purpose, technical solutions and beneficial effects of the present application in detail. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the scope of protection of the present application. It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of protection of the present application.

Claims

1. A method for real-time detection of cognitive status of attention glasses, characterized in that: include: Acquire multi-channel EEG data of the attention glasses, and acquire multi-dimensional features corresponding to the multi-channel EEG data; Performing time domain transformation according to the multidimensional features to obtain signal components and obtain state representations corresponding to the signal components; performing hierarchical processing on the state representations to determine feature levels; Label cognitive attributes based on the feature hierarchy; construct a state map based on the cognitive attributes, and generate a detection sequence based on the state map; Performing time series analysis on the detection sequence, extracting change features, marking state nodes based on the change features, constructing a cognitive sequence using the state nodes, and generating a state description according to the cognitive sequence; Performing trend analysis on the state description, performing feature extraction on the analysis result corresponding to the trend analysis, constructing conversion rules according to the extracted features, and generating a prediction sequence according to the conversion rules; Performing threshold division on the prediction sequence, marking state intervals, performing feature aggregation according to the state intervals, constructing a criterion sequence using aggregation results corresponding to the feature aggregation, and generating an evaluation result according to the criterion sequence; performing signal detection on the evaluation result to obtain detection information; Performing data conversion on the detection information, constructing a display sequence, organizing states based on the display sequence, and marking state types according to the organization results of the state organization; Cognitive state information is generated according to the state type.

2. The method according to claim 1, characterized in that The step of obtaining and generating multi-dimensional features corresponding to the multi-channel EEG data includes: Establishing a data cache of the multi-channel EEG data; Performing window division based on the data cache, generating a real-time sequence according to the division result of the window division; performing feature marking on the real-time sequence, and extracting a state indicator; A feature set is constructed based on the state indicator, and a multi-dimensional feature is generated according to the feature set.

3. The method according to claim 2, characterized in that The step of establishing a data cache of the multi-channel EEG data comprises: A double buffer mechanism is used for cache management to calculate the kurtosis and skewness of the divided multi-channel EEG data to evaluate the data quality.

4. The method according to claim 2, characterized in that: The step of marking the real-time sequence with features and extracting status indicators includes: Extract the signal amplitude variation characteristics of multi-channel EEG data in time domain analysis, decompose the signal energy distribution characteristics of multi-channel EEG data in frequency domain analysis, obtain the signal dynamic evolution characteristics of multi-channel EEG data in time-frequency domain joint analysis, and extract nonlinear dynamic characteristics; The state index is constructed according to the signal amplitude change characteristics, signal energy distribution characteristics, signal dynamic evolution characteristics and nonlinear dynamic characteristics.

5. The method according to claim 1, characterized in that The state nodes include slope change points, fluctuation abnormal points and fluctuation nodes; the state node marking based on the change characteristics includes: Perform piecewise linear fitting on the trend characteristics corresponding to the change characteristics and identify the slope change points; Adaptive threshold method is used to detect fluctuation abnormal points for fluctuation features corresponding to the change features; Calculate the local fluctuation intensity corresponding to the change characteristics and mark the fluctuation nodes that exceed the standard deviation of the historical mean.

6. The method according to claim 1, characterized in that The generating a state description according to the cognitive sequence comprises: Obtain the time intervals between adjacent states corresponding to the cognitive sequence to generate a basic description; Extracting the state transition path rule from the basic description to generate a behavior pattern description; In combination with the importance weight, the high-weight state transitions are filtered in the pattern description to generate a priority description; The detail level of the pattern description is dynamically adjusted according to the application scenario requirements to generate a state description including time series statistics and state evolution process.

7. The method according to claim 1, characterized in that The obtaining of the state representation corresponding to the signal component includes: performing frequency domain decomposition on the signal components to mark feature intervals through the frequency domain decomposition; A state representation is generated based on the feature interval.

8. The method according to claim 1, characterized in that The performing signal detection on the evaluation result to obtain detection information includes: Classifying and marking the detection results corresponding to the signal detection, and identifying abnormal features based on the classification marks; The detection information is generated according to the abnormal features.

9. A real-time cognitive status detection device, characterized in that: include: A data acquisition unit, used to acquire multi-channel EEG data of the attention glasses and to acquire multi-dimensional features corresponding to the multi-channel EEG data; Performing time domain transformation according to the multidimensional features to obtain signal components and obtain state representations corresponding to the signal components; performing hierarchical processing on the state representations to determine feature levels; Label cognitive attributes based on the feature hierarchy; construct a state map based on the cognitive attributes, and generate a detection sequence based on the state map; A timing analysis unit, configured to perform timing analysis on the detection sequence, extract change features, mark state nodes based on the change features, construct a cognitive sequence using the state nodes, and generate a state description according to the cognitive sequence; Performing trend analysis on the state description, performing feature extraction on the analysis result corresponding to the trend analysis, constructing conversion rules according to the extracted features, and generating a prediction sequence according to the conversion rules; A threshold division unit is used to perform threshold division on the prediction sequence, mark state intervals, perform feature aggregation according to the state intervals, construct a criterion sequence using the aggregation results corresponding to the feature aggregation, generate an evaluation result according to the criterion sequence; perform signal detection on the evaluation result to obtain detection information; A data conversion unit, configured to perform data conversion on the detection information, construct a display sequence, organize states based on the display sequence, and mark state types according to the organization results of the state organization; Cognitive state information is generated according to the state type.

10. A computer device, characterized in that: The method comprises a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program and implement the method according to any one of claims 1 to 8 when executing the computer program.

Images