Intelligent glasses electroencephalogram signal noise reduction and feature extraction method, device and equipment

Abstract

The invention discloses an electroencephalogram signal noise reduction and feature extraction method, device and equipment for intelligent glasses, and the method comprises the steps: collecting a multi-channel electroencephalogram signal of the intelligent glasses, obtaining an original data flow, and generating a standard data flow according to the original data flow; performing sub-band decomposition on the standard data stream to obtain frequency band data, and obtaining a noise reduction signal corresponding to the frequency band data; performing edge detection on the noise reduction signal to obtain a feature position, and obtaining a contour feature corresponding to the feature position; obtaining a feature template corresponding to the contour feature and a corresponding signal type, and constructing a signal sample according to the signal type; obtaining a frequency component corresponding to the signal sample, and performing reconstruction according to the frequency component to generate a pure waveform; obtaining a feature index corresponding to the pure waveform, generating feature mapping according to the feature index, and generating a feature matrix according to the feature mapping; and performing dimension mapping on the feature matrix to obtain a main component, outputting feature representation according to the main component, and completing noise reduction and feature extraction of the multichannel electroencephalogram signal.

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 noise reduction and feature extraction of EEG signals of smart glasses. Background Art

[0002] As a portable EEG acquisition device, the signal quality of smart glasses is easily affected by multi-source noise interference and user behavior. Traditional signal processing methods have the following limitations: First, the commonly used frequency domain filtering and statistical noise reduction techniques are difficult to effectively process the non-stationary signals collected by smart glasses in dynamic scenes, which easily leads to the loss of effective features; second, the anomaly detection method based on fixed thresholds cannot adapt to individual differences and environmental changes, reducing the noise reduction effect; third, the existing feature extraction methods mostly rely on preset feature templates, which are difficult to capture dynamic feature patterns under different cognitive tasks.

[0003] In practical applications, the EEG signals collected by smart glasses still face the following challenges: first, the signal quality between different channels varies greatly, and it is necessary to establish a unified quality assessment and processing standard; second, the signal characteristics have obvious temporal dependencies, and the time evolution characteristics of the characteristics need to be considered; third, the impact of environmental noise is complex and changeable, and it is necessary to develop adaptive noise identification and suppression methods. In addition, the portability of smart glasses requires the signal processing algorithm to have high computational efficiency.

[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 noise reduction and feature extraction of EEG signals for smart glasses, aiming to solve the following challenges in practical applications of EEG signals collected by smart glasses: first, the signal quality between different channels varies greatly, and it is necessary to establish a unified quality assessment and processing standard; second, the signal features have obvious temporal dependencies, and it is necessary to consider the time evolution characteristics of the features; third, the influence of environmental noise is complex and changeable, and it is necessary to develop an adaptive noise identification and suppression method. In addition, the portability of smart glasses requires the signal processing algorithm to have a high computational efficiency.

[0006] In a first aspect, the present application provides a method for denoising and extracting features of EEG signals in smart glasses, comprising:

[0007] Collecting multi-channel EEG signals of the smart glasses to obtain an original data stream, generating an initial feature set according to the original data stream, and generating a standard data stream according to the initial feature set;

[0008] Performing sub-band decomposition on the standard data stream to obtain frequency band data, obtaining energy features corresponding to the frequency band data, and obtaining a noise reduction signal according to the energy features; performing edge detection on the noise reduction signal to obtain a feature position, and obtaining a contour feature corresponding to the feature position;

[0009] Acquire a feature template and a corresponding signal type corresponding to the contour feature, and construct a signal sample according to the signal type; acquire a frequency component corresponding to the signal sample, and reconstruct according to the frequency component to generate a pure waveform;

[0010] Acquire characteristic indicators corresponding to the pure waveform, generate characteristic maps according to the characteristic indicators, and generate characteristic matrices according to the characteristic maps;

[0011] The feature matrix is ​​dimensionally mapped to obtain the main components, and feature representation is output according to the main components to complete noise reduction and feature extraction of the multi-channel EEG signal.

[0012] In a second aspect, the present application also provides a smart glasses EEG signal noise reduction and feature extraction device, comprising:

[0013] A signal acquisition module, used to collect multi-channel EEG signals of the smart glasses, obtain an original data stream, generate an initial feature set according to the original data stream, and generate a standard data stream according to the initial feature set;

[0014] A feature acquisition module is used to perform sub-band decomposition on the standard data stream to obtain frequency band data, obtain energy features corresponding to the frequency band data, and obtain a noise reduction signal according to the energy features; perform edge detection on the noise reduction signal to obtain feature positions, and obtain contour features corresponding to the feature positions;

[0015] A type acquisition module is used to acquire a feature template and a corresponding signal type corresponding to the contour feature, and construct a signal sample according to the signal type; acquire a frequency component corresponding to the signal sample, and reconstruct the signal sample according to the frequency component to generate a pure waveform;

[0016] An index acquisition module, used for acquiring characteristic indexes corresponding to the pure waveform, generating a characteristic map according to the characteristic indexes, and generating a characteristic matrix according to the characteristic map;

[0017] The extraction completion module is used to perform dimension mapping on the feature matrix to obtain the main components, output feature representations based on the main components, and complete noise reduction and feature extraction of the multi-channel EEG signal.

[0018] 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 method for denoising and extracting features of smart glasses EEG signals as described in the first aspect is implemented.

[0019] In a fourth aspect, the present application further provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the smart glasses EEG signal denoising and feature extraction method as described in the first aspect.

[0020] Compared with the prior art, this application has at least the following beneficial effects:

[0021] 1. Through hierarchical feature extraction and processing strategies, the method realizes adaptive processing of the signals collected by smart glasses. First, the initial feature set is constructed based on multi-channel signal features, including time domain statistical features, frequency band energy features and spatial correlation features; then, through anomaly detection and data correction based on probability density function, a dynamic threshold mechanism is established to achieve adaptation to individual differences; finally, multi-level sub-band decomposition and spectrum analysis are used to extract stable signal features, effectively solving the problem of signal instability in portable EEG acquisition.

[0022] 2. The method based on template matching and multi-dimensional feature analysis is used to achieve accurate recognition of different cognitive states. The method first uses multi-scale derivative operators to perform edge detection and accurately locate the key feature positions; then combines hierarchical clustering and spatial consistency constraints to establish a cross-channel feature segment alignment mechanism; further, through the improved wavelet packet reconstruction algorithm, the selective retention of different cognitive state features is achieved; finally, based on spectral analysis, the feature expression is optimized, which significantly improves the recognition accuracy of the user's cognitive state in complex environments.

[0023] 3. A hierarchical feature mapping and dimensionality reduction framework was established to achieve efficient expression of features. The method uses continuous wavelet transform technology to construct a multi-scale decomposition system to achieve a fine characterization of signal features; a low-dimensional representation of features is established based on the t-SNE algorithm and hierarchical mapping structure; a multi-level quality assessment mechanism, including reconstruction error analysis, temporal feature retention, and spatial feature verification, ensures the retention of key information in the dimensionality reduction process; and finally forms a feature representation method that retains cognitive state discrimination information and has high computational efficiency.

[0024] 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

[0025] Figure 1A schematic diagram of a process for noise reduction and feature extraction of EEG signals in smart glasses according to an embodiment of the present application;

[0026] Figure 2 This is a schematic diagram of the structure of a device for denoising and extracting features of EEG signals in smart glasses according to an embodiment of the present application;

[0027] Figure 3 A schematic diagram of the structure of a computer device shown in an embodiment of the present application. DETAILED DESCRIPTION

[0028] 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.

[0029] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0030] 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.

[0031] 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.

[0032] 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.

[0033] 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.

[0034] The technical solution of the embodiment of the present application is introduced below.

[0035] As a portable EEG acquisition device, the signal quality of smart glasses is easily affected by multi-source noise interference and user behavior. Traditional signal processing methods have the following limitations: First, the commonly used frequency domain filtering and statistical noise reduction techniques are difficult to effectively process the non-stationary signals collected by smart glasses in dynamic scenes, which easily leads to the loss of effective features; second, the anomaly detection method based on fixed thresholds cannot adapt to individual differences and environmental changes, reducing the noise reduction effect; third, the existing feature extraction methods mostly rely on preset feature templates, which are difficult to capture dynamic feature patterns under different cognitive tasks.

[0036] In practical applications, the EEG signals collected by smart glasses still face the following challenges: First, the signal quality between different channels varies greatly, and it is necessary to establish a unified quality assessment and processing standard; second, the signal features have obvious temporal dependencies, and the time evolution characteristics of the features need to be considered; third, the impact of environmental noise is complex and changeable, and it is necessary to develop adaptive noise identification and suppression methods. In addition, the portability of smart glasses requires that the signal processing algorithm has high computational efficiency. Therefore, it is necessary to develop a signal processing solution that adapts to the application characteristics of smart glasses to achieve full process optimization from signal acquisition, noise reduction to feature extraction.

[0037] To solve the above problems, please refer to Figure 1 , Figure 1 The flowchart of a method for denoising and extracting features of EEG signals in smart glasses provided in an embodiment of the present application is shown in FIG. The method for denoising and extracting features of EEG signals in smart 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 method for denoising and extracting features of smart glasses EEG signals in this embodiment includes steps S101 to S105, which are described in detail as follows:

[0038] Step S101, collecting multi-channel EEG signals of the smart glasses to obtain an original data stream, generating an initial feature set according to the original data stream, and generating a standard data stream according to the initial feature set.

[0039] Specifically, the multi-channel EEG signals of the smart glasses are collected. The smart glasses have an eight-channel dry electrode array built in, in which the Fp1 and Fp2 electrodes in the frontal lobe area are used to collect EEG activities related to the frontal lobe executive control, the T3 and T4 electrodes in the temporal lobe area collect EEG signals related to auditory cognition, the O1 and O2 electrodes in the occipital lobe area collect EEG data related to visual processing, and the P3 and P4 electrodes in the parietal lobe area collect EEG activities related to spatial attention. The sampling frequency is set to 1000HZ to capture the rapid changes of EEG signals, and the sampling accuracy is 24 bits to ensure the accurate collection of weak signals. For example, in an office scene, when the user concentrates on reading a document, the P3 and P4 electrodes record the spatial attention allocation state, the O1 and O2 electrodes record the visual processing process, and the Fp1 and Fp2 electrodes record the cognitive load level. The method monitors the contact impedance value between the electrode and the skin in real time. When the channel impedance is detected to exceed five kiloohms, a quality mark is added to the channel. For example, in a remote conference scenario, the T3 and T4 electrodes focus on collecting EEG changes during voice interaction, the Fp1 and Fp2 electrodes record the state of attention adjustment, and the O1 and O2 electrodes monitor changes in visual concentration. The EEG signals collected by each channel are synchronized in real time and aligned, and together with the channel quality mark information, form an original data stream, which contains the multi-channel EEG time series characteristics of users in different cognitive states.

[0040] In some embodiments, generating the initial feature set according to the original data stream includes: dividing the original data stream into time windows to obtain time series segments; and generating the initial feature set according to the time series segments.

[0041] Since the original data stream contains long-term signals collected continuously, appropriate segmentation processing is required to extract local cognitive state features. At the same time, the signal amplitude ranges of different channels are different, and unified amplitude mapping is required for subsequent analysis. Therefore, the method performs time window segmentation and normalization processing on the original data stream. Based on the temporal characteristics of human cognitive processing, the window length is set to 2 seconds to include the complete cognitive activity process, and each window contains 2000 sampling points; the window sliding step is 1 second, and adjacent windows overlap 1000 sampling points to maintain the continuity of signal analysis. For example, in the driving scene, when the driver looks at the road ahead, the O1 and O2 electrodes record continuous visual monitoring activities, the P3 and P4 electrodes collect spatial alertness, and the Fp1 and Fp2 electrodes monitor cognitive fatigue levels; when the driver looks at the rearview mirror, the O1, O2 and P3, P4 electrodes jointly record the visual attention conversion process, and the T3 and T4 electrodes collect spatial perception states. The method marks the specific driving behavior type for each time window, including forward gaze, rearview mirror viewing, instrument panel observation, etc. For example, in a walking scene, window segmentation can separate and record cognitive processes such as user movement imagination, spatial positioning, and environmental perception. P3 and P4 electrodes record EEG activities related to spatial navigation, O1 and O2 electrodes collect environmental visual search processes, T3 and T4 electrodes monitor sound positioning status, and Fp1 and Fp2 electrodes record execution control levels. Amplitude normalization is performed on the eight-channel signals in all segmentation windows, and the signals are uniformly mapped to the interval from negative one to positive one to generate a standardized time sequence segment group.

[0042] The signals in the standardized time sequence segment group have a unified time scale and amplitude range, which enables the method to use a consistent feature extraction method to process the data of all channels. Based on this, the method further constructs an initial feature set containing multi-dimensional features. First, the basic time domain statistical features are extracted from the signals in each window. The signal average is calculated to represent the baseline level, the variance reflects the signal fluctuation intensity, the waveform length represents the signal complexity, and the zero-crossing rate reflects the signal oscillation frequency. For example, in the reading scene, the 4-7HZ theta frequency band energy ratio is extracted from the Fp1 and Fp2 channels to represent the working memory load, the 8-13HZ alpha frequency band energy ratio represents the attention inhibition level, and the 14-30HZ beta frequency band energy ratio represents the cognitive activity intensity; by calculating these frequency band energy features, the changes in the user's cognitive state during the reading process can be quantified. For example, in the video viewing scene, the alpha frequency band energy changes of the O1 and O2 channels are mainly extracted to represent the visual attention intensity, the theta / alpha energy ratio of the P3 and P4 channels is analyzed to represent the attention allocation efficiency, and the beta frequency band energy of the T3 and T4 channels is calculated to reflect the auditory processing intensity. The cross-correlation matrix of each channel signal is calculated to record the phase synchronization characteristics between the symmetrical channels of the left and right brain, the information transmission characteristics of the front and back brain regions, and the local connectivity characteristics between adjacent electrodes. For example, in the motion scene, the alpha band power asymmetry index of the Fp1 and Fp2 channels is extracted to represent the attention bias, the gamma band energy changes of the T3 and T4 channels are analyzed to reflect the spatial perception intensity, and the beta / theta energy ratio of the P3 and P4 channels is calculated to evaluate the motor control level. All the extracted feature parameters are normalized and assigned unique identifiers, and the time domain features, frequency domain features, and spatial features are organized into a structured feature array, which finally forms an initial feature set that represents the multi-dimensional cognitive state of the user.

[0043] In some embodiments, generating a standard data stream based on the initial feature set includes: performing amplitude statistics on the initial feature set to obtain an amplitude distribution; obtaining an abnormal point position corresponding to the initial feature set based on the amplitude distribution; and correcting the initial feature set based on the abnormal point position to generate the standard data stream.

[0044] Based on the initial feature set formed in step S101, the method first performs amplitude statistical analysis on the multi-channel EEG features. The statistical distribution parameters of the time domain feature values ​​of each channel in the initial feature set, including zero-crossing rate, waveform length, root mean square value, etc., are calculated. For example, in the classroom concentration detection scenario, the method extracts the maximum, minimum, mean, standard deviation, skewness and kurtosis of the Fp1 and Fp2 channel features, where the mean reflects the baseline level of attention, the standard deviation represents the fluctuation range, the skewness describes the asymmetry of attention distribution, and the kurtosis represents the steepness of attention fluctuation. For the visual evoked potential characteristics of the O1 and O2 channels, the amplitude distribution of key components such as the P100 peak value, the N170 trough value, and the P300 peak value are respectively counted. At the same time, the probability distribution of the early component feature values ​​such as N100 and P200 related to auditory cognition in the T3 and T4 channels is analyzed. For the frequency domain characteristics, the method performs hierarchical statistics on the energy characteristics of theta, alpha, beta and other frequency bands, and establishes a distribution model of the frequency band energy ratio. For example, in the scenario of intelligent care, the method establishes a probability density function for the attenuation characteristics of the alpha frequency band energy in the frontal lobe of the elderly to characterize the statistical laws of cognitive ability. For spatial features, the distribution characteristics of the correlation coefficient between channels are analyzed, and a probability model of spatial connection strength is established. The method organizes all statistics into a multidimensional matrix, the row vectors of the matrix represent different electrode channels, and the column vectors contain various statistical parameters. At the same time, the probability density function of each channel is constructed based on the time window, and finally a complete amplitude distribution feature space containing time domain, frequency domain and spatial domain features is obtained.

[0045] For the obtained amplitude distribution feature space, the method adopts a multi-level detection strategy to locate abnormal points. First, the adaptive detection threshold is calculated based on the probability density function of each channel. The threshold parameter is determined by the interquartile range and variance to achieve sensitive detection of abnormalities of different scales. For example, in the driving fatigue monitoring scenario, the method detects the abnormal offset of the alpha band energy distribution of the Fp1 and Fp2 channels, calculates the Mahalanobis distance between the current feature and the standard distribution, and marks the points that deviate more than three times the standard deviation as fatigue risk points. For the visual evoked features of the O1 and O2 channels, a dynamic threshold based on the amplitude probability distribution is set to identify abnormal fluctuations caused by distraction. Secondly, the method establishes a joint probability model between channels to analyze the spatial correlation of feature distribution. For example, in the emotion recognition scenario, the asymmetric distribution anomaly caused by emotion fluctuations is identified by calculating the correlation coefficient matrix of the symmetrical channels of the left and right brains. For time series features, a sliding window is used to calculate the local probability density function, and features that deviate significantly from the historical distribution are marked as time series anomalies. The method also constructs an anomaly detection model based on high-order statistics to analyze the high-order moments such as skewness and kurtosis of feature distribution to capture subtle distribution anomalies. For example, in the motor imagery scenario, the abnormal points in the intention conversion process can be accurately located by calculating the high-order statistical features of the μ rhythm energy distribution of the C3 and C4 channels. Finally, all the detection results are integrated into an abnormal point location table, which records the timestamp, channel number, and abnormality degree of each abnormal point.

[0046] Signal correction is performed based on the location information and abnormality degree recorded in the abnormal point location table. The method determines the correction strategy according to the spatial distribution characteristics of the abnormal points. For points marked as single-channel abnormalities, the average value of adjacent time windows is used for correction; for points with simultaneous abnormalities in multiple channels, data is reconstructed based on the spatial correlation information in the abnormal point location table. For example, in a multi-person collaborative scenario, when it is detected that the Fp1 and Fp2 channels of a user simultaneously have abnormalities marked in the location table, the method combines the abnormality degree scores of the two channels for joint correction. The correction amplitude is determined by the abnormality metric value recorded in the location table. The larger the abnormality metric value, the stronger the corresponding correction strength. For example, in the gait monitoring scenario, the method analyzes the occurrence pattern of the periodic abnormal points marked in the location table and selects a suitable correction algorithm according to the time series pattern of the abnormal characteristics. For each corrected data point, the method retains the feature mapping relationship in the abnormal point location table to ensure the continuity and stability of the signal. During the signal correction process, the method also refers to the channel correlation information recorded in the location table, adopts a consistent correction strategy for the associated channels, and maintains the phase relationship between the channels. After correction processing based on the abnormal point location table, a standard data stream is finally generated.

[0047] Step S102, performing sub-band decomposition on the standard data stream to obtain frequency band data, obtaining energy features corresponding to the frequency band data, and obtaining a noise reduction signal according to the energy features; performing edge detection on the noise reduction signal to obtain feature positions, and obtaining contour features corresponding to the feature positions.

[0048] Specifically, the standard data stream obtained in step S101 is subjected to sub-band decomposition. The method uses a 6th-order Butterworth filter to construct a multi-stage filter bank, and decomposes the EEG signal into delta band (0.5 to 4 Hz), theta band (4 to 8 Hz), alpha band (8 to 13 Hz), beta band (13 to 30 Hz) and gamma band (30 to 50 Hz). To ensure the accuracy of filtering, a transition bandwidth of 0.5 Hz is set in each band, and the stopband attenuation is set to 60 decibels. For example, in the concentration detection scenario, the method performs sub-band decomposition on the Fp1 and Fp2 channel signals in the standard data stream to extract the theta band activity related to the cognitive load of the prefrontal lobe; at the same time, the correlation coefficient of the signal before and after passing through the filter is calculated to ensure signal integrity. When decomposing the O1 and O2 channel signals, the alpha band is focused on to extract the rhythmic activity in the visual processing process, and the decomposition effect is verified by calculating the frequency band energy ratio. In the decomposition of T3 and T4 channel signals, the gamma band is emphasized to obtain the high-frequency characteristics of auditory information processing. The sub-band decomposition adopts zero-phase filtering technology, eliminates phase distortion through forward and reverse filtering, and maintains the timing characteristics of the signal. For example, in the visual evoked scene, the method uses a sliding window decomposition for the P3 and P4 channel signals, with the window length set to 1 second and the sliding step length set to 0.5 seconds, to extract the instantaneous characteristics of each frequency band respectively. In the motion imagery scene, an adaptive filter group is introduced into the signals of the C3 and C4 channels, and the filter parameters are dynamically adjusted according to the spectral characteristics of the signal. The method calculates the signal-to-noise ratio of each frequency band signal after decomposition, and marks the frequency band with a signal-to-noise ratio lower than 10 decibels as the object to be processed. Finally, a multi-channel and multi-band frequency band data matrix is ​​obtained, and each element of the matrix contains the time domain sequence of the corresponding channel and the corresponding frequency band.

[0049] In some embodiments, obtaining the noise reduction signal according to the energy feature includes: performing threshold selection on the energy feature according to a preset threshold to obtain a valid signal corresponding to the energy feature; obtaining a reconstructed waveform corresponding to the valid signal; and obtaining the noise reduction signal according to the reconstructed waveform.

[0050] For the obtained frequency band data matrix, the method conducts multi-scale energy analysis. First, the short-time Fourier transform is used to calculate the instantaneous energy of each frequency band. The window length is set to 0.5 seconds, the sliding step is 0.25 seconds, and the Hanning window is used as the windowing function to reduce spectral leakage. For example, in the motor imagery scene, the method analyzes the energy change characteristics of the C3 and C4 channels in the 8 to 13 Hz frequency band, and evaluates the intensity of motor imagery by calculating the energy difference between the left and right motor cortices. When calculating the theta band energy of the Fz channel, the multi-resolution analysis method is used to extract the energy change pattern in the attention regulation process at different time scales. The method establishes an energy envelope curve for each frequency band, extracts the energy modulation characteristics through the Hilbert transform, and calculates the first-order derivative of the envelope to reflect the energy change rate. For example, in the working memory task, the theta band energy envelope of the Fp1 and Fp2 channels is decomposed by wavelet to extract the energy fluctuation characteristics at different scales; the phase synchronization of the alpha band energy envelope of the P3 and P4 channels is calculated to evaluate the spatiotemporal coordination of information processing. In the emotion recognition scenario, the asymmetry of the alpha band energy of the F3 and F4 channels is analyzed to construct an energy feature vector reflecting the emotional activity of the prefrontal lobe. The method also establishes an energy coupling model between frequency bands, and analyzes the energy interaction characteristics across frequency bands by calculating the cross-correlation function of the energy of different frequency bands. For example, in the attention switching task, the theta-gamma coupling intensity and the alpha-beta energy correlation are analyzed simultaneously to construct multi-band energy synergy characteristics. The method finally generates a multi-dimensional energy feature matrix containing the energy distribution, energy ratio, energy modulation characteristics, and frequency band coupling characteristics of each frequency band.

[0051] Adaptive threshold selection based on energy feature matrix. The improved OTSU algorithm is used to calculate the energy threshold of each frequency band, and the signal segments with energy significantly higher than the background are marked as active. The improvements include introducing inter-channel correlation weights and time continuity constraints to improve the stability of threshold selection. For example, in the visual attention task, the method sets a dynamic threshold for the alpha band energy of the O1 and O2 channels. When the energy exceeds the threshold, the corresponding time window is marked as the active period of visual processing; at the same time, the energy distribution of the P3 and P4 channels is considered, and a threshold adjustment mechanism based on spatial correlation is established. For the case of simultaneous activity of multiple channels, the method establishes a joint threshold criterion based on inter-channel energy correlation, and optimizes the threshold parameters by calculating the energy covariance matrix between channels. For example, in the motion control scenario, the beta band energy of the C3, C4 and FCz channels is monitored simultaneously, and the energy principal components of the three channels are calculated, and the joint threshold is set according to the principal component score. For each signal segment that passes the threshold test, its start and end time, channel, band type, energy intensity and spatial distribution characteristics are recorded to form a valid signal table. At the same time, the corresponding original band information in the band data matrix is ​​retained for subsequent signal reconstruction.

[0052] Signal reconstruction was performed using the spatiotemporal marking information of the effective signal table and the retained frequency band data. Methods First, amplitude normalization was performed on the frequency band data marked in the effective signal table, and the energy difference between different frequency bands was eliminated by piecewise linear mapping. The normalization parameters were optimized according to the physiological significance of each frequency band, such as enhancing the weight of cognitive-related frequency bands and suppressing the contribution of myoelectric interference frequency bands. Then, the signal of each frequency band was reconstructed into a complete waveform by weighted superposition, and the weight coefficient was determined by the energy contribution of each frequency band. For example, in the cognitive load assessment scenario, when reconstructing the signal of the Fz channel, the method assigned higher weight coefficients to the theta band and the beta band to highlight the EEG characteristics related to working memory. In the spatial attention analysis, the spatial characteristics of the alpha band of the P3 and P4 channels were retained, and the spatial consistency of the reconstructed signal was ensured by phase correction. Methods An improved wavelet packet reconstruction algorithm was used to selectively reconstruct each frequency band according to the guidance of the effective signal table. For example, in visual evoked potential analysis, the time domain characteristics of the alpha and gamma bands are retained when reconstructing the O1 and O2 channel signals, and the reconstruction quality is evaluated by calculating the cross-correlation function before and after reconstruction, and finally a noise-reduced signal is generated.

[0053] In some embodiments, obtaining the contour feature corresponding to the characteristic position includes: obtaining the signal boundary corresponding to the noise reduction signal according to the characteristic position; dividing the noise reduction signal according to the signal boundary to obtain characteristic segments; and performing feature extraction on the characteristic segments to obtain the contour feature.

[0054] Edge detection is performed on the denoised signal. The method uses a multi-scale derivative operator to calculate the first-order and second-order derivatives of the signal and establishes an edge detection factor library. When calculating the derivative, Gaussian smoothing kernels of different scales are used, and the scale parameters are set to 2 milliseconds, 5 milliseconds, and 10 milliseconds, respectively, to capture the signal change characteristics of different time spans. For example, in the analysis of visual evoked potentials, the local maximum gradient position of the denoised signals of the O1 and O2 channels is calculated to identify the rapidly changing edges of components such as P100 and N170. At the same time, an adaptive threshold mechanism is introduced, and the threshold parameter is determined by the local standard deviation and mean of the signal. For example, in the motor imagery task, the method extracts the position information of the peaks and troughs of the denoised signals of the C3 and C4 channels, and marks the edge points where the motor imagery starts and ends. The method calculates the significance score for each detected edge point, which is determined by the gradient amplitude and local contrast. In the attention switching task, the edge position of the energy change of the alpha band is detected, and the feature position sequence containing edge position, edge strength, and edge type is formed through weighted fusion of multi-scale edge detection results.

[0055] The signal boundary is marked based on the obtained feature position sequence. The method clusters the edge points in the feature position sequence and uses a hierarchical clustering method based on time distance to classify edge points with similar time as the same boundary area. The clustering threshold is dynamically adjusted according to the sampling rate of the signal. For signals with a sampling rate of 1000 Hz, edge points with a time interval of less than 20 milliseconds are classified into the same category. For example, in the attention switching task, the method analyzes the boundary features related to attention changes in the Fp1 and Fp2 channels, and determines the start and end positions of the boundary by calculating the time distribution density of the edge points. For cross-channel boundary marking, a spatial consistency constraint is introduced, requiring that the boundary position deviation of adjacent channels does not exceed the set threshold. For example, in the emotion recognition scenario, the method marks the boundaries of the emotion-related waveforms of the F3 and F4 channels and establishes the corresponding relationship between the left and right brain regions. The method also calculates the stability index of each boundary, which is determined by the density and intensity distribution of the edge points in the boundary area. Finally, a signal boundary table is generated, which records the time position, duration, channel to which it belongs, and boundary intensity information of each boundary.

[0056] Feature segments are divided according to the signal boundary table and the noise reduction signal. The method first verifies the validity of the signal boundary of each channel and removes the false boundaries caused by transient noise. For example, in the working memory task, the method divides the feature segments of the three stages of encoding, retention and extraction according to the boundary information of the theta rhythm of the Fz channel. For cross-channel collaborative activities, a segment alignment mechanism based on boundary synchronization is established. For example, in the visual search task, the method synchronously divides the signal segments of P3, P4 and O1, O2 channels to ensure the temporal correspondence between spatial attention and visual processing features. In the emotion-inducing scene, the method grades the segments according to the boundary strength, and the boundaries with strength above 75% are divided into key analysis segments. The method establishes an index for the divided feature segments to form a feature segment set containing the start and end time of the segment, the segment type, and the channel distribution.

[0057] Feature extraction is performed based on the time attributes and channel distribution information recorded in the feature segment set. Different extraction strategies are adopted for different types of feature segments: for segments marked as the encoding stage, the energy characteristics of the theta frequency band are extracted; for segments marked as the holding stage, the inhibition characteristics of the alpha frequency band are analyzed; for segments marked as the extraction stage, the activity characteristics of the beta frequency band are focused. In the time domain feature layer, the kurtosis, skewness, zero-crossing rate and other statistics of the signal are calculated based on the duration information of the segment; in the frequency domain feature layer, the spectral characteristics such as power spectrum density and band energy ratio are extracted based on the channel attributes of the segment; in the morphological feature layer, the morphological parameters such as symmetry, smoothness and complexity of the waveform are analyzed based on the boundary characteristics of the segment. For example, in the remote conference scenario, the theta / beta energy ratio of the Fp1 and Fp2 channel segments marked as the focus state in the feature segment set is analyzed to extract the characteristic indicators reflecting cognitive engagement. In the document reading task, for the visual processing segments of the O1 and O2 channels in the feature segment set, the P100-N170-P300 waveform sequence features are extracted based on their time span. The method also establishes a combined extraction mechanism of multi-channel features based on the spatial distribution information recorded in the segment set. For example, in the attention switching task, the alpha band energy changes of the frontal and parietal segments are analyzed synchronously to extract the collaborative features of the spatial attention network. Finally, a contour feature set containing time domain, frequency domain and morphological features is generated.

[0058] Step S103, obtaining a feature template and a corresponding signal type corresponding to the contour feature, and constructing a signal sample according to the signal type; obtaining a frequency component corresponding to the signal sample, and reconstructing according to the frequency component to generate a pure waveform.

[0059] Specifically, the contour features obtained in step S102 are subjected to pattern matching. The method first constructs a feature template library, which contains standard contour features of typical cognitive activities such as visual reading, conference conversation, and daily office work. The visual evoked potential template records the standard latency and amplitude thresholds of components such as P100, N170, and P300, where the latency range of P100 is set to 80 to 120 milliseconds, the amplitude threshold is 3 to 8 microvolts, and the latency range of N170 is 150 to 200 milliseconds, and the amplitude threshold is 5 to 10 microvolts. Differentiated matching strategies are set for different cognitive tasks. For example, video watching uses a matching criterion based on latency and amplitude, and the matching score is determined by the weighted sum of time deviation and amplitude deviation. The attention task uses a matching method based on frequency band energy to calculate the similarity of energy contours in the 8 to 13 Hz frequency band. For example, in a teleconference, the method extracts the P300 component from the contour features of the O1 and O2 channels, calculates their latency deviation scores, amplitude deviation scores, and waveform width deviation scores, and uses weights of 0.4, 0.4, and 0.2 to calculate the comprehensive matching score. The matching threshold is dynamically adjusted according to the difficulty of the task. The matching degree is required to reach more than 85% for simple browsing tasks, and it is reduced to 70% for complex document reading tasks. In daily office scenarios, the attention pattern features are calculated for the contour features of the Fp1 and Fp2 channels, including time series parameters such as the time of concentrated attention, the degree of distracted attention, and the frequency of task switching. The method uses a dynamic time warping algorithm to align the features to be matched and the template features, and the maximum allowed time deformation range is 20% of the original length. Finally, a feature template set containing the optimal matching template identifier, matching metric score, time alignment parameter, and deformation coefficient is generated for each contour feature.

[0060] Signal type labeling was performed using matching information in the feature template set. The matching metric score of each contour feature and its optimal matching template was analyzed. When the score exceeded the preset threshold, the cognitive type label corresponding to the template was assigned to the feature. To improve the reliability of labeling, a three-layer evaluation mechanism was used for the matching metric score: the first layer evaluated the time domain matching, including waveform correlation coefficient and root mean square error; the second layer evaluated the frequency domain matching, including power spectrum similarity and phase consistency; and the third layer evaluated the morphological matching, including peak and trough distribution and waveform complexity. For example, in the evaluation of meeting concentration, the method determined that the high matching (score>0.8) feature was highly focused, the medium matching (0.6-0.8) feature was generally focused, and the low matching (<0.6) feature was distracted based on the matching scores of the Fp1 and Fp2 channel features and the attention template. The combined scores of the three matching dimensions were considered at the same time. Only when the time domain matching>0.75, the frequency domain matching>0.7 and the morphological matching>0.8 were confirmed as high reliability labels. Based on the time alignment parameters in the feature template set, the method establishes cross-channel synchronization constraints. For example, in multimedia interaction scenarios, the matching time difference between the visual channel and auditory channel features and their respective templates is analyzed, and the alignment time difference of the associated channels is required to be no more than 50 milliseconds. The method also evaluates the reliability of type marking based on the deformation coefficient. For every 0.1 increase in the deformation coefficient, the reliability of the marking decreases by 10%. In continuous office tasks, the method establishes transfer constraints on the type marking of continuous time windows to limit the mutation probability of the marking type. When the type of adjacent windows changes, sufficient matching evidence is required to support it. Finally, a signal type table containing feature types, three-layer matching reliability and timing relationships is generated.

[0061] Signal samples are constructed based on the matching features of the feature template set and the label information of the signal type table. The method first selects high-quality samples according to the matching metric score, and selects the features with the top 30% comprehensive matching scores as benchmark samples. For each benchmark sample, the weighted feature expression is constructed by combining its type label and three-layer reliability information in the signal type table. The sample weight is determined by the matching metric score and reliability, and the calculation formula is: weight = matching score × (0.4 × time domain reliability + 0.3 × frequency domain reliability + 0.3 × morphological reliability). For example, in the document reading task, the method selects O1 and O2 channel features with high matching (>0.85) with the visual attention template, and organizes them into a continuous attention sample sequence according to the timing information in the signal type table. In the remote conference scenario, a complete attention sample set is established by analyzing the matching stability and type conversion rules of the Fp1 and Fp2 channel features with the concentration template. The sample set contains feature samples of different states such as concentration, distraction, and conversation. Each sample inherits the template matching information, type label attribute and reliability index of its corresponding feature. For cognitive indicators that need to be accurately quantified, this method combines the deformation parameters of the feature template set and the temporal relationship of the type mark to establish a sample scoring mechanism. The scoring formula is: sample score = basic score × (1-deformation coefficient) × temporal consistency, where the basic score is determined by the matching metric and the temporal consistency reflects the logical rationality of the sample sequence. The method implements quality control on the sample set, regularly calculates the representativeness and diversity indicators of the samples, and dynamically updates the sample library based on these indicators. Finally, a standardized signal sample set is formed.

[0062] In some embodiments, the reconstruction based on the frequency components to generate a pure waveform includes: performing bandpass filtering on the frequency components; obtaining frequency components corresponding to the frequency components based on filtering results corresponding to the bandpass filtering; and reconstructing signals based on the frequency components to generate the pure waveform.

[0063] The obtained signal samples were subjected to spectrum analysis. The method used a multi-scale derivative operator to calculate the first-order derivative and the second-order derivative of the signal, and established an edge detection factor library. Gaussian smoothing kernels of different scales were used to calculate the derivatives, and the scale parameters were set to 2 milliseconds, 5 milliseconds and 10 milliseconds respectively to capture the signal change characteristics of different time spans. The first-order derivative was used to detect the rapidly changing area of ​​the signal, and the second-order derivative was used to locate the peak and trough positions. The combination of the two can effectively identify the key change points of the signal. For example, in the analysis of visual evoked potentials, the local maximum gradient position of the O1 and O2 channel signals was calculated to identify the rapidly changing edges of components such as P100 and N170; the method set an adaptive gradient threshold, and the threshold size was determined by the mean and variance of the local 50 millisecond window of the signal. At the same time, a morphological filter was introduced to optimize the edge detection results and eliminate false edges caused by small fluctuations. In the analysis process, short-time Fourier transform was used to calculate the spectral characteristics of the signal, with a window length of 0.5 seconds and a sliding step of 0.25 seconds. The windowing function used the Hanning window to reduce spectrum leakage. The method calculates the power spectrum density estimation for each analysis window and extracts the energy distribution characteristics of different frequency bands. Finally, a frequency component matrix is ​​generated, and each element in the matrix contains information such as frequency value, energy value, phase value and spectrum peak position.

[0064] A bandpass screening strategy is designed based on the energy distribution and spectrum peak information in the frequency component matrix. Methods The key frequency band range is determined according to the peak distribution of the power spectral density, and a bandpass filter is set in the frequency band where the energy is concentrated. For each detected main spectrum peak, a bandpass filter with a center frequency corresponding to the spectrum peak position is designed, and the filter bandwidth is determined according to the energy distribution range of the spectrum peak. For example, in the working memory task, when the frequency component matrix shows that the theta band has a significant energy peak in the range of 4-8 Hz, the method adjusts the bandpass range accordingly to cover the energy concentration area. Methods The energy transition characteristics of adjacent frequency bands in the frequency component matrix are analyzed, and the transition band parameters are designed accordingly. A narrower transition band is used for the junction of frequency bands with steep energy changes. For example, in the concentration training scenario, the passband ripple and stopband attenuation parameters of the filter are dynamically adjusted based on the energy distribution characteristics of the alpha band. In the sleep quality monitoring scenario, the phase characteristics of each frequency band in the frequency component matrix are analyzed, and a zero-phase filter design is used for phase-sensitive frequency bands. The filter performance was evaluated by calculating the signal-to-noise ratio, frequency band energy proportion, out-of-band suppression and other indicators, and a screening result table including filter parameters and performance indicators was generated.

[0065] According to the filter parameters in the screening result table, the spectral components in the frequency component matrix are selectively reorganized. The method first sets the screening threshold of each frequency band based on the filter performance index, focusing on retaining the frequency components that show high signal-to-noise ratio and energy stability in the corresponding frequency band. For example, in the classroom learning scenario, for the theta band of the Fp1 and Fp2 channels, the method extracts the frequency components in the range of 4-8 Hz according to the band parameters in the screening result table, and allocates the reorganization weight according to the performance indicators in the table. In the visual search task, the method screens the spectral components of 8-13 Hz according to the filter characteristics of the alpha band of the O1 and O2 channels, and ensures that the selected frequency components have sufficient signal-to-noise ratio through performance index evaluation. For the cross-band synergy characteristics, the method designs the frequency band combination strategy with reference to the passband parameters in the screening result table. For example, in the attention switching task, the frequency components of the theta and gamma bands are selected simultaneously based on the passband characteristics of the filter, and the phase relationship between them is maintained. According to the screening and selection results, the method generates a detailed frequency component selection table to record the parameter information of each retained frequency component.

[0066] The selected frequency components are reconstructed based on the parameter information in the frequency component selection table. The improved inverse Fourier transform algorithm is used to reassemble the frequency components in the selection table according to their recorded amplitude and phase information. The reconstruction process adopts a segmented processing strategy, with each segment length of 2 seconds, adjacent segments overlapping for 1 second, and smooth transition achieved by weighted averaging. For example, in the speech concentration analysis, the method reconstructs the P3 and P4 channel signals based on the frequency component parameters of the alpha band in the selection table, and uses a cosine smoothing window with a window length of 500 milliseconds at the transition between speaking and listening states. The method dynamically adjusts the reconstruction weight according to the frequency component stability index recorded in the selection table, and gives a greater reconstruction weight to the frequency components with high stability. In the multi-channel EEG feedback task, the method synchronously reconstructs the signals of the Fp1, Fp2 and O1, O2 channels based on the parameter information in the selection table, and ensures the spatial characteristics by maintaining the phase relationship recorded in the table. For example, in the motor imagery analysis, the mu rhythm signal is reconstructed based on the frequency component parameters of the C3 and C4 channels in the selection table, and the phase difference is kept within 20 degrees. The method also established a multi-level quality assessment system. The first level calculated the correlation coefficient between the reconstructed signal and the original sample based on the parameters of the selection table, the second level analyzed the spectral purity of the reconstructed signal, and the third level evaluated the physiological rationality of the reconstructed signal. Finally, a pure waveform that retains key cognitive components is generated.

[0067] Step S104, obtaining characteristic indicators corresponding to the pure waveform, generating a characteristic map according to the characteristic indicators, and generating a characteristic matrix according to the characteristic map.

[0068] Specifically, the pure waveform generated in step S103 is subjected to sub-band division processing. The method adopts continuous wavelet transform technology to construct a multi-scale wavelet decomposition system, and divides the signal into characteristic bands such as delta (0.5-4 Hz), theta (4-8 Hz), alpha (8-13 Hz), beta (13-30 Hz) and gamma (30-50 Hz). In the division process, the complex wavelet basis function is used, and the wavelet scale is progressively set according to multiples of 3, with a total of 12 scales to achieve fine characterization of the frequency band. For example, in the driving fatigue monitoring scenario, the method performs sub-band division on the pure waveform of the Fp1 and Fp2 channels, focusing on the subdivision characteristics of the alpha frequency band, and subdividing the alpha frequency band of 8-13 Hz into two sub-bands of low alpha (8-10 Hz) and high alpha (10-13 Hz), and extracting the energy envelope and phase sequence of each sub-band. For visual-related activities of O1 and O2 channels, the beta frequency band was subdivided into three sub-bands: low beta (13-20 Hz), medium beta (20-25 Hz), and high beta (25-30 Hz), and the instantaneous amplitude and phase changes of each sub-band were calculated simultaneously. For the motor imagery task, the adaptive bandwidth partitioning strategy was adopted for the C3 and C4 channel signals, and the sub-band boundaries were dynamically adjusted according to the peak distribution of the power spectrum to obtain the time-frequency energy distribution of each sub-band. The power correlation coefficients between sub-bands were calculated to construct the frequency band coupling matrix, forming a characteristic band segment set containing the energy distribution, phase sequence, and coupling relationship of each sub-band.

[0069] Feature indicators are extracted based on the energy distribution and phase sequence information of the feature band segment set. In the time domain, statistical moments are calculated for the energy envelope of each sub-band, including mean, standard deviation, skewness and kurtosis; envelope smoothness and phase stability indicators are calculated for the phase sequence. For example, in the emotion recognition task, the method uses the energy distribution characteristics of the alpha sub-band of the F3 and F4 channels to calculate the asymmetry index, which is obtained by the normalized difference of the energy envelope of the left and right hemispheres. In the frequency domain feature layer, the power spectral density is estimated based on the time-frequency energy distribution of the sub-band, and spectral parameters such as center frequency, frequency standard deviation, and main frequency proportion are extracted. For example, in the working memory task, the relative power and absolute power of the energy distribution of the theta sub-band of the Fz channel are calculated, and the cognitive load index is constructed using the energy ratio of the theta and alpha sub-bands. The method also extracts cross-band features based on the frequency band coupling matrix, such as calculating theta-gamma coupling strength and alpha-beta phase synchronization. In the attention detection scenario, the instantaneous phase synchronization is calculated using the phase sequence of the beta sub-band of the P3 and P4 channels to evaluate the spatiotemporal coordination of the attention network. By methodically extracting various types of information from the feature band set, we finally obtain a multidimensional feature index set including time domain indicators, frequency domain indicators and coupling indicators.

[0070] The feature mapping relationship is constructed according to the parameters of the multidimensional feature index set. The method first analyzes the correlation of the feature index set, calculates the Pearson correlation coefficient matrix between the indicators, and selects the indicator subset with a correlation less than 0.7 for mapping. The t-SNE algorithm is used to project the selected high-dimensional feature indicators into the 3D feature space. The perplexity parameter in the mapping process is set to 30, the learning rate is set to 200, and the number of iterations is 1000. For example, in the multimodal learning scenario, the method selects the alpha rhythm indicators of the visual channel (O1, O2), the gamma rhythm indicators of the auditory channel (T3, T4) and the theta rhythm indicators of the executive control channel (Fp1, Fp2), and constructs a unified mapping space based on their time domain statistics and frequency domain features. For dynamic task scenarios, the method uses a sliding window method with a window length of 2 seconds and a step length of 0.5 seconds. The corresponding feature indicators are organized in each window for mapping to form a feature trajectory. In the sleep stage analysis, the energy indicators and phase indicators of multiple channels are used to establish a hierarchical mapping structure, and the characteristic distribution of different sleep stages is reflected by the combination of feature indicators. The method evaluates the topological preservation of the feature map by calculating the local adjacency relationship in the mapping space. When the stress value exceeds 0.2, the mapping parameters are adjusted to reconstruct the feature space. The resulting feature map retains the discriminant information of the original feature index and realizes the low-dimensional expression of the data.

[0071] In some embodiments, generating a feature matrix according to the feature map includes: performing grouping and clustering according to the feature map to obtain feature categories; constructing a feature set according to the feature categories; and generating the feature matrix according to the feature set.

[0072] The obtained feature maps were subjected to group clustering analysis. The method used a hierarchical clustering algorithm. First, the Euclidean distance matrix of the sample points in the feature space was calculated, and then the clusters were gradually merged based on the Ward minimum variance criterion. The number of clustering levels was set dynamically according to the size of the data set. For a data set containing 1000 samples, 4 to 6 levels were set for progressive clustering. For example, in the multi-task learning scenario, the method clustered the cognitive load feature maps of the Fp1 and Fp2 channels, and divided the working memory load into three categories: low load (focused attention), medium load (multitasking), and high load (cognitive overload). The category center and boundary of each category were determined by the distribution characteristics of its feature map. In the spatial attention analysis, the density clustering method was used for the feature maps of the P3 and P4 channels. The main mode of attention allocation was identified by setting the neighborhood radius to 0.5 and the minimum number of samples to 15. The method calculated the intra-class density and inter-class distance based on the clustering results, and formed a feature category table containing category labels, category centers, inter-class distance matrices, and intra-class density distribution.

[0073] The feature set is constructed using the category center and density distribution information in the feature category table. For each category, the method extracts sample points whose distance from the category center is less than the threshold, and determines the sample weight according to the intra-class density distribution. For example, in the emotion recognition task, for the three emotion categories of positive, negative and neutral obtained by clustering the F3 and F4 channels, the 30% sample points with the highest intra-class density are selected as core samples. For cross-channel feature categories, the method establishes spatial associations based on the inter-class distance matrix, and combines categories with similar distances to form a joint feature expression. In cognitive state assessment, the three types of working memory load features of Fp1 and Fp2 are combined with the attention category features of O1 and O2, and a multidimensional feature set is constructed based on intra-class density weighting. Finally, a multimodal feature set containing core samples of each category, weight coefficients and spatial associations is formed.

[0074] Feature combination is performed based on the core samples and weight information in the multimodal feature set. The method analyzes the feature correlation between core samples and combines the features with correlation coefficients less than 0.3 to ensure feature complementarity. For example, in the concentration evaluation scenario, the theta / beta ratio core samples of the Fp1 and Fp2 channels in the feature set are weighted combined with the alpha energy core samples of the P3 and P4 channels. The weight coefficient is determined by the density distribution of the samples in the original category. For dynamic task scenarios, the method constructs spatiotemporal constraints for feature combination based on the spatial correlation information in the feature set. In driving fatigue monitoring, the core feature samples of the frontal lobe, occipital lobe and temporal lobe are fused according to their weight coefficients to generate a multidimensional fatigue index. The method evaluates the stability of the combined feature calculation to ensure that the intra-class aggregation of the combined feature is not lower than that of the original feature set. Finally, a set of feature combination schemes including combination methods, weight coefficients and stability indicators are obtained.

[0075] The feature matrix is ​​generated according to the combination rules and weight parameters in the feature combination scheme. The method first performs weighted fusion on each group of combined features, and the weight value directly uses the stability index normalization result in the scheme. Then the matrix structure is constructed, the rows of the matrix represent different feature combinations, and the columns represent the feature dimensions after the combination. For example, in the multimodal learning task, the stable combination of visual attention features, auditory processing features, and motor control features is mapped to a 100-dimensional feature space based on the combination scheme. In the attention persistence test, the method constructs a feature sequence according to the spatiotemporal constraints of the combination scheme to maintain the temporal continuity of the combined features. The generated feature matrix is ​​analyzed by the condition number. When the condition number exceeds 100, the matrix structure is optimized by adjusting the weight parameters in the combination scheme. The feature matrix finally generated not only maintains the discriminant information of the original features, but also reflects the combination relationship between features.

[0076] Step S105, performing dimension mapping on the feature matrix to obtain main components, outputting feature representations according to the main components, and completing noise reduction and feature extraction of the multi-channel EEG signal.

[0077] Specifically, dimension mapping is performed on the feature matrix. The method uses a nonlinear dimensionality reduction algorithm to construct a feature projection framework. First, the distance matrix of sample points in the high-dimensional feature space is calculated, and the similarity between features is measured using the Euclidean distance. Then, the high-dimensional features are projected into a 3D representation space based on the t-SNE algorithm, where the perplexity parameter is set to 30, the learning rate is set to 200, and the maximum number of iterations is 1000. For example, in a remote meeting scenario, the method reduces the dimension of the focus index in the feature matrix, maps the high-dimensional representation containing features such as theta / beta ratio and alpha band energy into a trajectory in a 3D space, and realizes real-time tracking of the attention state of the participants. For long-term meetings, the method calculates feature mapping every 30 seconds to construct an attention change curve. In the document reading task, the method combines and maps the visual processing features of the O1 and O2 channels and the cognitive load features of the Fp1 and Fp2 channels to generate a state space reflecting the reading engagement. For the emotion-related features of the F3 and F4 channels, the feature aggregation is calculated based on the spatial distribution after mapping. The higher the aggregation, the more stable the emotional state. For example, in a multi-person collaboration scenario, the method maps the attention features of all participants synchronously, and evaluates the team collaboration status by calculating the spatial distribution features of the mapping points. In a daily office environment, the method integrates and maps the user's visual attention features, cognitive load features, and emotional state features to construct a multi-dimensional work status indicator. The method evaluates the mapping quality by calculating the local structure retention rate before and after projection, and triggers the parameter optimization mechanism when the retention rate is lower than 85%. At the same time, the pressure value evaluation is introduced to ensure that the topological relationship of high-dimensional features is maintained during the dimensionality reduction process. Finally, the main component matrix containing mapping coordinates, quality indicators, and topological features is obtained.

[0078] In some embodiments, outputting the feature representation based on the main component includes: performing spatial transformation on the main component according to a principal component analysis method to obtain a spatial transformation structure corresponding to the main component; performing dimensionality reduction processing on the spatial transformation structure to obtain the feature representation.

[0079] The spatial transformation is performed based on the main component matrix obtained. The principal component analysis method is used to perform orthogonal transformation on the mapping results, and the feature directions with a cumulative contribution rate of 95% are selected to construct a new feature space. During the transformation process, the variance contribution rate and eigenvalue of each principal component are calculated, and the eigenvalues ​​are sorted according to the size of the eigenvalues. For example, in the document reading task, the main components of the visual attention features of the O1 and O2 channels are spatially rotated so that the maximum variance direction is aligned with the attention intensity axis to achieve a quantitative evaluation of reading concentration. In the video viewing scenario, the attention component of the visual channel and the processing component of the auditory channel are jointly transformed to establish a multimodal attention evaluation model. For online learning tasks, the cognitive load component and the visual processing component are aligned in the feature space, and the learning efficiency is judged by the projection direction of the principal component. In the concentration evaluation scenario, the cognitive load components of the Fp1 and Fp2 channels are scaled, the spatial scale is adjusted according to the size of the eigenvalue, and a standardized attention scoring system is established. For the conference interaction scenario, the attention features of multiple participants are uniformly transformed in space so that the attention states of different individuals can be compared in the same metric space. The method also establishes a discriminant analysis model based on the Fisher criterion to optimize the inter-class separation and intra-class aggregation, and improve the accuracy of state recognition. In the browsing scene, the state index reflecting the efficiency of information acquisition is constructed by spatial transformation of visual scanning features. Finally, the spatial transformation structure including transformation matrix, eigenvalue, discriminant coefficient and normalization parameter is obtained.

[0080] The spatial transformation structure is applied to reduce the dimensionality of the main components. The method first analyzes the condition number of the transformation matrix to ensure numerical stability. Then the truncation threshold is determined according to the size of the eigenvalue, and the dimension with a contribution rate of explained variance exceeding 2% is retained. For example, in the attention persistence test, the fatigue features of the frontal leads are compressed, and the optimal compression ratio is determined by calculating the information entropy to achieve long-term monitoring of the attention level. In the fast reading task, the method establishes a feature compression model to compress the high-dimensional features of visual processing and cognitive load into a low-dimensional representation that is convenient for real-time calculation. For the multi-channel collaborative features in the meeting scene, a spatiotemporal dimensionality reduction model is established to reduce feature redundancy while maintaining temporal association, and to achieve real-time evaluation of the group interaction state. In the outdoor activity scene, the method extracts stable attention features through dimensionality reduction processing to reduce the impact of environmental noise. For multi-task work scenarios, the attention features in the task switching process are adaptively reduced in dimensionality to capture the key features of task switching. The reconstruction error of the dimensionality reduction result is evaluated at multiple levels: the first level calculates the signal reconstruction error to ensure that the dimensionality reduction does not lose key information; the second level evaluates the degree of preservation of temporal features and verifies the continuity of state transitions; the third level analyzes the fidelity of spatial features to ensure the consistency of the relationship between channels. When the evaluation index of any level exceeds the threshold, the method automatically adjusts the dimensionality reduction parameters until the quality requirements are met. Through a multi-level optimization strategy, a feature representation that retains key information of cognitive states and has high computational efficiency is finally generated.

[0081] The method provided in this application has at least the following beneficial effects:

[0082] 1Through hierarchical feature extraction and processing strategies, the method realizes adaptive processing of the signals collected by smart glasses. First, the initial feature set is constructed based on multi-channel signal features, including time domain statistical features, frequency band energy features and spatial correlation features; then, through anomaly detection and data correction based on probability density function, a dynamic threshold mechanism is established to achieve adaptation to individual differences; finally, multi-level sub-band decomposition and spectrum analysis are used to extract stable signal features, effectively solving the problem of signal instability in portable EEG acquisition.

[0083] 2. The method based on template matching and multi-dimensional feature analysis is used to achieve accurate recognition of different cognitive states. The method first uses multi-scale derivative operators to detect edges and accurately locate key feature positions; then, a cross-channel feature segment alignment mechanism is established by combining hierarchical clustering and spatial consistency constraints; further, the improved wavelet packet reconstruction algorithm is used to achieve selective retention of features of different cognitive states; finally, feature expression is optimized based on spectrum analysis, which significantly improves the recognition accuracy of user cognitive states in complex environments.

[0084] 3 A hierarchical feature mapping and dimensionality reduction framework was established to achieve efficient expression of features. Methods A multi-scale decomposition system was constructed through continuous wavelet transform technology to achieve a fine characterization of signal features; a low-dimensional representation of features was established based on the t-SNE algorithm and hierarchical mapping structure; a multi-level quality assessment mechanism, including reconstruction error analysis, temporal feature retention, and spatial feature verification, was used to ensure the retention of key information during the dimensionality reduction process; and finally a feature representation method was formed that not only retained cognitive state discrimination information but also had high computational efficiency.

[0085] In order to implement the smart glasses EEG signal denoising and feature extraction method corresponding to the above method embodiment, to achieve the corresponding functions and technical effects. Figure 2 , Figure 2 The structure block diagram of a smart glasses EEG signal denoising and feature extraction device 200 provided in an embodiment of the present application is shown. For the convenience of description, only the parts related to the present embodiment are shown. The smart glasses EEG signal denoising and feature extraction device 200 provided in an embodiment of the present application includes:

[0086] The signal acquisition module 201 is used to collect multi-channel EEG signals of the smart glasses, obtain an original data stream, generate an initial feature set according to the original data stream, and generate a standard data stream according to the initial feature set;

[0087] The feature acquisition module 202 is used to perform sub-band decomposition on the standard data stream to obtain frequency band data, obtain energy features corresponding to the frequency band data, and obtain a noise reduction signal according to the energy features; perform edge detection on the noise reduction signal to obtain feature positions, and obtain contour features corresponding to the feature positions;

[0088] The type acquisition module 203 is used to acquire the feature template and the corresponding signal type corresponding to the contour feature, and construct a signal sample according to the signal type; acquire the frequency component corresponding to the signal sample, and reconstruct according to the frequency component to generate a pure waveform;

[0089] An index acquisition module 204 is used to acquire a characteristic index corresponding to the pure waveform, generate a characteristic map according to the characteristic index, and generate a characteristic matrix according to the characteristic map;

[0090] The extraction completion module 205 is used to perform dimension mapping on the feature matrix to obtain main components, output feature representations based on the main components, and complete noise reduction and feature extraction of the multi-channel EEG signal.

[0091] The above-mentioned smart glasses EEG signal denoising and feature extraction device 200 can implement the smart glasses EEG signal denoising and feature extraction method of the above-mentioned method embodiment. The options 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.

[0092] 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 3 Only 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.

[0093] 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.

[0094] 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.

[0095] 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.

[0096] 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.

[0097] 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.

[0098] 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.

[0099] 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.

[0100] 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 denoising and extracting features of EEG signals in smart glasses, characterized in that: include: Collecting multi-channel EEG signals of the smart glasses to obtain an original data stream, generating an initial feature set according to the original data stream, and generating a standard data stream according to the initial feature set; Performing sub-band decomposition on the standard data stream to obtain frequency band data, obtaining energy features corresponding to the frequency band data, and obtaining a noise reduction signal according to the energy features; performing edge detection on the noise reduction signal to obtain a feature position, and obtaining a contour feature corresponding to the feature position; Acquire a feature template and a corresponding signal type corresponding to the contour feature, and construct a signal sample according to the signal type; acquire a frequency component corresponding to the signal sample, and reconstruct according to the frequency component to generate a pure waveform; Acquire characteristic indicators corresponding to the pure waveform, generate characteristic maps according to the characteristic indicators, and generate characteristic matrices according to the characteristic maps; The feature matrix is ​​dimensionally mapped to obtain the main components, and feature representation is output according to the main components to complete noise reduction and feature extraction of the multi-channel EEG signal.

2. The method according to claim 1, characterized in that The step of generating a standard data stream according to the initial feature set comprises: Performing amplitude statistics on the initial feature set to obtain amplitude distribution; Acquire the abnormal point position corresponding to the initial feature set according to the amplitude distribution; The initial feature set is corrected according to the position of the abnormal point to generate the standard data stream.

3. The method according to claim 1, characterized in that The obtaining of the noise reduction signal according to the energy feature comprises: Performing threshold selection on the energy feature according to a preset threshold to obtain a valid signal corresponding to the energy feature; Obtaining a reconstructed waveform corresponding to the effective signal; The noise reduction signal is obtained according to the reconstructed waveform.

4. The method according to claim 1, characterized in that The obtaining of the contour feature corresponding to the feature position includes: Acquire a signal boundary corresponding to the noise reduction signal according to the characteristic position; Dividing the noise reduction signal according to the signal boundary to obtain feature segments; Feature extraction is performed on the feature segment to obtain the contour feature.

5. The method according to claim 1, characterized in that The reconstructing according to the frequency components to generate a pure waveform includes: performing bandpass screening on the frequency components; Acquire the frequency component corresponding to the frequency component according to the screening result corresponding to the bandpass screening; The clean waveform is generated by reconstructing a signal according to the frequency components.

6. The method according to claim 1, characterized in that The step of generating a feature matrix according to the feature mapping comprises: Perform grouping and clustering according to the feature map to obtain feature categories; Constructing a feature set according to the feature category; The feature matrix is ​​generated according to the feature set.

7. The method according to claim 1, characterized in that The outputting a feature representation according to the main component comprises: Performing spatial transformation on the main components according to the principal component analysis method to obtain the spatial transformation structure corresponding to the main components; Perform dimensionality reduction processing on the spatial transformation structure to obtain the feature representation.

8. The method according to claim 1, characterized in that The generating an initial feature set according to the original data stream comprises: Performing time window segmentation on the original data stream to obtain time series segments; The initial feature set is generated according to the time sequence segment.

9. A device for reducing noise and extracting features of EEG signals for smart glasses, characterized in that: include: A signal acquisition module, used to collect multi-channel EEG signals of the smart glasses, obtain an original data stream, generate an initial feature set according to the original data stream, and generate a standard data stream according to the initial feature set; A feature acquisition module is used to perform sub-band decomposition on the standard data stream to obtain frequency band data, obtain energy features corresponding to the frequency band data, and obtain a noise reduction signal according to the energy features; perform edge detection on the noise reduction signal to obtain feature positions, and obtain contour features corresponding to the feature positions; A type acquisition module is used to acquire a feature template and a corresponding signal type corresponding to the contour feature, and construct a signal sample according to the signal type; acquire a frequency component corresponding to the signal sample, and reconstruct the signal sample according to the frequency component to generate a pure waveform; An index acquisition module, used for acquiring characteristic indexes corresponding to the pure waveform, generating a characteristic map according to the characteristic indexes, and generating a characteristic matrix according to the characteristic map; The extraction completion module is used to perform dimension mapping on the feature matrix to obtain the main components, output feature representations based on the main components, and complete noise reduction and feature extraction of the multi-channel EEG signal.

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