Intelligent glasses multi-channel electroencephalogram signal spatial filtering and feature enhancing method and device

Abstract

The invention discloses a multi-channel electroencephalogram signal spatial filtering and feature enhancement method and device for intelligent glasses, and the method comprises the steps: taking a multi-channel electroencephalogram signal collected by the intelligent glasses as an original signal, generating a corresponding electrode mapping relation for analysis, determining a spatial structure and a filtering scale, and generating a filter group according to the filtering scale; performing component separation on the original signal according to the filter bank, obtaining component information and a corresponding signal source, and constructing an independent component sequence according to the signal source; spatial mapping is carried out on the independent component sequence, main components are determined, a weight matrix is constructed for spatial filtering, a spatial feature sequence is generated for spatial reconstruction, and enhanced features are generated; performing multi-scale decomposition on the filter bank and the enhanced features, acquiring multi-scale features to construct a feature combination, acquiring corresponding time-frequency features to perform spatial mapping, acquiring interference modes corresponding to the time-frequency features to perform feature enhancement, acquiring corresponding feature representations to perform spatial reconstruction, and acquiring spatial enhanced features.

Description

Technical Field

[0001] The present application relates to the field of brain-computer interface technology, and in particular to a method and device for spatial filtering and feature enhancement of multi-channel EEG signals of smart glasses. Background Art

[0002] The multi-channel EEG acquisition system in smart glasses can simultaneously acquire EEG signals from multiple locations, making it possible to perform spatial analysis of EEG signals. However, due to the limited number of electrodes collected by smart glasses and the restrictions on their distribution by the structure of the glasses, traditional spatial filtering methods are difficult to apply effectively. At the same time, problems such as spatial crosstalk between electrodes, correlation caused by common reference electrodes, and local electromyographic interference seriously affect the extraction of effective EEG features. Especially in real application scenarios, the user's movements, emotional changes, and cognitive activities will introduce complex interference patterns, resulting in a significant decrease in signal quality.

[0003] Existing spatial filtering technologies are mainly designed for traditional EEG caps, which require more electrode coverage and regular electrode distribution. When these methods are directly applied to smart glasses, problems such as over-smoothing, decreased spatial resolution, and suppressed target features often occur. In dynamic scenes, the performance of spatial filtering will be further reduced due to slight changes in electrode positions and the influence of motion artifacts. In addition, existing methods lack targeted consideration of the specific wearing position and electrode distribution characteristics of smart glasses, and cannot establish effective electrode mapping relationships and multi-scale filtering strategies. At the same time, these methods often use a unified parameter configuration when processing time-frequency features, which makes it difficult to adapt to the feature requirements of different cognitive tasks, and cannot achieve dynamic enhancement and spatial reconstruction of target EEG features.

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

[0005] The embodiment of the present application provides a method and device for spatial filtering and feature enhancement of multi-channel EEG signals of smart glasses, aiming to solve the problem that existing methods lack targeted consideration of the specific wearing position and electrode distribution characteristics of smart glasses, and cannot establish effective electrode mapping relationships and multi-scale filtering strategies. At the same time, these methods often use a unified parameter configuration when processing time-frequency features, which is difficult to adapt to the feature requirements of different cognitive tasks, and cannot achieve the problem of dynamic enhancement and spatial reconstruction of target EEG features.

[0006] In a first aspect, an embodiment of the present application provides a method for spatial filtering and feature enhancement of multi-channel EEG signals of smart glasses, comprising:

[0007] Taking the multi-channel EEG signals collected by the smart glasses as original signals, establishing an electrode spatial position model according to the original signals, and generating an electrode mapping relationship according to the electrode spatial position model;

[0008] Performing a topological analysis on the electrode mapping relationship to determine a spatial structure and a corresponding filtering scale, and generating a filter bank according to the filtering scale;

[0009] Separating the original signal into components according to the filter group, obtaining component information, obtaining a signal source corresponding to the component information, and constructing an independent component sequence according to the signal source; the component information includes at least independent components, cluster components and principal components;

[0010] Performing spatial mapping on the independent component sequence to determine the main components, constructing a weight matrix according to the main components, performing spatial filtering according to the weight matrix to generate a spatial feature sequence, and performing spatial reconstruction on the spatial feature sequence to generate enhanced features;

[0011] Performing multi-scale decomposition on the filter group and the enhanced features to obtain multi-scale features, constructing a feature combination according to the multi-scale features, and obtaining time-frequency features corresponding to the feature combination;

[0012] Performing spatial mapping on the time-frequency features to obtain an interference pattern corresponding to the time-frequency features, performing feature enhancement according to the interference pattern to obtain a feature representation corresponding to the interference pattern;

[0013] The spatial reconstruction is performed according to the feature representation, and the feature fusion is performed on the reconstruction result to obtain the spatial enhancement feature.

[0014] In a second aspect, the present application further provides a spatial filtering and feature enhancement device, comprising:

[0015] A signal acquisition module, used to use the multi-channel EEG signals acquired by the smart glasses as original signals, establish an electrode spatial position model according to the original signals, and generate an electrode mapping relationship according to the electrode spatial position model;

[0016] A topological analysis module, used to perform topological analysis on the electrode mapping relationship, determine the spatial structure and the corresponding filtering scale, and generate a filter group according to the filtering scale;

[0017] A component separation module, used to separate the components of the original signal according to the filter bank, obtain component information, obtain a signal source corresponding to the component information, and construct an independent component sequence according to the signal source; the component information includes at least independent components, cluster components and principal components;

[0018] A spatial mapping module is used to perform spatial mapping on the independent component sequence, determine the main components, construct a weight matrix according to the main components, perform spatial filtering according to the weight matrix, generate a spatial feature sequence, perform spatial reconstruction on the spatial feature sequence, and generate enhanced features;

[0019] A scale decomposition module, used to perform multi-scale decomposition on the filter group and the enhanced features, obtain multi-scale features, construct a feature combination according to the multi-scale features, and obtain time-frequency features corresponding to the feature combination;

[0020] A spatial mapping module, used to perform spatial mapping on the time-frequency features, obtain an interference pattern corresponding to the time-frequency features, perform feature enhancement according to the interference pattern, and obtain a feature representation corresponding to the interference pattern;

[0021] The spatial reconstruction module is used to perform spatial reconstruction according to the feature representation, perform feature fusion on the reconstruction result, and obtain spatial enhancement features.

[0022] In a third aspect, the present application also provides a computer device, including 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 spatial filtering and feature enhancement of multi-channel EEG signals of smart glasses as described in the first aspect is implemented.

[0023] 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 method for spatial filtering and feature enhancement of multi-channel EEG signals of smart glasses as described in the first aspect.

[0024] Compared with the prior art, this application has at least the following beneficial effects:

[0025] 1. This solution establishes a complete electrode mapping and multi-scale filtering system, overcoming the limitation of the limited number of electrodes in smart glasses. By establishing the electrode spatial position model and mapping relationship, performing topological analysis, and determining the appropriate filtering scale, the optimal utilization of limited electrode signals is achieved.

[0026] 2. This solution builds a complete processing chain from signal decomposition to feature enhancement, effectively solving the problem of multi-source interference. By achieving spatial decomposition, component separation, and feature enhancement, and then going through the steps of interference pattern recognition and spatial template construction, the method significantly improves its anti-interference ability while ensuring the integrity of target features.

[0027] 3. This solution achieves dynamic adaptation of feature extraction and enhancement, meeting the needs of different application scenarios. Through multi-scale decomposition and time-frequency analysis, and then through spatial reconstruction and feature fusion, it achieves accurate response to different cognitive tasks and ensures the temporal and spatial consistency of enhanced features.

[0028] 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

[0029] Figure 1 This is a flow chart of a method for spatial filtering and feature enhancement of multi-channel EEG signals of smart glasses shown in an embodiment of the present application;

[0030] Figure 2 This is a schematic diagram of the structure of a spatial filtering and feature enhancement device shown in an embodiment of the present application;

[0031] Figure 3 A schematic diagram of the structure of a computer device shown in an embodiment of the present application. DETAILED DESCRIPTION

[0032] 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.

[0033] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or combinations thereof.

[0034] 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.

[0035] 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.

[0036] 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.

[0037] 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.

[0038] The technical solution of the embodiment of the present application is introduced below.

[0039] The multi-channel EEG acquisition system in smart glasses can simultaneously acquire EEG signals from multiple locations, making it possible to perform spatial analysis of EEG signals. However, due to the limited number of electrodes collected by smart glasses and the restrictions on their distribution by the structure of the glasses, traditional spatial filtering methods are difficult to apply effectively. At the same time, problems such as spatial crosstalk between electrodes, correlation caused by common reference electrodes, and local electromyographic interference seriously affect the extraction of effective EEG features. Especially in real application scenarios, the user's movements, emotional changes, and cognitive activities will introduce complex interference patterns, resulting in a significant decrease in signal quality.

[0040] Existing spatial filtering technologies are mainly designed for traditional EEG caps, which require more electrode coverage and regular electrode distribution. When these methods are directly applied to smart glasses, problems such as over-smoothing, decreased spatial resolution, and suppressed target features often occur. In dynamic scenes, the performance of spatial filtering will be further reduced due to slight changes in electrode positions and the influence of motion artifacts. In addition, existing methods lack targeted consideration of the specific wearing position and electrode distribution characteristics of smart glasses, and cannot establish effective electrode mapping relationships and multi-scale filtering strategies. At the same time, these methods often use a unified parameter configuration when processing time-frequency features, which makes it difficult to adapt to the feature requirements of different cognitive tasks, and cannot achieve dynamic enhancement and spatial reconstruction of target EEG features.

[0041] To solve the above problems, please refer to Figure 1 , Figure 1 The flowchart of a method for spatial filtering and feature enhancement of multi-channel EEG signals of smart glasses provided in an embodiment of the present application is shown in FIG. The method for spatial filtering and feature enhancement of multi-channel EEG signals of 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 1As shown, the method for spatial filtering and feature enhancement of multi-channel EEG signals of smart glasses in this embodiment includes steps S101 to S107, which are described in detail as follows:

[0042] Step S101, taking the multi-channel EEG signals collected by the smart glasses as original signals, establishing an electrode spatial position model according to the original signals, and generating an electrode mapping relationship according to the electrode spatial position model.

[0043] Specifically, the multi-channel EEG acquisition system of smart glasses acquires the user's EEG signals through a dry electrode array arranged at specific positions of the frame. The system is equipped with 12 acquisition electrodes, which are located at key positions in the forehead, temples and both sides of the head. The sampling frequency is set at 1000Hz to ensure the time resolution of the signal. Each acquisition channel is equipped with a dedicated signal conditioning circuit, including a preamplifier, a bandpass filter and a digital isolation module. The signal is amplified 5000 times and converted to 16 bits to form a multi-channel digital signal stream. During the acquisition process, the system monitors the contact impedance of each electrode in real time and ensures that it is stable below 5kΩ. The common mode interference is effectively suppressed through active shielding technology and differential amplifier circuits. The adaptive baseline correction algorithm is used to eliminate signal drift in real time, and the digital low-pass filter is used to suppress high-frequency noise and power frequency interference. The system also integrates a motion artifact detection module, which can mark and remove abnormal signal segments caused by blinking, chewing and other actions. For different wearing scenarios, an intelligent gain control strategy has been developed, which can automatically adjust the amplification factor according to the signal amplitude to ensure that the signal is always in the optimal acquisition range. At the same time, the system adopts a high-precision clock synchronization mechanism to ensure the precise alignment of multi-channel data sampling, and realizes real-time transmission and temporary storage of data through a cache management mechanism.

[0044] Based on the collected multi-channel EEG signals, the system establishes an accurate electrode spatial position model. The model uses the smart glasses frame as the standard reference system and establishes a spatial rectangular coordinate system with the center point of the bridge of the nose as the coordinate origin. In practical applications, the system integrates a miniature three-axis accelerometer and an angle sensor to obtain the spatial posture information of the glasses in real time. Through the sensor data fusion algorithm, the method can compensate for the position deviation caused by head movement and ensure the stability of the electrode spatial position. Considering the differences in head shapes of different users, the method uses a deformable mesh model to describe the spatial distribution of the electrode array and dynamically deforms it according to the actual wearing situation. The position model also contains spatial constraints based on anatomical structures to ensure that the electrode position description corresponds to the actual brain area distribution. Through a two-layer position optimization strategy, the method first performs a rough alignment based on rigid body transformation, and then uses a local deformation algorithm to achieve fine matching, and finally obtains a high-precision representation of the electrode spatial distribution. For dynamic usage scenarios, a real-time position calibration mechanism is developed to dynamically update the electrode position information by analyzing the sensor data stream. The method also establishes a position error compensation model, which can predict and correct the position deviation caused by the movement of glasses based on historical data. In order to improve the stability of the position model, a spatial filtering algorithm is introduced to smooth the original position data and effectively suppress the position jitter caused by sensor noise.

[0045] After obtaining the accurate spatial position of the electrodes, the method constructs the mapping relationship between the electrodes through multi-level analysis. First, the initial connection relationship is established based on the geometric distance between the electrodes, and the spatial coupling strength is calculated using an adaptive weight function. Considering the conductive properties of the scalp tissue, the method introduces a conduction model based on physiological characteristics to estimate the diffusion characteristics of the signal in space. A special spatial correlation analysis method is developed for the electrode distribution pattern unique to smart glasses. This method comprehensively considers the signal conduction path, tissue impedance distribution, and electrode contact characteristics to construct an electrode mapping relationship that is more in line with the actual situation. For example, when the user is sitting and working in an office environment, the method mainly focuses on the electrode mapping of the frontal region, and improves the accuracy of signal detection by adjusting the weight configuration of the forehead electrode; when the user is walking or exercising lightly, the method automatically adjusts to the exercise mode, focusing on the mapping relationship of the electrodes in the temple area, and suppressing the influence of motion artifacts by optimizing the mapping parameters; when the user is resting, the method adopts a balanced mapping strategy to maintain similar weight distribution for all brain region electrodes. In order to cope with the dynamic measurement environment, the method adopts a sliding window strategy to update the mapping weights in real time, and the window length is adaptively adjusted according to the signal stability. By analyzing the signal correlation and phase consistency of adjacent electrodes, the method can identify and suppress abnormal connections caused by local electromyographic interference or loose electrodes. In the process of mapping relationship construction, special attention is paid to the spatial constraints caused by the structure of glasses. By calculating the relative position changes of electrodes under different wearing conditions, a mapping optimization scheme considering physical constraints is established. The method realizes multi-scale mapping analysis function, which can describe the correlation characteristics between electrodes at different spatial scales, ensuring that a stable and reliable spatial mapping relationship can be obtained in various practical usage scenarios.

[0046] Step S102, topological analysis is performed on the electrode mapping relationship to determine the spatial structure and the corresponding filtering scale, and a filter bank is generated according to the filtering scale.

[0047] Specifically, based on the mapping relationship between electrodes obtained in the steps, the method first conducts an in-depth topological analysis to determine the spatial structural characteristics. By calculating the connection strength, connection density and adjacency relationship of each electrode node, the key electrode connection pattern is identified. The connection strength is measured by the signal correlation and phase synchronization between electrodes, and the connection density reflects the degree of interconnection between electrodes in the local area.

[0048] In some embodiments, the electrode mapping relationship is topologically analyzed to determine the spatial structure and the corresponding filtering scale, including: obtaining the connection strength, connection density and adjacency relationship of each electrode node according to the electrode mapping relationship; determining the key electrode connection pattern according to the connection strength, connection density and adjacency relationship; determining the spatial structure and constructing a connection feature matrix according to the key electrode connection pattern; dividing the connection feature matrix according to a multi-level clustering strategy to obtain the filtering scale.

[0049] The average connection strength within the region and the ratio of the connection density between regions were calculated for the distribution of electrodes in different functional areas such as the frontal lobe and temple. For example, when the user was concentrating on reading, the internal connection strength of the electrodes in the frontal lobe region reached above 0.7, and the connection density in the region was 85%, while the connection strength with other regions was less than 0.3, and the connection density was only 30%; when the user was doing simple movements, the electrodes in the temporal lobe region showed a significant asymmetric distribution, with the average connection strength of the left electrode group being 0.65 and the connection density being 75%, which was about 30% higher than that on the right side. The method established a connection feature matrix to record the connection strength, directionality, and stability indicators of all electrode pairs, providing a quantitative basis for subsequent scale division.

[0050] Exemplarily, generating a filter group according to the filtering scale includes: constructing a parameter template according to the filtering scale; the parameter template includes at least kernel function parameters, spatial coverage range, weight distribution and directional weight template; generating a filter group according to the parameter template.

[0051] It should be noted that, in some embodiments, constructing a parameter template according to the filtering scale includes: obtaining the average connection strength, inter-regional connection density and connection strength matrix corresponding to each electrode unit of the smart glasses; at a local scale, generating kernel function parameters according to the average connection strength; wherein, the higher the average connection strength, the smaller the spatial attenuation rate of the kernel function parameters; at a medium scale, adjusting the spatial coverage and weight distribution corresponding to the filter according to the inter-regional connection density ratio; at a global scale, constructing a directional weight template according to the eigenvector of the connection strength matrix; in the directional weight template, the weight coefficient of the direction with a larger eigenvalue of the eigenvector is higher; generating the parameter template according to the kernel function parameters, spatial coverage, weight distribution and directional weight template.

[0052] According to the obtained connection feature matrix, the method uses a multi-level clustering strategy to divide the filter scale. First, based on the connection strength threshold (0.7) and the connection density threshold (75%), highly coupled electrode groups are identified and divided into local scale units. For example, two highly coupled local units were detected in the prefrontal region, containing 3 and 4 electrodes respectively. Then, according to the ratio of the connection density between regions (the threshold is set to 50%), the functionally related local units are combined into medium-scale modules. For example, the two local units detected in the prefrontal region are merged into a medium-scale module because the connection density between them reaches 65%. Finally, based on the eigenvalue decomposition of the inter-regional connection strength matrix, the global scale division is determined, and the remote electrode pairs with a connection strength of more than 0.3 are included in the same large-scale processing unit. This multi-level division based on connection features forms three spatial scale levels: local (2-4cm), medium (4-6cm), and global (>6cm).

[0053] Based on the divided spatial scale hierarchy, the method combines specific connection features to construct parameter templates. At the local scale, the kernel function parameters are set according to the average connection strength within the electrode unit. The higher the connection strength, the smaller the spatial attenuation rate of the kernel function. For example, the high coupling unit in the frontal lobe (connection strength 0.7) uses a Gaussian kernel with an attenuation rate of 0.3, while the low coupling unit in the temporal lobe (connection strength 0.5) uses a Gaussian kernel with an attenuation rate of 0.6. At the medium scale, the spatial coverage and weight distribution of the filter are adjusted based on the ratio of the connection density between regions. For example, the internal connection density of the medium-scale module in the frontal lobe is 75%, based on which an effective coverage radius of 5 cm is set, and a center weight of 0.8 is used. At the global scale, the eigenvector of the connection strength matrix is ​​used to construct a directional weight template, and the direction with a large eigenvalue is given a higher weight coefficient. The parameter template obtained in this way contains a complete set of parameters such as spatial scale, kernel function type, and weight distribution.

[0054] According to the parameter templates of each scale, the method generates the corresponding filter group. At the local scale level, each high coupling unit is configured with a set of spatial filters with matching kernel function parameters. For example, the filter of the prefrontal unit uses a Gaussian kernel (attenuation rate 0.3) with a processing range of 3cm, which is mainly used to extract local synchronization features; the filter of the temporal unit uses a Gaussian kernel (attenuation rate 0.6) with a processing range of 2cm, focusing on suppressing local interference. At the medium scale level, regional filters are generated based on the connection density distribution, and the coverage range is determined by the effective radius in the parameter template. For example, the filter coverage radius of the prefrontal region is 5cm, the center weight is 0.8, which is used to extract regional cognitive features; the filter coverage radius of the movement-related region is 4cm, the center weight is 0.7, which is used to monitor movement-related activities. At the global scale level, a directional filter is constructed according to the connection strength eigenvector. The direction with a larger eigenvalue (such as the front-to-back direction, eigenvalue 0.6) is configured with a response gain of 1.2, and the direction with a smaller eigenvalue (such as the left-right direction, eigenvalue 0.3) is configured with a response gain of 0.8. This multi-level filter group realizes the selective extraction of features at different spatial scales. In practical applications, the filter bank can adaptively adjust parameters according to the dynamic changes of connection features. For example, when the connection strength of the frontal lobe region is detected to drop from 0.7 to 0.5, the kernel function attenuation rate of the local filter is adjusted accordingly; when the connection feature value in a certain direction is found to increase significantly, the response gain of the filter in the corresponding direction is promptly increased.

[0055] Step S103, separating the components of the original signal according to the filter bank, obtaining component information, obtaining the signal source corresponding to the component information, and constructing an independent component sequence according to the signal source; the component information at least includes independent components, cluster components and principal components.

[0056] Specifically, the method performs spatial decomposition by combining the original signal and the filter bank. First, the original signal is segmented according to the sampling time window (2 seconds), and each segment is convolved with the multi-scale filter bank. At the local scale, the signal is spatially filtered using a Gaussian kernel filter (attenuation rate 0.3-0.6), and 2-4 main spatial components are obtained in each local area, among which the energy share of the frontal lobe region is the highest, reaching 65%, followed by the temporal lobe region, accounting for 35%, and the occipital lobe region is about 25%; at the medium scale, a regional filter (covering radius 4-5cm) is used to extract regional activity features, and 3-5 significant regional components are obtained, with the frontal-temporal lobe connection region accounting for 40% of the energy, and the parietal-occipital lobe connection region accounting for 30%; at the global scale, a directional filter (gain 0.8-1.2) is used to obtain large-scale spatial patterns, usually obtaining 2-3 stable global components, with energy shares of 15%, 10%, and 5%, respectively. The spatial power spectrum and phase spectrum are calculated for each scale level to obtain the energy distribution in the 0-30Hz frequency band. In the resting state, the energy of the spatial component in the alpha band (8-13Hz) is dominant, while in cognitive tasks, the energy of the component in the theta band (4-7Hz) increases significantly.MethodsThe decomposition results of 50 consecutive time windows were statistically analyzed to establish the energy distribution histogram of the spatial component, which was used for the subsequent component separation threshold setting.

[0057] Based on the energy distribution characteristics obtained by spatial decomposition, the method separates the components of each scale. At the local scale, the spatial components with an energy share of more than 25% are first selected as candidates, and the mutual information and correlation coefficients between these components are calculated. When the mutual information is less than 0.2 and the correlation coefficient is less than 0.3, it is marked as an independent component; when the correlation coefficient is between 0.3-0.6, the conditional mutual information is further calculated, and if it is less than 0.15, it is still considered an independent component. For example, three independent components are usually separated in the prefrontal region, corresponding to the activity patterns of the left, middle, and right frontal lobes, and their energy shares are 35%, 20%, and 15%, respectively. At the medium scale, feature clustering is performed based on the regional components with an energy share of more than 30%. A two-step clustering strategy is adopted: first, preliminary grouping is performed based on spatial correlation (threshold 0.6), and then the phase synchronization index of the components within the group is calculated. When the synchronization index exceeds 0.5 and the duration is greater than 500ms, it is confirmed as the same cluster. In this way, two stable cluster components are usually obtained in the frontal-temporal region, reflecting the activity patterns related to cognitive processing and emotion regulation, respectively. For the global scale component, according to the previously obtained energy distribution, the component with cumulative energy accounting for 80% is selected for principal component analysis, and the principal component with eigenvalue greater than 1 is extracted as the global signal feature.

[0058] The method uses the separated independent components, cluster components and principal components to mark the signal source. For local independent components, the source is located by combining their spatial position and spectral characteristics. When the spatial distribution center of the component is located in the frontal area and the energy proportion of the θ band exceeds 40%, it is marked as an attention-related source; when it is located in the temporal lobe and the energy of the α band is significant, it is marked as a perceptual processing source. The method also calculates characteristic parameters such as source strength (based on energy proportion), source stability (based on temporal persistence) and source specificity (based on spatial distribution kurtosis). For medium-sized cluster components, their spatial connection patterns and temporal synchronization characteristics are analyzed. When components in multiple brain regions show stable phase synchronization (synchronization index>0.7 and duration>1 second), they are marked as functional network sources, and topological parameters such as network centrality, clustering degree and efficiency are calculated. For example, in the working memory task, the fronto-parietal working memory network (number of nodes 4-6, average path length 2.3) and the temporo-parietal attention network (number of nodes 3-5, clustering coefficient 0.65) can be identified. The principal components of the global scale were labeled according to the distribution characteristics of their spatial patterns and the laws of their temporal evolution. When the principal components showed a front-to-back symmetrical spatial distribution and existed stably in the resting state, they were labeled as default network sources; when they showed left-right asymmetry and were related to the task state, they were labeled as cognitive control sources.

[0059] Based on the information of the labeled signal sources, the method constructs multi-level independent component sequences. At the local level, the significance score of each independent component is first calculated, and the score is obtained by weighting the three indicators of energy contribution rate (weight 0.4), temporal stability (weight 0.3) and spatial specificity (weight 0.3). Components with scores exceeding 0.7 are selected to construct local sequences, such as the prefrontal attention source (score 0.85) and the temporal auditory processing source (score 0.78). At the medium-scale level, network sequences are constructed based on the characteristic parameters of the functional network sources. The overall score of the network is calculated, including the connection strength (weight 0.35), spatial coverage (weight 0.35) and temporal stability (weight 0.3). Network sources with scores exceeding 0.65 are included in the sequence, such as the working memory network (score 0.82) and the attention network (score 0.73). At the global level, important components are screened according to the explained variance (threshold 10%) and temporal consistency (consistency index>0.6) of the principal component sources to construct sequences reflecting large-scale brain activity patterns. The method dynamically maintains these multi-level sequences, and triggers the sequence update mechanism when a significant change in source features is detected (such as intensity change > 20% or spatial offset > 2 cm). For example, in the attention switching task, when the subject's attention shifts from visual to auditory stimulation, the method can capture the energy redistribution of the prefrontal attention source (the energy of the visual-related source decreases by 30%, and the energy of the auditory-related source increases by 25%) and the dynamic reorganization of the attention network (the visual-parietal connection weakens, and the auditory-parietal connection strengthens), and updates the independent component sequences at each level accordingly.

[0060] Step S104, spatially map the independent component sequence to determine the main components, construct a weight matrix based on the main components, perform spatial filtering based on the weight matrix to generate a spatial feature sequence, spatially reconstruct the spatial feature sequence, and generate enhanced features.

[0061] Specifically, the obtained independent component sequence was spatially mapped. At the local scale, the independent components with scores exceeding 0.7 were mapped to the corresponding brain area space to establish an accurate spatial distribution map. For example, the prefrontal attention source (score 0.85) formed an activation center with an intensity of 0.8 in the forehead midline area after mapping, showing a ring-like attenuation outward, covering an area of ​​5 cm²; the temporal auditory processing source (score 0.78) showed a strong activation area of ​​4 cm² on the left (intensity 0.75) and a secondary activation area of ​​3 cm² on the right (intensity 0.65). At the medium scale, the mapping of the working memory network (score 0.82) formed a ribbon connection between the prefrontal and parietal lobes, with a width of 3 cm and a total coverage area of ​​12 cm², and the connection strength showed a gradient distribution of 0.7-0.5; the attention network (score 0.73) formed a ring distribution in the temporal-parietal region, covering an area of ​​10 cm², and the activation intensity at the main nodes remained above 0.6. On a global scale, the mapping of the default network source forms a symmetrical distribution in the front and back brain regions, covering 8cm² in the front and 12cm² in the back, with an intensity distribution of 0.5-0.4. When the user performs a video viewing task, the method detects a 6cm² persistent activation (intensity 0.7) in the occipital visual area, and forms a 4cm wide connection with the parietal attention area.

[0062] Based on the specific numerical characteristics of spatial mapping, the method constructs a weight matrix. At the local scale, according to the 5cm² coverage feature of the prefrontal attention source, a 5×5 local weight grid is set, the weight of the center position is set to 0.85 (consistent with the mapping intensity), and it decays by 0.15 every 1cm outward to form an accurate weight gradient; for the temporal auditory source, based on the left-right asymmetric distribution characteristics (4cm² and 3cm²), 4×4 and 3×3 weight grids are set respectively, with the maximum weight of 0.75 on the left and 0.65 on the right. At the medium scale, the weight distribution of the working memory network follows the band distribution feature. In the coverage area of ​​12cm², the weight of the prefrontal node is set to 0.7 (corresponding to the mapping intensity), the weight of the parietal node is set to 0.6, and the weight on the connecting belt is linearly decreased by 0.1 according to the width of 3cm; the attention network sets a polar coordinate weight distribution in the ring area of ​​10cm², the node position weight is 0.6, the ring spacing is 1cm, and the weight decay step is 0.1. On a global scale, a large-scale weight field was constructed based on the total coverage area of ​​20 cm², and the corresponding weights were set for the front 8 cm² area and the rear 12 cm² area according to the measured intensity range of 0.5-0.4.

[0063] The constructed weight matrix was used to perform spatial filtering on the signal. In the 5cm² high-weight area of ​​the prefrontal lobe (weight>0.7), a 2cm×2cm sliding window was used, and the window weight was multiplied by the matrix element point by point and then summed up, retaining the attention-related θ wave characteristics; in the asymmetric area of ​​the temporal lobe, a 2.5cm×2.5cm window was used in the 4cm² area on the left and a 2cm×2cm window was used in the 3cm² area on the right to extract the characteristics of auditory processing. For the 12cm² area covered by the working memory network, a 3cm×1cm anisotropic filtering window was used along the 3cm-wide connection band direction to enhance the information transfer characteristics between nodes; an adaptive window in polar coordinate form was used in the 10cm² area of ​​the attention network, and the window size varied between 2-4cm with the circumferential position. In the 20cm² coverage area of ​​the default network source, a 3cm×3cm window was used in the front and a 4cm×4cm window was used in the back to achieve large-scale spatial smoothing. For example, when the user switches from rest to cognitive tasks, the method detects that the theta wave energy in the prefrontal area increases by 35%, and obtains stable attention signatures through local fine filtering; when switching between dual tasks, the direction of information flow on the working memory network connection belt changes, and the method captures this dynamic change process through anisotropic filtering.

[0064] According to the results of spatial filtering, the method generates spatial feature sequences. For the 5cm² high-weight area of ​​the frontal lobe, a feature unit is divided every 2cm², and three groups of features, namely energy distribution, phase consistency and directionality, are extracted to form a 15-dimensional local feature vector; in the asymmetric area of ​​the temporal lobe, a 12-dimensional feature vector is extracted from the left 4cm² area, and a 9-dimensional feature vector is extracted from the right 3cm² area. The two form a cross-hemispheric feature sequence through position correspondence. The 12cm² coverage area of ​​the working memory network is uniformly sampled along the connection belt direction, and a set of feature parameters including node activity intensity, connection flow direction and phase difference are extracted every 1cm² to form a 36-dimensional network feature sequence; the 10cm² area of ​​the attention network sets 8 sampling points on the ring structure, and 5-dimensional features are extracted from each point, resulting in a 40-dimensional ring feature sequence. For the 20cm² area of ​​the default network source, wide-area synchronization, energy gradient and phase coupling features are extracted from the front 8cm² and the back 12cm², respectively, and synthesized to form a 100-dimensional global feature sequence. For example, in the video viewing task, the method extracted an 18-dimensional visual feature sequence from the 6cm² activation area of ​​the occipital lobe, and at the same time extracted a 12-dimensional attention regulation feature sequence from the 4cm connection zone with the parietal lobe, which together constituted the spatiotemporal feature expression of visual attention. The method analyzed these feature sequences in real time, and when the local energy distribution deviation exceeded 25% or the network connection feature changed by more than 30%, the feature extraction parameters were adjusted in time to ensure the timeliness and reliability of the feature sequence.

[0065] In some embodiments, the spatial reconstruction of the spatial feature sequence to generate enhanced features includes: performing local projection on the spatial feature sequence to obtain a feature area; performing spatial reconstruction based on the feature area; and generating the enhanced features based on a reconstruction result corresponding to the spatial reconstruction.

[0066] The generated spatial feature sequence is locally projected. In the prefrontal region, the energy-oriented main direction decomposition is performed on the local feature vector to obtain the projection results of three feature axes. For example, when the user performs a continuous reading task, the first feature axis reflects the energy distribution of the θ band, showing the characteristics of center focus in the stage of maintaining attention; when the attention is shifted to different positions on the page, the energy center moves with the movement of the line of sight; when the attention is distracted, the energy distribution shows a clear diffusion trend. The second feature axis describes the propagation characteristics of phase consistency, which is manifested as an orderly propagation from the forehead to the parietal lobe during deep reading comprehension; when encountering difficult words and sentences, this propagation is briefly interrupted and reorganized; and when the mind is free, it presents disordered phase changes. The third feature axis captures the lateral directional changes, especially showing a significant left-right alternation pattern when switching between dual tasks. In the temporal lobe area, the left and right hemisphere features are projected asymmetrically. In language processing tasks, it was observed that the projection intensity of the left hemisphere was significantly higher than that of the right hemisphere. This asymmetry was most significant in native language reading and was relatively weakened in foreign language reading. When appreciating music, the projection intensity of the right hemisphere gradually increased, especially in music clips with a strong sense of rhythm.

[0067] Based on the amplitude distribution of local projections, the method marks the feature areas. In the frontal region, the regions are divided into three categories according to the projection amplitude: primary activity area, secondary activity area and background area. For example, when focusing on solving mathematical problems, the primary activity area is concentrated in the forehead midline, showing strong theta wave activity; when the difficulty of the problem increases step by step, the range of the secondary activity area gradually expands, reflecting the increase in cognitive load; when encountering a key breakthrough point, the primary activity area will have a short intensity peak. In long-term continuous calculation tasks, the secondary activity area will periodically alternate with the primary activity area, showing the dynamic deployment process of cognitive resources. The marking of the temporal lobe area reflects the lateralization characteristics of auditory processing. In the speech recognition task, the activity intensity of the left main feature area is significantly higher than that of the right side; when processing multi-person conversations, the left activity area will change dynamically with the switching of speakers; when processing environmental sounds, the difference in activity intensity between the two sides decreases, but it shows stronger temporal synchronization. The markers of the working memory network showed task-related dynamic changes. In the sequence memory task, the node areas of the frontal and parietal lobes showed regular alternating activation; when the sequence length increased, the duration of node activation was prolonged and the intensity increased accordingly; in the spatial working memory task, a persistent activation pattern of the parietal area was observed, accompanied by periodic information exchange with the visual area.

[0068] According to the characteristic distribution of the marked area, the method performs spatial reconstruction. In complex cognitive tasks, corresponding reconstruction strategies are adopted based on the characteristics of the above three types of areas. High-density reconstruction is used for the main activity area of ​​the prefrontal lobe to capture the rapid changes of theta wave activity, and its reconstruction accuracy is dynamically adjusted with the marked activity intensity. For example, in the multi-target tracking task, when the main activity area has an intensity peak, the reconstruction accuracy is increased to the highest; the reconstruction accuracy of the secondary activity area is proportional to the cognitive load level of its mark, and the sampling density is gradually increased when the load increases; low-density reconstruction is used in the background area to maintain computational efficiency. In the emotion processing task, the reconstruction parameters are set based on the asymmetric marking results of the temporal lobe. The left main feature area adopts a higher reconstruction accuracy due to its stronger activity intensity, while the right side is configured with a lower accuracy accordingly. The reconstruction strategy of the working memory network is directly based on the previously marked node activity pattern. Time-varying reconstruction accuracy is adopted in the identified alternating activation area, and a smooth transition of accuracy is achieved on the connecting pathway. For example, in the n-back task, the method dynamically adjusts the reconstruction density according to the marked node activity intensity, maintains a stable reconstruction accuracy in the information maintenance stage, and improves the local reconstruction accuracy at the information update point to capture transient changes. The reconstruction process simultaneously takes into account the spatial characteristics of different brain regions, focusing on signal reconstruction in the depth direction in the prefrontal region and enhancing feature recovery in the horizontal direction in the temporal lobe region.

[0069] Using the spatial distribution of reconstruction, the method generates enhanced features, and the degree of enhancement is directly dominated by the reconstruction accuracy. In the prefrontal region, the main activity area with high-precision reconstruction adopts an adaptive enhancement strategy, and the enhancement strength is proportional to the density of reconstruction points. For example, in driving fatigue monitoring, the method modulates the enhancement amplitude of the θ wave features according to the reconstruction accuracy of the main activity area. The area with higher reconstruction accuracy (such as the middle frontal area) obtains 35% feature enhancement, while the peripheral area with lower reconstruction accuracy only provides 15% enhancement. In the classroom attention monitoring scenario, the degree of feature enhancement of the working memory network changes with the spatial resolution of reconstruction. Strong feature enhancement is provided in the node area with high-precision reconstruction, while a gradual enhancement intensity is used in the transition area to ensure the spatial continuity of the feature. For the language processing features of the temporal lobe, the method designs an enhancement strategy based on the asymmetric reconstruction results. The high-precision reconstruction area of ​​the left hemisphere obtains stronger feature enhancement, thereby better highlighting the activity characteristics of the language-dominant hemisphere. In the multi-tasking scenario, the dynamic adjustment of feature enhancement also strictly follows the distribution of reconstruction accuracy. For example, when the reconstruction accuracy of the visual search task exceeds the threshold, the method increases the feature enhancement strength of the occipital lobe area accordingly; when the reconstruction accuracy of the auditory task is improved, the feature expression of the temporal lobe area is strengthened. This adaptive enhancement mechanism based on reconstruction accuracy not only ensures the accuracy of feature enhancement, but also maintains the structural integrity of spatial features. In long-term working scenarios, the method continuously monitors the reconstruction quality. When the reconstruction accuracy of a certain area changes significantly (such as the signal quality degradation caused by fatigue), the feature enhancement strategy of the corresponding area is adjusted in time to ensure the reliability of the enhanced features. For example, in a continuous cognitive task, if the reconstruction accuracy of the prefrontal lobe area is detected to be reduced, the method will reduce the feature enhancement strength accordingly to avoid incorrectly amplifying the noise features.

[0070] Step S105, performing multi-scale decomposition on the filter group and the enhanced features to obtain multi-scale features, constructing feature combinations according to the multi-scale features, and obtaining time-frequency features corresponding to the feature combinations.

[0071] Specifically, the filter group obtained in the step (local scale Gaussian kernel filter, medium scale bandpass filter and global scale low-pass filter) and the enhanced features obtained in the step (the signal of the main activity area of ​​the frontal lobe is enhanced by 35%, the secondary activity area is enhanced by 20%, the left side of the temporal lobe is enhanced by 40%, the right side is enhanced by 25%, and the working memory network node is enhanced by 30%), and the method is used for multi-scale decomposition. In the frontal lobe area, the local scale uses a Gaussian kernel filter (attenuation rate 0.3, window 2cm) to process the enhanced main activity area signal; the medium scale uses a bandpass filter (attenuation rate 0.5, window 4cm) to process the enhanced secondary activity area signal; the global scale uses a low-pass filter (attenuation rate 0.7, window 6cm) to process the background area signal. In the concentration test, when the attention is focused, the processed local component accounts for 65% of the energy in the central area and 15% in the peripheral area; when the attention is distracted, the central energy drops to 35% and the peripheral energy rises to 45%. For the temporal lobe region, the method applied directional filter banks to the left and right hemispheres to process the enhanced asymmetric signals. The decomposition of the left region with a 40% enhancement showed 7-13Hz activity related to language, and the right region with a 25% enhancement showed 3-7Hz slow wave activity. On the working memory network, the node signal with a 30% enhancement was processed using a connection band filter (width 1.5cm), and the 4-7Hz theta band activity component was obtained. Finally, the characteristic components of three scales, local, medium and global, were obtained, each of which contained the spatiotemporal activity characteristics of a specific frequency band.

[0072] Hierarchical extraction was performed on the above multi-scale components. At the local scale, time-frequency analysis was performed on the central component with an energy share of 65% and the peripheral component with an energy share of 15%, and the characteristics of the attention state were obtained, including the instantaneous intensity sequence of the central theta wave (0.5-2.5μV) and the modulation sequence of the peripheral alpha wave (20%-60%). Spatial clustering was performed on the asymmetric components of the temporal lobe, and the characteristics of three main language processing nodes on the left side (with intensities of 1.8μV, 1.5μV, and 1.2μV) and two secondary nodes on the right side (with intensities of 1.0μV and 0.8μV) were obtained. Connectivity analysis was performed on the 4-7Hz component of the working memory network, and the characteristics of four information transmission pathways were obtained, with connection strengths of 0.75, 0.68, 0.62, and 0.55, respectively. These characteristics showed significant differences in different tasks: when focusing, the central-peripheral intensity ratio remained above 3:1; when performing difficult tasks, the pathway connection strength increased by 30%; when performing language comprehension, the difference in strength between the left and right sides reached 40%. In actual application scenarios, these features show obvious task specificity: in the classroom concentration assessment, the central theta wave intensity continuously exceeding 2.0μV indicates deep engagement; in language learning tasks, the sequential activation pattern of the left node reflects the degree of understanding; in memory tests, the dynamic changes in the strength of pathway connections indicate memory load. Through these analyses, a complete set of hierarchical features was finally obtained, including local attention hierarchical feature sequences (sampling rate 10Hz), regional language processing hierarchical features (spatial resolution 1cm), and network connection hierarchical features (update rate 2Hz).

[0073] Based on the extracted hierarchical features, the method generates multi-scale feature expressions. First, the central theta wave sequence (intensity 1.5-2.5μV) and the peripheral alpha wave sequence (intensity 0.8-1.2μV) are combined into an attention distribution map with a spatial resolution of 1cm and a temporal resolution of 100ms. Then, the five nodes (3 left and 2 right) of the temporal lobe and their activity intensity are integrated into a language processing network, and the phase relationship between the nodes is characterized by the synchronization index (range 0.3-0.8). For the four pathways of working memory, the direction of information flow and transmission efficiency are calculated to obtain a 4×4 connection matrix, and the matrix elements represent the instantaneous state of the pathway. These multi-scale features show dynamic changes in actual tasks: classroom attention monitoring shows that when students are focused, the central intensity of the attention distribution map is stable at more than 2.0μV, and the peripheral alpha wave modulation is reduced to the minimum; during reading comprehension, the left nodes of the language network are activated in sequence, and the synchronization index gradually increases to more than 0.7; in the sequence memory task, the working memory pathway is activated in sequence according to the stimulus sequence, and the connection strength increases with the increase of load. Through long-term data accumulation and statistical analysis, we finally obtained the characteristic distribution law: attention characteristics showed a center-periphery gradient distribution (gradient coefficient 0.4-0.8), language processing characteristics showed task-related lateralization (asymmetry index 0.3-0.6), and working memory characteristics showed load-dependent network reorganization (reorganization coefficient 0.2-0.5).

[0074] According to the above distribution law, the method constructs a feature combination. Dynamic weights are configured for local scale features (central theta wave energy 2.0μV, peripheral alpha wave energy 0.8μV, gradient coefficient 0.4-0.8), and the weight increases to 0.6 when attention is focused and decreases to 0.3 when distracted. The medium-scale features (left temporal lobe average 1.5μV, right 0.9μV, asymmetry index 0.3-0.6) are weighted in a lateralized manner, and the left weight is increased to 0.5 during language tasks. Network dependency weights are set for global scale features (average connection strength 0.65, recombination coefficient 0.2-0.5), and the weight increases dynamically when cognitive load increases. The method adjusts the combination parameters in real time according to signal quality (signal-to-noise ratio threshold 8dB) and task relevance (correlation coefficient threshold 0.7). Through this dynamic combination process, we finally obtained a complete feature combination: local scale combination (center-periphery energy distribution characteristics, phase modulation characteristics), medium scale combination (bilateral language processing node characteristics, functional connection characteristics), and global scale combination (large-scale network synchronization characteristics). Each combination contains corresponding weight configuration and dynamic adjustment mechanism.

[0075] In some embodiments, obtaining the time-frequency features corresponding to the feature combination includes: performing time domain analysis on the feature combination to obtain time series features; performing frequency domain transformation on the time series features according to a preset multi-level transformation strategy to obtain frequency domain statistical results; and constructing the time-frequency features based on the frequency domain statistical results.

[0076] The obtained feature combinations were analyzed in the time domain. Sliding window processing was used for different scale combinations, with a window length of 2 seconds and an overlap rate of 50%. For the local scale combination, a dual-core processing unit was designed: the center detection core (radius 2cm) was used to extract the energy time envelope of the central area theta wave (2.0μV), and the edge detection core (radius 4cm) was used to obtain the inhibitory modulation characteristics of the peripheral area alpha wave (0.8μV). Both detection cores used a sampling rate of 100Hz and an amplitude resolution of 0.1μV. For the medium scale combination, a bilateral analysis module was constructed: the activation sequence was extracted for the left language processing node (1.5μV) and the right node (0.9μV) with a time accuracy of 5ms, and the functional connectivity changes between nodes were calculated at the same time, with a connection update rate of 10Hz. For the global scale combination, a network synchronization analyzer was used to track large-scale network patterns (average connection strength 0.65), and global phase consistency (resolution 2°) and network reorganization features (temporal resolution 50ms) were calculated. For example, when switching cognitive tasks, a rapid change in the energy of the central theta wave from 2.0μV to 1.4μV was observed at the local scale, and a compensatory enhancement of the peripheral alpha wave occurred; at the medium scale, the activation sequence of bilateral nodes was reorganized, and the connection strength had a 200ms adjustment period; at the global scale, a transient decrease and reconstruction of network synchronization was recorded. Methods A family of response curves was established for each type of feature, including energy response curves (dynamic range 40dB), phase response curves (phase range -π to π) and connection response curves (normalized scale 0-1). Through multidimensional analysis, a set of multidimensional temporal features containing timing, polarity and spatial distribution information was finally obtained.

[0077] The frequency domain transformation was performed on the multidimensional time series features, and a multi-level transformation strategy was adopted. The energy time series (100Hz sampling) was short-time Fourier transformed, and the transformation window was selected as 1 second to match the cognitive rhythm characteristics. The window function used an improved Hanning window (the main lobe width was optimized to 1Hz), and the step size was set to 100ms to ensure the time-frequency resolution. The transformation results covered the range of 0.1-30Hz, and overlapping FFT channels were set: 4-7Hz (8 frequency points) in the θ band, 8-13Hz (10 frequency points) in the α band, and 13-30Hz (16 frequency points) in the β band, with an overlap rate of 30% for adjacent channels. The Hilbert transform was performed on the phase time series (200Hz sampling), and the instantaneous phase was extracted by an accurate phase unwrapping algorithm (error <0.1°), and the frequency modulation characteristics were obtained by using an adaptive threshold instantaneous frequency estimator (time resolution 5ms). Methods The phase locking value (16×16 matrix), group delay characteristics (frequency resolution 0.5Hz) and phase-amplitude coupling index were calculated simultaneously. The connection time series (updated at 10Hz) was subjected to wavelet transformation. The Morlet basis wavelet (center frequency 5.5, bandwidth parameter 1.2) was selected, and the decomposition scale was set to an exponential sequence of 1-64. Energy normalization and phase correction modules were configured at each scale. For example, when the cognitive task was switched, the spectrum analysis of the energy time series showed that the energy of the theta band decreased from 45% to 25%, and the energy of the alpha band increased accordingly. The phase time series revealed the phase reorganization process between the prefrontal and parietal lobes. Through these transformations and statistical analysis, the frequency domain statistical results were finally obtained: full-band energy distribution map (frequency resolution 0.5Hz, dynamic range 60dB), phase coupling matrix (64×64, confidence 0.95) and time-varying connection pattern spectrum (number of patterns 32, update rate 5Hz).

[0078] According to the above frequency domain statistical results, the method constructs time-frequency features and adopts a multi-scale synthesis strategy. At the energy feature level, a high-precision time-frequency energy map is constructed. The time axis resolution is inherited from the sliding window (100ms), and the frequency axis is divided into 128 frequency points using a logarithmic scale (0.1-30Hz). Adaptive threshold technology is introduced in the spectrum generation process. The noise floor is determined by wavelet denoising and median estimation (signal-to-noise ratio threshold 10dB), and the signal peak detection adopts an improved maximum suppression algorithm (the suppression radius is adaptively adjusted with the frequency, ranging from 0.5-2Hz). At the phase feature level, a multi-level phase coupling map is established, including an instantaneous phase difference map (resolution 5°, update rate 10Hz), a phase locking value map (resolution 0.05, significance threshold p<0.01) and a cross-band phase coupling strength map (band interval 1Hz). The method performs a Rayleigh statistical test on the phase data, removes insignificant phase relationships (p>0.01), and retains stable phase patterns. At the network feature level, a dynamic connection time-frequency map was formed to describe the reorganization characteristics of the functional network in different frequency bands. Adaptive segmentation processing (segment length 50-200ms) was used to calculate the connection strength matrix (32×32) and network topology indicators (clustering coefficient, path length) in each segment. For example, during the cognitive state transition, the energy time-frequency map showed the energy migration process from theta wave to the alpha wave, which lasted for about 200ms; the phase coupling map captured the short-term synchronization enhancement of the beta band, which lasted for 100-150ms; the network time-frequency map reflected the dynamic reorganization pattern of the large-scale functional network. The method configured a timestamp (accuracy 1ms), a frequency tag (accuracy 0.1Hz), an intensity value (dynamic range 60dB) and a reliability indicator (confidence 0.95) for each feature. Through these processing, we finally get a complete set of time-frequency features, including: energy time-frequency spectrum (signal-to-noise ratio>10dB, 128 frequency points), phase time-frequency relationship (64 coupling modes), network time-frequency dynamics (32 connection modes), and all features have traceable and quantifiable statistical properties.

[0079] Step S106, spatially mapping the time-frequency features to obtain interference patterns corresponding to the time-frequency features, performing feature enhancement according to the interference patterns to obtain feature representations corresponding to the interference patterns.

[0080] Specifically, the obtained time-frequency feature set (including energy time-frequency spectrum, phase time-frequency relationship and network time-frequency dynamics) is spatially mapped. The method first maps the energy time-frequency spectrum to the electrode space, constructs a time-space-frequency stereogram, and establishes signal representation in three dimensions: time, space and frequency. Then, the phase time-frequency relationship is spatially projected to establish a phase propagation map to describe the phase difference and propagation delay characteristics between electrodes. The network time-frequency dynamics are spatially reconstructed to form a spatiotemporal distribution map of functional connections, which contains connection strength and direction information. For example, in the visual attention task, the energy time-frequency feature shows a gamma-band response in the occipital region that changes with the intensity of visual stimulation; in the working memory task, the phase time-frequency feature reflects the theta-band information transmission between the frontal and parietal lobes. Through these spatial mappings, the method identifies three main types of interference patterns: local high-frequency bursts (manifested as transient abnormal enhancement of energy), phase jumps (manifested as mutations in phase relations), and connection breaks (manifested as local destruction of network structure). These interferences show specific spatiotemporal distribution characteristics in different cognitive tasks, and finally form a complete interference pattern spectrum.

[0081] In some embodiments, the time-frequency characteristics at least include an energy time-frequency spectrum, a phase time-frequency relationship, and a network time-frequency dynamics; the spatial mapping of the time-frequency characteristics to obtain the interference pattern corresponding to the time-frequency characteristics includes: mapping the energy time-frequency spectrum to the electrode space to construct a space-time-frequency stereogram for establishing signal representation in three dimensions of time, space, and frequency; spatially projecting the phase time-frequency relationship to establish a phase propagation map for describing the phase difference and propagation delay characteristics between electrodes; spatially reconstructing the network time-frequency dynamics to obtain a space-time distribution map of functional connections, the space-time distribution map including connection strength and direction information; generating an interference pattern spectrum based on the space-time-frequency stereogram, the phase propagation map, and the space-time distribution map to determine the interference pattern based on the interference pattern spectrum.

[0082] Based on the identified interference pattern map, the method constructs a spatial template. For local high-frequency bursts, an adaptive suppression template is designed, and the suppression strength is dynamically adjusted with the interference amplitude, and the spatial coverage range changes adaptively according to the degree of interference diffusion. For the phase jump area, a phase correction template is constructed, and the phase relationship of the surrounding stable electrodes is used for interpolation reconstruction to achieve a smooth transition of the phase. For the connection break, a connection repair template is established, and the pattern is supplemented by combining the historical connection pattern and the surrounding complete connection. These templates are in the form of three-dimensional tensors, containing spatial weights, frequency selectivity, and time modulation characteristics. For example, in the processing of electrooculographic artifacts, the suppression template forms a directional inhibition field in the forehead area, and the intensity decays step by step with distance; in the processing of electromyographic interference, the correction template realizes selective inhibition of the high frequency band. In the process of cognitive task switching, the repair template can effectively maintain the continuity of the functional network and ensure the stable extraction of signal features. The method also optimizes the template parameters for specific scenarios, such as enhancing the directional selectivity of the suppression template in the motion state and optimizing the time response characteristics of the connection repair template in social interaction.

[0083] The constructed spatial template is used to enhance the features. The method adopts a two-layer processing strategy of template matching and feature reconstruction. In the first layer, the suppression template is convolved with the burst region to achieve accurate suppression of interference while maintaining the integrity of the effective signal. In the second layer, the correction template is applied to restore the continuity of the signal, including the smooth transition of the phase relationship and the stable reconstruction of the functional connection. An adaptive threshold mechanism is introduced in the enhancement process, and the threshold parameter is dynamically adjusted according to the local signal-to-noise ratio to ensure the spatial consistency of the enhancement effect. In classroom concentration monitoring, this enhancement strategy can effectively handle a variety of interferences: transient blink artifacts are significantly suppressed, signal drift caused by head movement is corrected, and the connection characteristics of the attention network remain stable. In the driving fatigue monitoring scenario, the method can accurately separate the electromyographic interference generated by driving operations and retain the slow wave characteristics reflecting the fatigue state. In multi-person interaction tasks, the enhanced signal can clearly reflect the neural synchronization pattern between individuals, providing reliable data support for social cognition research.

[0084] Exemplarily, the method of obtaining the feature representation corresponding to the interference pattern includes: constructing a spatial template based on the interference pattern map; the spatial template includes an adaptive suppression template, a phase correction template and a connection repair template; and performing feature enhancement on the spatial template according to a two-layer processing strategy of template matching and feature reconstruction to obtain the feature representation corresponding to the interference pattern.

[0085] Based on the enhancement results, the method generates feature representation. First, a multi-level feature descriptor is constructed, which includes time dimension, frequency dimension and space dimension to achieve comprehensive expression of signal features. Then, feature encoding is performed to convert the enhanced signal into a compact feature vector, and a sparse representation strategy is used to retain the main feature components. The method establishes a feature index structure to support fast retrieval based on time, frequency and spatial position. In practical applications, this feature representation shows excellent performance: in cognitive state monitoring, it can accurately characterize the spatiotemporal characteristics of attention shifting, including the spatial migration of frontal lobe theta wave activity, the phase reorganization of parietal alpha wave, and the dynamic reconstruction of functional connectivity. In the emotion recognition task, the dynamic characteristics of emotion-related neural networks are represented, such as the time-varying pattern of prefrontal asymmetry and the activation sequence of limbic methods. In the process of skill learning, the plastic changes of the motor control network are captured, including the evolution of the coordinated pattern of sensorimotor areas and the regulatory characteristics of cortical excitability. The final feature representation not only has a high signal-to-noise ratio and spatial resolution, but also maintains the physiological significance and temporal continuity of the signal.

[0086] Step S107, performing spatial reconstruction according to the feature representation, performing feature fusion on the reconstruction result, and obtaining spatial enhancement features.

[0087] Specifically, the method performs spatial reconstruction by combining the obtained multi-level feature representation (including high signal-to-noise ratio energy features, phase time-frequency relationships, and network dynamic features). First, a reconstruction mapping matrix is ​​established to map the feature vector back to the electrode space to maintain the spatial continuity of the signal. For the frontal lobe area, a high-precision reconstruction strategy is adopted to restore the attention-related features of the θ band through a local interpolation algorithm; for the temporal lobe area, a lateralized reconstruction method is used to maintain the left-right asymmetry of language processing features; for the parietal lobe area, a network reconstruction is implemented to reconstruct the local connection pattern related to the spatial task. The method configures a quality monitoring module for each reconstructed area to evaluate the reconstruction accuracy and spatial consistency in real time. For example, in attention monitoring, the reconstruction results show the θ wave activity center in the center of the forehead and the surrounding α wave inhibition ring, and the correlation between the spatial distribution and the original features is above 0.85; in the language processing task, the activation pattern of left hemisphere dominance is reconstructed, and the activity intensity ratio deviates from the original feature by less than 10%. In the motor imagery task, the reconstruction process can accurately restore the μ rhythm inhibition pattern of the sensorimotor area and maintain the functional connection characteristics with the premotor area. Through multi-region collaborative reconstruction, a complete spatial activity map was obtained, including local feature distribution, inter-regional connectivity and large-scale network structure.

[0088] In some embodiments, spatial reconstruction is performed based on the feature representation, and feature fusion is performed on the reconstruction result to obtain spatial enhancement features, including: establishing a reconstruction mapping matrix, mapping the feature vector corresponding to the feature representation to the electrode space, and obtaining a spatial activity map; the spatial activity map includes local feature distribution, inter-regional connection relations and large-scale network structure; and feature fusion is performed on the spatial activity map to obtain the spatial enhancement features.

[0089] The reconstructed spatial activity map was subjected to feature fusion, and a multi-level fusion strategy was designed. At the spatial level, the reconstructed features of different brain regions were adaptively weighted and combined, and the weight coefficients were dynamically adjusted according to signal reliability and task relevance. For example, in the concentration assessment, the weight of the prefrontal θ wave activity was adjusted between 0.4 and 0.8 with the change of attention level; in the language comprehension task, the weight of the left temporal lobe would be increased to above 0.6 according to the language processing requirements. At the frequency level, the method fused the activity features of multiple frequency bands to establish the interaction pattern across frequency bands, such as theta-alpha coupling strength, beta-gamma energy ratio, etc. At the temporal level, multi-scale sliding window analysis was used to achieve continuous integration of features, with short windows (200ms) used to capture transient changes and long windows (2s) used to extract steady-state features. In practical applications, this multi-level fusion strategy has shown excellent performance: in working memory tasks, the method integrates the frontal theta wave, parietal alpha wave and phase synchronization features between the two regions, accurately reflecting the dynamic changes of memory load; in the process of motor imagery, it integrates the μ rhythm inhibition of the sensorimotor area and the beta wave changes in the premotor area to achieve early recognition of motor intention. Through this multi-dimensional feature fusion, a set of combined features reflecting brain states is finally obtained, which not only retains the unique attributes of each component, but also reflects the synergistic relationship between them.

[0090] According to the combined features obtained by fusion, the method generates and outputs spatial enhancement features. First, the fused features are standardized to establish a unified cross-modal quantitative standard, including amplitude normalization, phase normalization, and connection strength normalization. Then, feature enhancement is performed, and an adaptive gain control strategy is adopted. The gain coefficient is dynamically adjusted according to feature significance and task relevance. The method sets the enhancement parameters for different regions. The frontal lobe region focuses on the enhancement of the θ band features, the temporal lobe region strengthens the interaction between frequency bands, and the parietal lobe region highlights the network connection features. In continuous monitoring scenarios, this spatial enhancement strategy shows good adaptability: in classroom attention monitoring, it can track the changes in cognitive engagement in real time, and the accuracy of detecting distraction is 90%; in driving fatigue monitoring, by enhancing the slow wave activity features, the occurrence of fatigue state can be warned 30 seconds in advance; in multi-person collaborative tasks, the neural synchronization pattern between team members can be accurately captured, providing a quantitative basis for the evaluation of interaction efficiency. The method also optimizes the enhancement strategy for different application scenarios: in the process of skill learning, the focus is on enhancing the plasticity characteristics that reflect the progress of learning, including changes in cortical excitability and functional network reorganization; in the emotion regulation task, the emotion-related neural circuit characteristics are highlighted, such as prefrontal asymmetry and limbic approach activity; in creative thinking training, the dynamic interaction characteristics of the default network and executive control network are enhanced. Through these targeted enhancement processing, the spatial enhancement features of the final output not only maintain the physiological authenticity of the signal, but also provide outstanding task-related features, achieving accurate characterization and reliable evaluation of brain states.

[0091] The provided method has at least the following beneficial effects:

[0092] 1. This solution establishes a complete electrode mapping and multi-scale filtering system, overcoming the limitation of the limited number of electrodes in smart glasses. By establishing the electrode spatial position model and mapping relationship, performing topological analysis, and determining the appropriate filtering scale, the optimal utilization of limited electrode signals is achieved.

[0093] 2. This solution builds a complete processing chain from signal decomposition to feature enhancement, effectively solving the problem of multi-source interference. By achieving spatial decomposition, component separation, and feature enhancement, and then going through the steps of interference pattern recognition and spatial template construction, the method significantly improves its anti-interference ability while ensuring the integrity of target features.

[0094] 3. This solution achieves dynamic adaptation of feature extraction and enhancement, meeting the needs of different application scenarios. Through multi-scale decomposition and time-frequency analysis, and then through spatial reconstruction and feature fusion, it achieves accurate response to different cognitive tasks and ensures the temporal and spatial consistency of enhanced features.

[0095] In order to implement the smart glasses multi-channel EEG signal spatial filtering and feature enhancement 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 spatial filtering and feature enhancement 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 spatial filtering and feature enhancement device 200 provided in an embodiment of the present application includes:

[0096] The signal acquisition module 201 is used to use the multi-channel EEG signals collected by the smart glasses as original signals, establish an electrode spatial position model according to the original signals, and generate an electrode mapping relationship according to the electrode spatial position model;

[0097] A topological analysis module 202 is used to perform topological analysis on the electrode mapping relationship, determine the spatial structure and the corresponding filtering scale, and generate a filter bank according to the filtering scale;

[0098] A component separation module 203 is used to separate the components of the original signal according to the filter bank, obtain component information, obtain the signal source corresponding to the component information, and construct an independent component sequence according to the signal source; the component information at least includes independent components, cluster components and principal components;

[0099] The component determination module 204 is used to perform spatial mapping on the independent component sequence, determine the main components, construct a weight matrix according to the main components, perform spatial filtering according to the weight matrix, generate a spatial feature sequence, perform spatial reconstruction on the spatial feature sequence, and generate enhanced features;

[0100] A scale decomposition module 205 is used to perform multi-scale decomposition on the filter group and the enhanced features to obtain multi-scale features, construct a feature combination according to the multi-scale features, and obtain time-frequency features corresponding to the feature combination;

[0101] A spatial mapping module 206 is used to perform spatial mapping on the time-frequency feature, obtain an interference pattern corresponding to the time-frequency feature, perform feature enhancement according to the interference pattern, and obtain a feature representation corresponding to the interference pattern;

[0102] The space reconstruction module 207 is used to perform space reconstruction according to the feature representation, perform feature fusion on the reconstruction result, and obtain space enhancement features.

[0103] The above-mentioned spatial filtering and feature enhancement device 200 can implement the smart glasses multi-channel EEG signal spatial filtering and feature enhancement 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.

[0104] 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.

[0105] 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.

[0106] 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.

[0107] 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.

[0108] 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.

[0109] 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.

[0110] 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.

[0111] 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.

[0112] 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 spatial filtering and feature enhancement of multi-channel EEG signals in smart glasses, characterized in that: include: Taking the multi-channel EEG signals collected by the smart glasses as original signals, establishing an electrode spatial position model according to the original signals, and generating an electrode mapping relationship according to the electrode spatial position model; Performing a topological analysis on the electrode mapping relationship to determine a spatial structure and a corresponding filtering scale, and generating a filter bank according to the filtering scale; Separating the original signal into components according to the filter group, obtaining component information, obtaining a signal source corresponding to the component information, and constructing an independent component sequence according to the signal source; the component information includes at least independent components, cluster components and principal components; Performing spatial mapping on the independent component sequence to determine the main components, constructing a weight matrix according to the main components, performing spatial filtering according to the weight matrix to generate a spatial feature sequence, and performing spatial reconstruction on the spatial feature sequence to generate enhanced features; Performing multi-scale decomposition on the filter group and the enhanced features to obtain multi-scale features, constructing a feature combination according to the multi-scale features, and obtaining time-frequency features corresponding to the feature combination; Performing spatial mapping on the time-frequency features to obtain an interference pattern corresponding to the time-frequency features, performing feature enhancement according to the interference pattern to obtain a feature representation corresponding to the interference pattern; The spatial reconstruction is performed according to the feature representation, and the feature fusion is performed on the reconstruction result to obtain the spatial enhancement feature.

2. The method according to claim 1, characterized in that The topological analysis of the electrode mapping relationship to determine the spatial structure and the corresponding filtering scale includes: Acquire the connection strength, connection density and adjacency relationship of each electrode node according to the electrode mapping relationship; Determining a key electrode connection pattern based on the connection strength, connection density and adjacency relationship; Determine the spatial structure and construct a connection feature matrix according to the key electrode connection pattern; The connection feature matrix is ​​divided according to a multi-level clustering strategy to obtain the filtering scale.

3. The method according to claim 2, characterized in that The generating of the filter bank according to the filtering scale comprises: Constructing a parameter template according to the filtering scale; the parameter template at least includes kernel function parameters, spatial coverage, weight distribution and directional weight template; A filter bank is generated according to the parameter template.

4. The method according to claim 3, characterized in that The step of constructing a parameter template according to the filtering scale includes: Obtaining an average connection strength, an inter-region connection density, and a connection strength matrix corresponding to each electrode unit of the smart glasses; On a local scale, a kernel function parameter is generated according to the average connection strength; wherein the higher the average connection strength, the smaller the spatial decay rate of the kernel function parameter; At the medium scale, the spatial coverage and weight distribution corresponding to the filter are adjusted according to the ratio of the connection density between the regions; On a global scale, a directional weight template is constructed according to the eigenvector of the connection strength matrix; in the directional weight template, the direction with a larger eigenvalue of the eigenvector has a higher weight coefficient; The parameter template is generated according to the kernel function parameters, spatial coverage, weight distribution and directional weight template.

5. The method according to claim 1, characterized in that The spatial feature sequence is spatially reconstructed to generate enhanced features, including: Performing local projection on the spatial feature sequence to obtain a feature region; Performing spatial reconstruction according to the feature region; The enhanced feature is generated according to a reconstruction result corresponding to the spatial reconstruction.

6. The method according to claim 1, characterized in that The obtaining of the time-frequency features corresponding to the feature combination includes: Performing time domain analysis on the feature combination to obtain time series features; Perform frequency domain transformation on the time series features according to a preset multi-level transformation strategy to obtain frequency domain statistical results; The time-frequency feature is constructed according to the frequency domain statistical result.

7. The method according to claim 1, characterized in that The time-frequency features at least include energy time-frequency spectra, phase time-frequency relationships, and network time-frequency dynamics; and spatially mapping the time-frequency features to obtain interference patterns corresponding to the time-frequency features includes: Mapping the energy time-frequency spectrum to the electrode space to construct a time-space-frequency stereogram for establishing signal representation in three dimensions: time, space, and frequency; Performing spatial projection on the phase time-frequency relationship to establish a phase propagation diagram for describing the phase difference and propagation delay characteristics between electrodes; Spatially reconstructing the network time-frequency dynamics to obtain a spatiotemporal distribution map of functional connections, wherein the spatiotemporal distribution map includes connection strength and direction information; An interference pattern spectrum is generated according to the space-time-frequency stereogram, the phase propagation diagram and the space-time distribution diagram, so as to determine the interference pattern according to the interference pattern spectrum.

8. The method according to claim 7, characterized in that The obtaining of the feature representation corresponding to the interference pattern includes: Constructing a spatial template according to the interference pattern map; the spatial template includes an adaptive suppression template, a phase correction template and a connection repair template; The spatial template is feature enhanced according to a double-layer processing strategy of template matching and feature reconstruction to obtain a feature representation corresponding to the interference pattern.

9. The method according to claim 1, characterized in that: Performing spatial reconstruction according to the feature representation, and performing feature fusion on the reconstruction result to obtain spatial enhancement features, including: Establishing a reconstruction mapping matrix, mapping the feature vector corresponding to the feature representation to the electrode space, and obtaining a spatial activity map; the spatial activity map includes local feature distribution, inter-regional connection relationship and large-scale network structure; Perform feature fusion on the spatial activity graph to obtain the spatial enhancement feature.

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 9 when executing the computer program.

Images