Method, device and equipment for online adaptive classification of EEG signals in AI glasses

Abstract

The present application relates to the field of brain-computer interface technology, and in particular to a method, device, and apparatus for online adaptive classification of EEG signals in AI glasses. The method comprises: obtaining real-time EEG signals of AI glasses, performing quality inspection as raw data, generating data streams for category analysis, obtaining classification benchmarks, generating classification sequences for sample processing, marking key points for feature extraction, establishing storage sequences, generating replay data sequences, and generating adaptive sequences; performing category statistics based on the classification sequences, obtaining distribution features, marking equilibrium points based on the distribution features, constructing selection rules based on the equilibrium points, generating equilibrium sequences based on the selection rules, generating classification indicators based on the adaptive sequences and equilibrium sequences, performing resource analysis, obtaining resource status for task division, and generating scheduling plans; performing confidence analysis on the scheduling plans, constructing output sequences, and generating classification information. A complete processing flow from raw signal quality control to feature domain construction is realized.

Description

Technical Field

[0001] The present application relates to the field of brain-computer interface technology, and in particular to a method, device and equipment for online adaptive classification of EEG signals in AI glasses. Background Art

[0002] Traditional EEG signal classification methods suffer from significant deficiencies in raw signal processing, lack effective signal quality assessment mechanisms, and their feature extraction processes fail to fully capture the multidimensional characteristics of EEG signals. Existing classification methods also fail to adequately account for the dynamic nature of user status over time and struggle to cope with changes in environmental conditions such as lighting and motion. Furthermore, balancing real-time performance, accuracy, and resource efficiency is a significant challenge on portable devices with limited computing resources.

[0003] In the actual application of AI glasses, users' cognitive states fluctuate across different scenarios, and EEG signals exhibit significant non-stationary characteristics over time, leading to a continuous decline in the performance of fixed classification models. Existing methods lack the ability to recognize novel patterns and perform adaptive adjustments, making it difficult to maintain classification accuracy. Furthermore, the lack of a comprehensive resource scheduling solution and a reliable classification output mechanism hinders the effectiveness of the system in real-world scenarios.

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

[0005] The embodiments of the present application provide a method, device, and apparatus for online adaptive classification of EEG signals in AI glasses. The method aims to solve the problem that in the actual application of AI glasses, the user's cognitive state in different scenarios is changeable, and the EEG signals show obvious non-stationary characteristics over time, resulting in a continuous decline in the performance of fixed classification models. Existing methods are insufficient in the ability to recognize novel patterns and adaptively adjust, making it difficult to maintain classification accuracy. In addition, the lack of a complete resource scheduling solution and a reliable classification output mechanism affects the application effect of the system in actual scenarios.

[0006] In a first aspect, an embodiment of the present application provides an online adaptive classification method for EEG signals in AI glasses, comprising:

[0007] Acquire real-time EEG signals from the AI ​​glasses as raw data, perform quality testing on the raw data, generate a basic sequence based on the quality testing results, and generate a data stream corresponding to the basic sequence; perform category analysis on the data stream to obtain a classification benchmark, and generate a classification sequence based on the classification benchmark;

[0008] Performing sample processing on the classification sequence to mark key points; performing feature extraction on the key points, establishing a storage sequence based on features obtained by the feature extraction, generating a replay data sequence based on the storage sequence, and generating an adaptive sequence based on the replay data sequence;

[0009] Perform category statistics according to the classification sequence to obtain distribution characteristics, mark equilibrium points based on the distribution characteristics, construct selection rules according to the equilibrium points, generate equilibrium sequences according to the selection rules, and generate classification indicators according to the adaptive sequence and the equilibrium sequence;

[0010] Resource analysis is performed according to the classification indicators to obtain resource status, and tasks are divided based on the resource status to build a processing queue. A scheduling plan is generated based on the processing queue. Confidence analysis is performed on the scheduling plan to obtain a credible interval. Classification results are filtered based on the credible interval to construct an output sequence. Classification information is generated based on the output sequence to complete the online adaptive classification of EEG signals in AI glasses.

[0011] In a second aspect, the present application also provides an online adaptive classification device for EEG signals in AI glasses, comprising:

[0012] A signal acquisition unit is configured to acquire real-time EEG signals from the AI ​​glasses as raw data, perform quality inspection on the raw data, generate a basic sequence based on the quality inspection results, and generate a data stream corresponding to the basic sequence; perform category analysis on the data stream to obtain a classification benchmark, and generate a classification sequence based on the classification benchmark;

[0013] a sample processing unit, configured to perform sample processing on the classification sequence and mark key points; perform feature extraction on the key points, establish a storage sequence based on features obtained by the feature extraction, generate a replay data sequence based on the storage sequence, and generate an adaptive sequence based on the replay data sequence;

[0014] a category statistics unit, configured to perform category statistics based on the classification sequence, obtain distribution characteristics, mark equilibrium points based on the distribution characteristics, construct selection rules based on the equilibrium points, generate equilibrium sequences based on the selection rules, and generate classification indicators based on the adaptive sequence and the equilibrium sequence;

[0015] The classification completion unit is used to perform resource analysis based on the classification indicators, obtain resource status, divide tasks based on the resource status to build a processing queue, and generate a scheduling plan based on the processing queue; perform confidence analysis on the scheduling plan to obtain a credible interval, filter the classification results based on the credible interval to build an output sequence, generate classification information based on the output sequence, and complete the online adaptive classification of EEG signals in the AI ​​glasses.

[0016] In a third aspect, the present application also provides a computer device comprising a processor and a memory, wherein the memory is used to store a computer program, and when the computer program is executed by the processor, the online adaptive classification method for EEG signals in AI glasses as described in the first aspect is implemented.

[0017] This method uses AI glasses to collect real-time EEG signals, perform quality checks, remove noise interference, and generate high-quality basic sequences and data streams. The data stream is then analyzed for categories, and classification benchmarks (such as feature space and category attributes) are extracted to generate preliminary classification sequences. Key points in the classification sequence are marked and features extracted, and a storage sequence and replay data sequence are constructed to generate adaptive sequences (such as dynamically adjusting model parameters). The distribution characteristics of the classification sequence are statistically analyzed, and equilibrium points are marked (such as for class imbalance correction). A balanced sequence is constructed, and classification metrics (such as classification rules and discrimination thresholds) are generated in combination with the adaptive sequence. Computational resource requirements (such as priority and reliability) are analyzed based on the classification metrics, task queues are divided, and dynamic scheduling plans are generated. Confidence analysis is used to screen credible results and output final classification information (such as cognitive state recognition).

[0018] By processing EEG signals online, eliminating the need for offline storage, the system meets the low-power and high-real-time requirements of wearable devices. Dynamic model adjustment through replaying data sequences and updating step sizes accommodates individual differences and signal drift (e.g., changes in EEG characteristics due to fatigue). Building balanced sequences based on distributional characteristics addresses class imbalance (e.g., missed detection of minority class samples) and improves classification accuracy. Dynamic resource scheduling (e.g., parallel task partitioning and verification mechanisms) optimizes computing resource allocation, ensuring processing efficiency in complex scenarios. Confidence interval filtering reduces false positives (e.g., eliminating low-confidence noise interference) and improves classification reliability.

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

[0020] Figure 1 This is a flow chart of an online adaptive classification method for EEG signals in AI glasses according to an embodiment of the present application;

[0021] Figure 2 This is a schematic diagram of the structure of the online adaptive classification device for EEG signals in AI glasses shown in an embodiment of the present application;

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

[0023] In the following description, specific details such as specific system structures and techniques are provided for purposes of illustration rather than limitation to facilitate a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application may 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 avoid obscuring the description of the present application with unnecessary detail.

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

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

[0026] As used in this specification and the appended claims, the term "if" can be interpreted as "when" or "upon" 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 "upon determination" or "in response to determining" or "upon detection of [described condition or event]" or "in response to detecting [described condition or event]," depending on the context.

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

[0028] References to "one embodiment" or "some embodiments" in this specification mean that a particular feature, structure, or characteristic described in conjunction with that embodiment is included in one or more embodiments of the present application. Thus, phrases such as "in one embodiment," "in some embodiments," "in other embodiments," and "in other embodiments" appearing in various places in this specification do not necessarily refer to the same embodiment, but rather mean "one or more but not all embodiments," unless otherwise specifically emphasized. The terms "including," "comprising," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0029] The technical solutions of the embodiments of this application are introduced below.

[0030] Traditional EEG signal classification methods suffer from significant deficiencies in raw signal processing, lack effective signal quality assessment mechanisms, and their feature extraction processes fail to fully capture the multidimensional characteristics of EEG signals. Existing classification methods also fail to adequately account for the dynamic nature of user status over time and struggle to cope with changes in environmental conditions such as lighting and motion. Furthermore, balancing real-time performance, accuracy, and resource efficiency is a significant challenge on portable devices with limited computing resources.

[0031] These issues are particularly prominent in the practical application of AI glasses. Users' cognitive states fluctuate across different scenarios, and EEG signals exhibit significant non-stationary characteristics over time, leading to a continuous decline in the performance of fixed classification models. Existing methods lack the ability to recognize novel patterns and perform adaptive adjustments, making it difficult to maintain classification accuracy. Furthermore, the lack of a comprehensive resource scheduling solution and a reliable classification output mechanism hinders the effectiveness of the system in real-world scenarios.

[0032] Please refer to Figure 1 , Figure 1 The flowchart of the method for online adaptive classification of EEG signals in AI glasses provided in the embodiment of the present application is provided. The method for online adaptive classification of EEG signals in AI glasses provided in the 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 and cloud servers. Figure 1 As shown, the online adaptive classification method of EEG signals in AI glasses of this embodiment includes steps S101 to S104, which are described in detail as follows:

[0033] Step S101: Acquire the real-time EEG signal of the AI ​​glasses as raw data, perform quality inspection on the raw data, generate a basic sequence based on the quality inspection results, and generate a data stream corresponding to the basic sequence; perform category analysis on the data stream to obtain a classification benchmark, and generate a classification sequence based on the classification benchmark.

[0034] Specifically, during the acquisition phase, the AI ​​glasses continuously collect EEG signals from the user's scalp surface at a sampling frequency of 1000Hz through a built-in dry electrode array. The electrode array adopts a standard 10-20 system layout, including 9 acquisition sites in the frontal lobe (F3, F4, Fz), central area (C3, C4, Cz) and parietal lobe (P3, P4, Pz). Each acquisition site uses a dry electrode made of gold-plated ceramic composite material, which has a low electrode-skin contact impedance. The electrode signal is initially amplified by a pre-amplifier with an amplification factor of 100 times and has a high input impedance (>100MΩ). The signal of each channel is digitized by a 24-bit analog-to-digital converter to convert the analog signal into a digital signal with a sampling accuracy of 0.1μV. The system synchronously records auxiliary information such as the timestamp, electrode impedance and head movement during signal acquisition. The acquisition circuit integrates a 50Hz notch filter to suppress power-frequency interference, and a 0.1-100Hz bandpass filter to preserve the primary frequency band of the EEG signal. The filter uses a Butterworth structure to achieve a relatively flat passband response. After filtering and digitization, the system outputs a 9-channel raw digital signal sequence in real time.

[0035] The system performs multi-dimensional quality checks on the 9-channel raw digital signal sequence. A sliding window with a 2-second duration and 50% overlap is used to calculate the signal quality index for each acquisition point. The quality index calculation combines three types of features: amplitude features, including peak-to-peak value, root mean square value, and maximum slope; frequency domain features, including energy distribution in the α band (8-13Hz), β band (13-30Hz), θ band (4-8Hz), and δ band (0.1-4Hz), as well as the energy ratio between frequency bands; and time domain features, including zero-crossing rate, signal duration, and number of mutation points. If the peak-to-peak value exceeds ±100μV or the root mean square value falls below 0.5μV, the window is marked as low quality. The system uses a fast Fourier transform to calculate the signal's power spectral density and uses the Shannon entropy formula to calculate spectral entropy to assess signal complexity. Excessively high entropy values ​​indicate signal disorder, while low entropy values ​​indicate a monotonous signal. A detailed analysis of the power spectrum indicates the presence of power frequency interference when the energy contribution of the 50Hz component and its harmonics exceeds 5%. The system monitors energy changes in the high-frequency band (above 30Hz) in real time, calculates short-time energy variance, and identifies it as an electromyographic artifact when the mutation amplitude exceeds 3 times the standard deviation of the baseline and the duration exceeds 100ms. The detection of electrooculographic artifacts is based on the slow wave component analysis of the forehead electrode, and is identified by calculating the low-frequency energy ratio and waveform characteristics. At the same time, the system has established an artifact detection model based on deep learning as an auxiliary means. The model has been trained with a large amount of labeled data and can identify a variety of common artifact types. By weighted fusion of various detection indicators, the system finally outputs the quality detection result matrix for each time window, which contains the quality score and abnormal marking information of each channel.

[0036] The system restores and organizes the original signal sequence based on the quality inspection result matrix. For signal regions with quality scores below the threshold, the system first analyzes the anomaly type and affected range. Power frequency interference is suppressed using an adaptive notch filter, with the filter center frequency fine-tuned based on the actual interference frequency. Short-duration, high-frequency noise is denoised using a wavelet transform, employing the db4 wavelet basis function for a five-layer decomposition. High-frequency coefficients are processed using a soft thresholding method. When the signal is completely lost or severely contaminated, spherical spline interpolation is used for reconstruction. This interpolation process considers the temporal continuity of adjacent valid data and introduces spatial constraints to ensure the rationality of the reconstructed signal. A weight matrix is ​​constructed based on the actual physical distance between electrodes, with weights exponentially decaying with increasing distance, typically with an attenuation coefficient of 0.2. This is used to create a spatial filter to suppress common-mode interference. The signals of the nine channels are strictly aligned to a common time axis, and linear interpolation is used to address sampling asynchrony, ensuring time alignment accuracy of multi-channel data better than 0.5 ms. The processed signal retains the quality score and processing record, including the original quality score, anomaly type identifier, selected processing method, and processing parameter settings. The system divides the processed signal into 1-second time segments, with adjacent segments overlapping by 0.5 seconds, and each segment contains 1000 sampling points. Characteristic metrics are calculated for each time segment, including basic statistics (mean, standard deviation, skewness, kurtosis), absolute and relative energy in each frequency band, inter-channel correlation coefficients and phase synchronization metrics, and signal complexity metrics (sample entropy and permutation entropy). Based on these processing results, the system ultimately generates a base sequence, which includes quality-controlled multi-channel signal data, a complete processing record chain, and a rich set of characteristic metrics.

[0037] In some embodiments, generating the data stream corresponding to the basic sequence includes: performing feature division on the basic sequence and marking feature points; constructing a feature domain based on the feature points; and generating the data stream according to the feature domain.

[0038] The above steps perform feature processing on the base sequence output from the previous steps. The multi-channel signal data, processing records, and feature index set contained in the base sequence provide the basis for feature partitioning. The system employs a hierarchical strategy for feature partitioning, dividing features into three dimensions: time domain, frequency domain, and spatial domain. Time domain features include waveform fluctuations, duration, and amplitude variation. By calculating the first-order and second-order difference sequences of sampling points, the waveform change rate and acceleration are determined. For example, when a user focuses on a flickering stimulus on the screen, a stable visual evoked potential (VEP) appears in the EEG signal. This response manifests as waveform fluctuations at a specific frequency in the time domain. Frequency domain features focus on the rhythmic activity and spectral characteristics of the signal. Wavelet transforms are used to perform time-frequency decomposition of the base sequence. Morlet wavelet functions are used to perform multi-scale analysis within the 0.1-100 Hz range, with a scale interval of 0.5 Hz. Spatial features are based on inter-channel correlations. A spatial feature network is constructed by calculating the correlation coefficient matrix, phase locking value (PLV), and Granger causality strength between channel pairs. The system normalizes the features at the three levels and clusters similar features using a density-based clustering algorithm (DBSCAN), with a cluster radius of 0.15 and a minimum sample size of 5. These processes ultimately output a feature partitioning result with a clear hierarchical structure, including feature category labels and partition boundary information for each time window.

[0039] Feature points are marked based on the category labels and boundary information in the feature partitioning results. For temporal features, the system uses a local extrema detection algorithm to identify peaks and valleys within each partition, setting a minimum peak-to-peak interval of 100ms and an amplitude threshold of 1.5 standard deviations of the local mean. For example, when a user performs a motor imagery task, characteristic μ-rhythm suppression occurs in the motor regions of the brain. This phenomenon manifests in the temporal domain as a series of waveforms with gradually decreasing amplitudes. The system tracks these amplitude changes to mark the start and end of motor imagery. For each detected peak and valley, characteristic parameters such as duration, waveform asymmetry, and peak power are calculated. For frequency features, energy mutation points are marked within each frequency band, and the start and end of rhythmic activity are identified by calculating the short-term energy ratio, with a threshold of 2.5. For spatial features, the system calculates inter-channel synchronization metrics, marking moments when synchronization strength changes significantly, with a threshold of 30%. In practice, when a user shifts attention, the synchronization patterns between different brain regions change significantly. The system tracks these changes in synchronization strength to track dynamic changes in attention. Each feature point is assigned a three-dimensional feature vector containing its time-domain, frequency-domain, and spatial eigenvalues. The vector elements undergo MinMax normalization to ensure that their values ​​fall within the [0, 1] range. All landmarks are constructed into a directed graph structure based on their temporal relationships and feature similarities. Edge weights are determined by the Euclidean distance between the feature vectors. This process ultimately outputs a feature point network containing complete spatiotemporal information, where nodes represent feature points and edges indicate the strength of associations between them.

[0040] The system uses a network of feature points and their correlation strength information to construct a feature domain and generate a data stream. First, a spectral clustering algorithm is applied to the feature point network, partitioning the entire network into multiple subgraphs, each representing a feature domain. In real-world applications, for example, when users perform continuous cognitive tasks, brain activity exhibits distinct characteristic patterns, forming relatively independent clusters in the feature space. The system partitions the feature domain based on this clustering structure. The number of clusters is adaptively determined by the silhouette coefficient of the eigenvalue distribution, typically ranging from 5 to 10. A kernel-based density estimation method is used within the feature domain to construct a continuous feature distribution. The kernel function uses a Gaussian kernel, and the bandwidth parameter is determined through cross-validation. To generate the data stream, the system designs an adaptive sampling strategy within the feature domain. The sampling interval is inversely proportional to the feature gradient, and the sampling density can reach 5 times the base sampling rate in areas with sharp feature changes. The sampled data points are interpolated using cubic spline interpolation to generate a continuous data stream, with the number of interpolation nodes being 1.5 times the number of original sampling points. Each time point in the data stream contains complete feature information for the current feature domain, and the data dimension matches the dimension of the feature vector. Using compressed sensing, the data stream is compressed to a ratio of 3:1, while maintaining a reconstruction error of less than 5%. The resulting data stream sequence not only retains the key features of the original signal but also enables efficient data transmission and storage.

[0041] In some embodiments, generating the classification sequence according to the classification benchmark includes: constructing a feature space based on the classification benchmark; marking category attributes according to the feature space, and generating the classification sequence according to the category attributes.

[0042] First, the compressed data stream output from the above steps undergoes category analysis. The system extracts the feature value sequence from the data stream, decompresses it and restores it using a 3:1 compression ratio, reconstructing the complete time series features. Time series analysis is then performed on the decompressed data stream, calculating the feature change rate and trend between adjacent time windows and identifying stable and fluctuating segments within the data stream. For example, in an attention classification task, the system extracts dynamic variations in the frontal lobe theta band energy and the parietal lobe alpha band energy from the restored data stream. Based on the time series features in the data stream, the system uses a hierarchical clustering approach to identify naturally occurring feature categories and calculates inter- and intra-class distance ratios to evaluate clustering effectiveness. When a user uses AI glasses for work, the system identifies characteristic patterns corresponding to different cognitive tasks, such as word processing, web browsing, and video conferencing, from the decompressed data stream. For each identified feature category, the system analyzes the reconstructed sample distribution within the data stream and calculates its statistical characteristics, including the mean vector, covariance matrix, and distribution density. Prior constraints are established based on the physiological characteristics of the EEG signals extracted from the data stream, such as the symmetry of motor areas in a motor imagery task. Cross-validation is used to evaluate the classification performance of different feature combinations in the data stream, and the subset of features with the strongest discriminative power is selected as the classification benchmark. The system uses a dynamic threshold strategy to adaptively adjust the decision boundary based on changes in the feature distribution, thereby obtaining a classification benchmark that includes feature selection schemes, decision rules, and threshold parameters.

[0043] The system constructs a feature space based on the classification benchmark obtained in the previous stage. Using the feature selection scheme specified in the benchmark, the system selects the most discriminative feature subset from the original feature dimensions for mapping. For each category sample that meets the benchmark's judgment rules, the system uses the threshold parameters specified in the benchmark to construct a radial basis function mapping, projecting the selected key features into a more discriminative feature space. As the user visits the museum, the system uses the benchmark's feature selection scheme to map the different neural network features activated by appreciating artworks and reading captions to different regions of the feature space. The system strictly adheres to the judgment rules established in the benchmark, constructing decision boundaries between categories based on the selected feature subsets. The boundary parameters are determined by optimizing inter-class distances and intra-class compactness. Within the dynamic threshold range specified in the benchmark, the system constructs a state transition network, using threshold parameters to describe the transition probabilities and patterns between categories. For example, in the reading task, the system maps periodic fluctuations in the user's attention level into specific transition paths in the feature space according to the benchmark's judgment rules. Density analysis identifies high-density and sparse regions that meet the benchmark's requirements, providing a spatial reference for subsequent category attribute labeling. The resulting feature space contains the projection matrix, decision boundary, and category distribution information built based on the classification benchmark.

[0044] The feature space constructed in the previous stage is used to label data samples with their category attributes. The system first projects each sample into a high-dimensional space using the projection matrix defined in the feature space. Then, based on a predefined decision boundary, it calculates the distance between the sample and the center of each category. The system then analyzes the projected sample's trajectory, referring to the category distribution information in the feature space, to identify state transition points and sustained stable intervals. In practice, as a user attends to stimuli in different regions of the display screen, the system projects EEG features into a high-dimensional space using the projection matrix in the feature space and uses the predefined decision boundary to determine the distribution of attention. For example, the P300 component in the frontal and parietal regions, after being mapped according to the projection matrix, systematically shifts with changes in the target of attention. For each projected sample, the system calculates a confidence metric based on its position in the predefined feature space and the decision boundary, including distance from the boundary and local density characteristics. For samples that fall near the decision boundary, the system retains a certain degree of uncertainty based on the distribution characteristics of the feature space to avoid arbitrary judgments. The system dynamically updates the category feature template based on the actual distribution of the sample in the feature space. The final sample labeling result fully records the category label, location information in the feature space, attribution probability and confidence index.

[0045] Based on the sample labeling information obtained in the previous stage, the system begins generating a classification sequence that reflects changes in cognitive state. Specifically, the system uses the category labels and timestamp information in the labeling results to establish an initial sequence and optimizes it based on the attribution probabilities and confidence metrics in the sample labels. The system uses a sliding window method to smooth category decisions with similar confidence levels, with the window length dynamically adjusted based on the stability of the labeled samples. A hierarchical classification structure is constructed based on the probability distribution of the sample labels, clustering samples with similar characteristics into main categories. Subcategories are then divided based on the detailed features of the labels. The system uses the temporal correlations in the labeling results and the position of the samples in the feature space to perform a spatial-temporal consistency check on the classification results, ensuring that samples with similar time and location have coherent category attributes. The system calculates statistical characteristics for each category in the classification sequence, including frequency of occurrence, duration, and transition patterns. In classroom learning scenarios, the system can distinguish whether students are actively thinking or passively receiving information by analyzing the probability distribution characteristics of the sample labels. The system employs differentiated classification strategies for varying confidence levels in the labeled results, maintaining stable results in high-confidence intervals and performing probabilistic classification in low-confidence intervals. The system assesses the uncertainty of the classification results by calculating the entropy of the labeled samples and adjusts the sampling strategy accordingly. The resulting classification sequence integrates all key information from the sample labels and encompasses the complete evolution of the categories.

[0046] Exemplarily, generating an adaptive sequence based on the replay data sequence includes: performing novelty analysis on the replay data sequence to obtain change characteristics; marking an adjustment interval based on the change characteristics; determining an update step according to the adjustment interval; and generating an adaptive sequence according to the update step.

[0047] The classification sequence output from the above steps undergoes sample processing. The system first conducts an in-depth analysis of the sequence's category labels, temporal information, confidence metrics, and the complete category evolution process. Based on the category evolution characteristics recorded in the classification sequence, key points in the sample sequence are identified, including category transition points, steady-state intervals, and abnormal fluctuation points. The system analyzes each key state in the evolution process and calculates its duration and transition characteristics in the classification sequence. For each identified key point, a sliding window matching the sequence is used to analyze its context. The window length is adaptively adjusted within a range of 250-1000ms based on the state characteristics recorded in the classification sequence. A dynamic feature description of the key point is established by calculating the feature change rate and state transition probability in the classification sequence. For example, when a user switches from focused document editing to video conferencing, the system extracts characteristic manifestations of the state transition point from the classification sequence, including changes in the prefrontal theta band power and a significant decrease in the alpha band ratio. These characteristic changes accurately correspond to the cognitive task switching process in the classification sequence. The system also determines the reliability level of each key point based on the confidence distribution of the classification sequence, focusing on analyzing and verifying points with lower confidence. The system sets dynamic detection thresholds based on the categorical features in the classification sequence, focusing on the corresponding feature combinations in different scenarios. The result is a set of labeled keypoints containing spatiotemporal features and contextual information, each of which directly corresponds to a state transition in the classification sequence.

[0048] The system extracts features from a set of marked keypoints. The system first processes the spatiotemporal features of the keypoints. For each keypoint, based on its representation in the classification sequence, it calculates multi-scale temporal features, including peaks, valleys, zero-crossing rates, amplitude variance, and various statistical moments of the EEG waveform within the corresponding time period. The spatial dimension of these spatiotemporal features is extracted by analyzing the multi-channel signal relationships at the keypoints, calculating the strength of inter-channel interactions reflected in the classification sequence. The system also leverages the contextual information of the keypoints to analyze state trends in preceding and following time windows, extracting gradual features reflecting cognitive state transitions. In the frequency domain, the system analyzes the dynamic energy patterns of various frequency bands within the time window corresponding to the keypoints, focusing specifically on characteristic bands associated with cognitive states identified in the classification sequence. In a museum scenario, when a user transitions from appreciating an artwork to reading a caption, the system extracts the characteristic pattern of that keypoint, capturing the significant transition from occipital alpha waves to frontal and parietal beta waves recorded in the classification sequence. This feature extraction process is directly based on the categorical variation characteristics recorded in the classification sequence, effectively capturing the neural signatures of two distinct cognitive activities: visual appreciation and text processing. Appreciation of artwork exhibits a dominant alpha wave activity, reflecting a relaxed and immersed state; whereas, reading text exhibits enhanced beta waves, indicating increased cognitive processing. The system normalizes and de-noises the extracted features, employing independent component analysis to remove electrooculographic and myoelectric artifacts. A template library is constructed based on the cleaned features. Each template contains the typical characteristic patterns and variation ranges exhibited by keypoints of that category in the classification sequence. The system maintains the template library through an adaptive update mechanism, regularly updating it based on emerging characteristic patterns in the classification sequence. The system outputs a processed feature set, with each feature vector directly corresponding to a state feature in the classification sequence.

[0049] Based on the processed feature set, the system constructs a storage sequence. This storage sequence adopts a hierarchical structure that fully corresponds to the feature extraction hierarchy: the bottom layer stores the raw feature data extracted from key points, the middle layer stores feature statistics and category information, and the top layer maintains temporal associations and state transition records. The system uses differential coding to compress feature data, recording only the amount of change for continuously changing features identified in the feature set. A complete record is used for sudden changes and key events extracted from key points. The system performs cluster analysis on the features in the storage sequence, automatically determining the number of clusters using a density peak algorithm and defining cluster feature centers and boundaries. During extended reading, the system constructs a storage sequence based on the extracted feature set, recording cyclical fluctuations in attention. This pattern reflects the cyclical changes in theta waves extracted from the feature set, as well as the fluctuations in alpha wave energy in the temporal and parietal regions. These fluctuations are then correlated with reading comprehension efficiency. For example, indicators in the feature set show that increased alpha wave activity is often associated with decreased reading speed and increased depth of comprehension; whereas, a dominant beta wave pattern indicates rapid scanning and information retrieval. The system optimizes the storage structure based on the temporal and spatial distribution characteristics of the feature patterns in the feature set to ensure efficient data organization and retrieval. The system outputs a compressed and encoded storage sequence that fully preserves the data, cluster information, and temporal relationships in the feature set.

[0050] The system generates replay data based on the compressed stored sequence. The replay process first interpolates the stored sequence data, reconstructing a continuous signal using a cubic spline function to ensure that the restored data is consistent with the original features in the stored sequence. The system incorporates a multi-scale replay mechanism that reproduces the cognitive state changes recorded in the stored sequence at different time scales. The replay data retains key features and categorical information from the stored sequence, enhancing analyzability. The system monitors signal quality in real time, evaluating the reconstructed signal's signal-to-noise ratio, feature consistency, and spatial correlation against benchmarks in the stored sequence. An adaptive optimization algorithm dynamically processes the replay data, adjusting replay parameters based on the feature distribution in the stored sequence. For example, in an office setting, the replay data fully reproduces the user's cognitive state changes throughout the day as recorded in the stored sequence: During the morning workday, high beta wave activity is observed, reflecting a state of deep thought; after noon, theta waves increase, indicating a decrease in attention; and during the afternoon meeting, alpha and beta waves alternate, reflecting a shift in attention between active engagement and passive reception. Replay at different time scales restores cognitive patterns based on the hierarchical structure of stored sequences: minute-level replay reflects task switching characteristics, hour-level replay reveals the accumulation of fatigue, and day-level replay reveals overall fluctuations in work efficiency. The system supports various replay modes, including fast, slow, and selective replay. It calculates reconstruction error, feature fidelity, and temporal consistency based on quality metrics in the stored sequences. The system ultimately outputs a quality-verified replay data sequence that contains the complete state transition process and quality assessment results.

[0051] The replay data sequences output from the above steps are analyzed for novelty. The system first uses the quality assessment results of the replay sequences to screen valid data segments. Reliable samples are identified based on quality metrics such as reconstruction error (<5%), temporal consistency (>0.8), and feature fidelity (>0.9). State transitions are extracted from qualified replay data, including features such as high beta wave activity during morning work, increased theta wave activity after noon, and alternating patterns of alpha and beta waves during afternoon meetings. Newly emerging characteristic patterns are identified by calculating the differences between the feature vectors in the replay data and these known patterns. These differences are primarily manifested in three dimensions: rhythmic changes in time, shifts in energy distribution in frequency, and changes in channel correlation patterns in space. In practical applications, when a user switches from traditional text reading to mixed reality reading using AR-enhanced displays, the system compares the state changes in the replay data and identifies feature combinations that differ significantly from known patterns, particularly distinctive synergistic patterns in the alpha and prefrontal theta bands in the visual cortex. The system performs a multi-level analysis of these differential features to establish a novelty assessment system that encompasses temporal stability, spatial consistency, and spectral characteristics. Temporal stability reflects the reproducibility of features in the replayed data at different time scales, spatial consistency describes the distribution of features across multiple brain regions in the replayed sequence, and spectral characteristics characterize the composition of EEG rhythms in the reconstructed signal. The system ultimately outputs a set of changing features that describe the novel characteristics of the replayed data. Each feature includes quantitative indicators of its significance level, frequency of occurrence, spatial distribution characteristics, and task relevance.

[0052] Based on the set of changing features, the system marks adjustment intervals. The system analyzes the impact of each novel feature in the replay data, including the time span, intensity distribution, and spatial coverage recorded in the set. Based on the significance level of the changing features, the system establishes a hierarchical threshold system, categorizing feature changes into three levels of intensity: significant (significance > 0.8), moderate (significance 0.5-0.8), and subtle (significance < 0.5). When multiple features in the set overlap in time or space, the system evaluates their interactions based on the significance level and frequency of each feature, identifying dominant and subordinate features. For gradually changing features in the set (such as the slow decay of alpha wave energy), the system uses a larger adjustment interval (typically > 2 seconds) to capture the complete change process. For abrupt features (such as the sudden increase in beta waves), a local fine-grained interval (typically < 500 ms) is used to accurately locate the moment of change. In a museum setting, when users view AR explanations of artworks through AI glasses, the system accurately marks the intervals where users transition from simple visual appreciation to deep understanding based on the spatial distribution of the change feature set. The system also determines the spatial propagation pattern of feature changes based on the task relevance of the change features and establishes an interval optimization mechanism. By merging adjacent weak change intervals within the change feature set and segmenting complex mixed change intervals, a multi-dimensional set of adjustment intervals is ultimately formed, each with clear boundary definitions, feature attributes, change levels, and adjustment priorities.

[0053] The system uses a set of adjustment intervals to determine the step size for feature updates. The system first uses the boundaries of the adjustment intervals to define the precise time window within which updates are performed. It then selects appropriate update dimensions and parameters based on the feature attributes. A baseline step size is calculated based on the level of change in each interval: significant change intervals use the smallest step size (typically 0.2 times the baseline value) to ensure accurate capture of feature changes; moderate change intervals use a medium step size (typically 0.5 times the baseline value); and subtle change intervals use a larger step size (typically 1 times the baseline value). The system also allocates computing resources based on the priority of the adjustment intervals, with high-priority intervals receiving more frequent updates and finer step size control. In rapidly changing cognitive scenarios, the system employs an adaptive step size strategy based on the level of change in the adjustment intervals: for significant change intervals, 50% of the current step size is initially used to quickly adapt to the new pattern, then gradually increased to the original size based on the learning curve; for moderate change intervals, the step size is maintained constant; and for subtle change intervals, the step size is gradually increased to 150% of the original size. The system dynamically adjusts the step size by monitoring the changing trends in classification accuracy within each adjustment interval: when accuracy steadily increases, the current step size is maintained; when accuracy fluctuates, the step size is reduced by 20%; and when accuracy continuously decreases, updates are paused and the system reverts to the previous stable state. A multi-scale update strategy is designed for different feature dimensions within the set of adjustment intervals: larger step sizes are used for rapidly changing surface features (such as transient alpha wave suppression) and smaller step sizes are used for stable deeper features (such as theta wave rhythm). Independent step size parameters are set for each adjustment interval, enabling refined update control. Ultimately, a complete update scheme is output, comprising a baseline step size, adaptive rules, and protection strategies.

[0054] Based on the update scheme, the system generates an adaptive sequence. First, based on the step size parameters in the update scheme, a mapping relationship for feature updates is constructed: for intervals of significant change, the system uses the minimum step size (0.2 times the baseline value) to update features every 100ms; for intervals of moderate change (0.5 times the baseline value), updates occur every 250ms; and for intervals of weak change, the system uses a larger step size (1 times the baseline value) to update features every 500ms. The system employs a progressive update strategy, applying different update rates to features at different levels based on the adaptive rules in the update scheme. In an office environment, as users gradually adapt to the multi-screen collaborative work mode provided by AI glasses, the adaptive sequence, based on the step size settings in the update scheme, accurately reflects the user's cognitive model evolution from initial adaptation to mastery. This shift is reflected in the activity patterns of the prefrontal executive control network, which gradually shift from initial high-intensity activity to a more efficient collaborative mode. The system establishes a dynamic feature importance assessment mechanism and allocates update resources to different features based on the step size allocation strategy in the update scheme. For feature patterns with long-term stability, the system uses the protection strategy in the update scheme to maintain their core characteristics during large-step updates. To address rapidly changing task characteristics, the system employs adaptive rules within its update scheme, making rapid adjustments within a small step size. An online verification mechanism continuously evaluates the effectiveness of the updates, including classification accuracy, generalization, and interference resistance. The resulting adaptive sequence maintains model stability while ensuring adaptability to new patterns.

[0055] Step S102: perform sample processing on the classification sequence and mark key points; perform feature extraction on the key points, establish a storage sequence based on the features obtained by feature extraction, generate a replay data sequence based on the storage sequence, and generate an adaptive sequence based on the replay data sequence.

[0056] Specifically, category statistics are performed on the classification sequence output from the above steps. The system analyzes the category label distribution, state transition records, confidence changes, and the complete category evolution process recorded in the sequence. The temporal distribution characteristics of each category label in the sequence are calculated, including the frequency and duration of occurrence at different time scales. State transitions during the category evolution process are analyzed, and a conditional probability matrix based on the actual transition records is constructed to quantify the regularity of transitions between states. Combined with the confidence indicators in the classification sequence, the characteristic differences in category distributions at different confidence levels are analyzed, and a corresponding relationship between confidence and distribution patterns is established. The system calculates the actual distribution density of categories in the feature space and identifies areas of category clustering and sparseness based on the transition records in the classification sequence. For example, in a long-term learning task, the system discovered a negative correlation pattern between focused and fatigued states in the classification sequence, as well as co-occurrence characteristics between mild fatigue and distracted states. Based on these co-occurrence patterns reflected in the sequences, the system constructs a category association graph to quantitatively describe the strength of the mutual influence between categories. Through multi-dimensional statistical analysis, the system outputs a distribution feature set including sample frequency distribution, state transition probability, temporal change pattern, spatial density distribution and category association strength.

[0057] Based on the distribution feature set obtained in the previous stage, the system identifies equilibrium points. First, the system uses frequency distribution features to identify the main activity intervals for each category. Based on the sample distribution in the distribution feature set, an adaptive threshold is set to identify potential equilibrium locations at the intersection of these intervals. Based on state transition probability data, the system identifies the main path of category transitions and marks key transition points along these high-frequency transition paths. By analyzing temporal variation patterns, the system uses autocorrelation analysis to identify periodicity in the distribution features and marks temporal equilibrium points during the stable phase of cyclical variation. Using spatial density distribution information, the system identifies spatial equilibrium points at the boundaries and overlapping regions of the category distribution. Based on category association strength data, the system identifies stable regions where multiple categories coexist. When the distribution features indicate that the co-occurrence probability of two or more categories exceeds a threshold, the region is marked as a significant equilibrium location. In a museum scenario, when observing users transition between focused appreciation and browsing, the system accurately locates equilibrium points for state transitions based on metrics from the distribution feature set. The system calculates a stability score for each candidate equilibrium point, including local variance, duration, and recurrence frequency. By comprehensively analyzing the indicators of various dimensions of distribution characteristics, the system finally determines a set of equilibrium points, each of which contains its spatial location, temporal characteristics, stability score and associated category information.

[0058] Based on the set of equilibrium points identified in the previous stage, the system constructs selection rules. The system treats each equilibrium point as a decision node, defines a spatial decision boundary based on its location attributes, sets state persistence constraints based on temporal characteristics, uses stability scores to determine transition thresholds, and prioritizes state transitions based on associated category information. For equilibrium points with high stability scores and few associated categories, the system sets higher transition thresholds to maintain state stability. For equilibrium points located at the intersection of multiple categories, the thresholds are lowered based on their characteristics to facilitate flexible switching. For example, in an office environment, when a user needs to switch between tasks such as document editing and video conferencing, the system can adjust the switching strategy based on the specific characteristics of the equilibrium point. Based on the set of equilibrium points, the system constructs location-based regional rules, time-based sequencing rules, stability-based threshold rules, and association-based transition rules. Specific rule templates are developed for different application scenarios. Ultimately, a multi-dimensional selection rule set is formed, containing clear judgment conditions and execution strategies.

[0059] Using the selection rule set established in the previous stage, the system generates a balanced sequence. First, regional rules are applied to spatially reorganize samples, strictly adhering to the spatial decision boundaries defined in the rule set to ensure spatial balance in the class distribution. Sequential rules are then used to adjust the temporal pattern of state transitions, ensuring smooth transitions based on the state persistence constraints defined in the rule set. The system then filters unstable transitions based on threshold rules, retaining only state changes that meet the stability requirements defined in the rule set. Finally, transition rules are used to optimize the switching paths between classes, achieving dynamic balance in the overall distribution according to the priorities defined in the rule set. In practical applications of AI glasses, this multi-rule collaborative approach can achieve overall balance in the class distribution while maintaining the natural evolution of cognitive states during long learning tasks. The system implements multi-level balance control, including local, regional, and global balance. The spatiotemporal distribution characteristics of the sequence are continuously monitored, and multi-dimensional evaluation metrics, including distribution uniformity, transition smoothness, and stability, are established. The resulting balanced sequence achieves equilibrium across multiple dimensions, including spatial distribution, temporal characteristics, and class relationships.

[0060] Step S103: perform category statistics based on the classification sequence to obtain distribution characteristics, mark equilibrium points based on the distribution characteristics, construct selection rules based on the equilibrium points, generate equilibrium sequences based on the selection rules, and generate classification indicators based on the adaptive sequence and the equilibrium sequence.

[0061] Specifically, the system first performs a feature combination analysis on the adaptive sequence and the equilibrium sequence of the aforementioned steps. Dynamic features such as update step size, adaptability, and novelty are extracted from the adaptive sequence. These features reflect the model's responsiveness to new patterns. Equilibrium features such as equilibrium point location, stability indicators, and transition rules are obtained from the equilibrium sequence, reflecting the reliability of the class distribution.

[0062] In some embodiments, generating classification indicators based on the adaptive sequence and the balanced sequence includes: combining features based on the adaptive sequence and the balanced sequence to construct a classification chain; extracting attributes based on the classification chain; establishing discrimination rules based on the attributes obtained by the attribute extraction, and generating classification indicators based on the discrimination rules.

[0063] The system establishes a sample-level feature matching mechanism, integrating the features of the two sequences through time alignment. It employs a multi-level feature fusion strategy: at the sample level, after aligning timestamps, the updated features of the adaptive sequence are vector-concatenated with the balanced features of the balanced sequence; at the window level, the combination pattern is calculated by combining the adaptive sequence's step size parameter and the balanced sequence's stability index; and at the sequence level, the coupling relationship between the changing trends of the adaptive features and the balanced features is analyzed. For example, when a user learns a new interaction method through AI glasses, the system combines the learning curve in the adaptive sequence with the state distribution in the balanced sequence to form a complete behavioral profile. The system calculates a feature correlation matrix based on the specific indicators of the two sequences and uses the mutual information criterion to select the most discriminative feature combinations. For dynamic features, the system analyzes their variation patterns under different step sizes in the adaptive sequence, while also considering the equilibrium constraints in the balanced sequence. Ultimately, a classification chain structure is constructed that incorporates temporal correlations, feature weights, combination patterns, and evolutionary laws.

[0064] Based on the constructed classification chain structure, the system extracts attributes. First, the temporal correlation information in the classification chain is used to determine the time window for feature analysis. Extraction priorities are set based on the feature weights recorded in the chain. For each temporal correlation node, the system extracts local statistical features based on the combination patterns in the classification chain. These include basic statistics such as mean, variance, and kurtosis, as well as higher-order features such as autocorrelation coefficient and conditional entropy. The system analyzes the synergistic effects of features based on the evolutionary patterns in the classification chain and calculates the joint distribution characteristics of feature combinations. In a museum setting, when the system detects changes in user attention patterns through the classification chain, it can quickly locate key feature change points based on the weights and associations recorded in the chain. The system evaluates the extracted attributes in multiple dimensions: attribute stability is assessed based on the temporal characteristics in the classification chain, discrimination is assessed based on feature weights, and reliability is assessed using combination patterns. The system uses hierarchical clustering to group attributes, ensuring that attributes within each group have similar functional characteristics. Ultimately, the system outputs a complete attribute set that includes statistical features, dynamic characteristics, association structure, and stability scores.

[0065] The system establishes discrimination rules based on the extracted attribute set. Basic classification thresholds are set based on the statistical features of the attribute set, incorporating dynamic characteristics into the threshold's adaptive adjustment mechanism. For example, when detecting the user's attention distribution characteristics in the AR environment, the system sets a baseline threshold based on the α / β wave ratio in the statistical features. It also adjusts the threshold boundary using the volatility of the dynamic characteristics, making discrimination more flexible. A multi-level decision tree is constructed based on the attribute association structure, with attributes of varying stability levels used at different levels for discrimination. Highly stable attributes such as prefrontal theta wave power are used at the top level for coarse classification, while highly discriminative but highly volatile attributes such as transient visual evoked potentials are used at the bottom level for fine classification. For each decision node, the system calculates the optimal segmentation threshold based on specific indicators in the attribute set, optimizing the decision boundary by minimizing classification error. To address the time-varying nature of the attribute set, the system has designed a dynamic threshold adjustment mechanism that adaptively updates the discrimination criteria based on state changes. In classroom learning scenarios, as students' attention span increases, the system gradually adjusts the classification threshold based on the fatigue accumulation characteristics of the attribute set, making discrimination more consistent with physiological laws. For combined judgments on multiple attributes, the system uses a weighted voting method based on attribute stability scores. The system also establishes a rule evaluation mechanism that optimizes rule parameters by analyzing the accuracy of historical judgment results. Ultimately, a complete judgment rule set is formed, including basic rules, combination rules, dynamic adjustment strategies, and conflict resolution mechanisms.

[0066] Based on the established discriminant rule set, the system generates classification metrics. First, the basic rules in the rule set are applied to the real-time data stream, generating preliminary classification judgments strictly according to the rule thresholds. Then, the combined rules in the rule set comprehensively evaluate multidimensional features to generate metrics such as category attribution, state stability, and transition probability. When different rules produce conflicting results, the system applies the conflict resolution mechanism in the rule set to resolve conflicts through rule priority, evidence weight, and historical consistency analysis, ensuring the consistency of classification results. The system expresses classification results using a probabilistic output defined by the rule set. In addition to providing the most likely category, it also provides a complete probability distribution. Through the dynamic adjustment strategy in the rule set, the system can adaptively adjust classification criteria based on environmental changes and task requirements. The system establishes a multi-level confidence assessment mechanism, calculating confidence levels based on multiple dimensions, including feature stability, discriminant consistency, and historical accuracy. By continuously monitoring classification performance, the system can promptly detect and address anomalies. Ultimately, it outputs a complete set of classification metrics, including category labels, probability distributions, confidence levels, quality scores, and trend predictions.

[0067] In step S104, resource analysis is performed based on the classification indicators to obtain resource status, and tasks are divided based on the resource status to build a processing queue. A scheduling plan is generated based on the processing queue; a confidence analysis is performed on the scheduling plan to obtain a credible interval, and the classification results are filtered based on the credible interval to build an output sequence. Classification information is generated based on the output sequence to complete the online adaptive classification of EEG signals in the AI ​​glasses.

[0068] Specifically, the system performs resource analysis based on the classification metrics output from the previous steps. The class labels in the classification metrics determine the type of computing resources required for each cognitive state. For example, a "focused" state requires more feature extraction resources, while a "relaxed" state requires more noise filtering resources.

[0069] In some embodiments, performing resource analysis based on the classification indicator to obtain resource status includes: determining the computing resource demand type based on the category label in the classification indicator; obtaining the fluctuation range of resource demand based on the probability distribution characteristics in the classification indicator; determining the reliability requirement of resource allocation based on the confidence evaluation result of the classification indicator; determining the processing priority based on the quality score characteristics in the classification indicator; and generating the resource status based on the demand type, fluctuation range, reliability requirement and processing priority.

[0070] In AI glasses' interactive scenarios, when the "Fast Visual Search" state is detected, the system allocates 60% of computing resources to visual feature processing, 20% to context analysis, and 20% to response generation. The probability distribution of metrics is analyzed to estimate the fluctuation range of resource requirements. High-probability states (probability > 0.8) receive 75% of resource allocation priority, states with probabilities between 0.5 and 0.8 receive 20%, and the remaining states share 5%. The confidence level of the classification metric is used to assess the reliability requirements of resource allocation. For states with a confidence level below 0.6, verification resources are increased by 50%, and for states with a confidence level between 0.6 and 0.8, verification resources are increased by 25%. Processing priority is determined by quality score. States with a score greater than 0.9 use double-precision floating-point arithmetic, states with a score between 0.7 and 0.9 use single-precision floating-point arithmetic, and states with a score below 0.7 use fixed-point arithmetic to conserve resources. Based on trend predictions from classification indicators, resources are pre-scheduled. The system pre-shifts 30% of computing resources from the current task unit to the predicted task unit. When the predicted state transition probability exceeds 75%, this proportion is increased to 50%. The system constructs a resource load model, calculates CPU usage, memory usage, and cache utilization for each functional module, and establishes a baseline level of resource utilization. Resource utilization efficiency is evaluated, resource bottlenecks and idle nodes are identified, and a resource balance index is calculated to ensure that the load difference between components does not exceed 30%. The system ultimately outputs a resource status description that includes resource requirements, load distribution, utilization efficiency, and bottleneck locations.

[0071] Based on the output resource status description, the system divides tasks. First, tasks are classified according to resource requirements, and compute-intensive (CPU>70%), memory-intensive (memory>2GB), and IO-intensive (IO wait>30%) tasks are processed separately. In the AR-assisted maintenance scenario, the system assigns object recognition (compute-intensive), scene reconstruction (memory-intensive), and action guidance (IO-intensive) tasks to different processing units, and establishes a mapping table between task types and processing resources. The system analyzes the load distribution characteristics in the resource status, evenly distributes tasks across various resources, and adjusts the task allocation ratio through a dynamic load factor (floating within the range of 0.6-1.2). For the bottleneck position identified in the resource status, the system adopts a task splitting strategy to decompose large tasks into multiple subtasks that can be processed in parallel. The computational load of each subtask is controlled within 20% of the original task, and the data dependency between subtasks does not exceed 15% of the total data volume. In the museum scenario, the system allocates tasks such as feature extraction, pattern recognition, and display updates to different processing units based on the load distribution of resource status, ensuring resource utilization remains within the optimal range of 75% and processing latency is controlled within 100ms. The system establishes a task optimization mechanism based on resource efficiency indicators, caching tasks used more than 10 times per second, achieving a cache hit rate exceeding 85%. The system implements an adaptive task decomposition mechanism, dynamically adjusting task granularity based on current resource status. Task decomposition is increased during high loads (>85%) and reduced during low loads (<40%) to reduce scheduling overhead. The system also considers inter-task dependencies, constructs a task dependency graph to ensure a logical execution order, calculates critical path lengths, and prioritizes tasks along these paths. Ultimately, a task partitioning plan is generated, which includes task types, resource requirements, execution order, and processing strategies.

[0072] Using the task partitioning scheme, the system constructs processing queues. First, a four-level priority queue is set up based on the task types determined in the task partitioning scheme. Urgent tasks (score > 9 points) enter the highest priority queue, important tasks (score 7-9 points) enter the high priority queue, routine tasks (score 5-7 points) enter the medium priority queue, and background tasks (score < 5 points) enter the low priority queue. Each queue uses a different scheduling algorithm: the highest priority queue uses preemptive scheduling, the high priority queue uses time slice round-robin (10ms), the medium priority queue uses weighted round-robin, and the low priority queue uses idle scheduling. The system calculates the execution cost of tasks based on the resource requirements specified in the partitioning scheme. When the expected execution time exceeds 100ms, the task is split into multiple subtasks, with a 5ms execution interval between subtasks to free up resources. Scheduling relationships between queues are established according to the execution order specified in the partitioning scheme to ensure data flow. Priority inversion issues are addressed through a priority inheritance mechanism. The system implements a dynamic queue adjustment mechanism, including adaptive queue length adjustment (fluctuating between 10-100), task timeout processing (the timeout threshold is twice the expected execution time), and load balancing (load differences between queues are controlled within 20%). When the number of backlogged tasks in a queue exceeds 70% of the threshold, a task migration operation is triggered, migrating the last 25% of tasks to a less loaded queue. A fast response channel is designed for sudden tasks to ensure that the response time does not exceed 50ms. The processing capacity of this channel is limited to 10% of the total processing capacity to prevent resource preemption. The system establishes a queue status monitoring mechanism to track changes in queue length, processing delay, and throughput in real time, sampling the queue status every 100ms and calculating trends. The final output is a processing queue framework that includes a multi-level queue structure, scheduling relationships, and a monitoring solution.

[0073] Exemplarily, generating a scheduling scheme based on the processing queue includes: dividing the processing queue into parallel task units according to the demand type in the resource status; establishing a mapping relationship between task units and computing nodes based on the results of resource utilization efficiency evaluation; setting the execution order of the task units according to processing priority characteristics; setting a dynamically adjusted time window according to the resource fluctuation range; configuring a verification mechanism for the task unit according to reliability requirements, and establishing a resource transfer channel between task units based on advance scheduling characteristics; generating the scheduling scheme according to the execution order, time window, verification mechanism and resource transfer channel.

[0074] Based on the processing queue framework, the system generates a scheduling plan. First, a basic scheduling policy based on queue priority is established: the highest-priority queue receives 40% of the processing time, the high-priority queue receives 30%, the medium-priority queue receives 20%, and the low-priority queue receives 10%. In actual implementation, an adaptive weight adjustment algorithm is used, and the actual time allocation of queues fluctuates within ±15% of the base value. The system dynamically adjusts scheduling parameters based on the current state of the queues. When a queue's length exceeds 80% of the threshold, its processing priority is temporarily increased, increasing its processing time by 25%. When a queue is empty, its time slice is proportionally allocated to other queues. Multiple scheduling algorithms are implemented: round-robin scheduling (fair distribution of processing time) is used when the load is below 50%, weighted round-robin scheduling (weights are proportional to priority) is used when the load is between 50% and 80%, and priority scheduling (higher-priority tasks are prioritized) is used when the load is above 80%. The system implements a real-time guarantee mechanism, reserving 25% of the processing time slice for time-sensitive tasks, ensuring a response latency of no more than 20ms for these tasks. In the video processing scenario, the frame rate stabilization task is marked as time-sensitive, ensuring that one frame is processed every 33ms, and maintaining a smooth 30fps even under high load conditions. The system implements a task migration and load balancing mechanism. When the load of a processing unit exceeds 90%, the task is migrated to a unit with a load less than 60%, and the integrity of the task status is maintained during the migration process. The system has established a performance monitoring and evaluation mechanism, including indicators such as response time, throughput, and resource utilization. A performance report is generated every 500ms, and the scheduling parameters are dynamically adjusted according to the report results. For abnormal situations such as processing unit failures, a task reallocation mechanism is implemented to complete the migration and recovery of affected tasks within 100ms. The final output is a complete scheduling plan that includes scheduling strategies, parameter configurations, and exception handling.

[0075] The system performs a confidence analysis on the scheduling plan output from the above steps. It assesses the execution reliability of different tasks based on the time slice allocation and priority settings in the scheduling plan. For the highest-priority queue, which receives a 40% time slice, the baseline confidence level is set at 0.9, increasing to 0.95 when execution is stable. For the high-priority queue, which receives a 30% time slice, the baseline confidence level is 0.8, rising to 0.85. For the medium-priority queue (20% time slice), the baseline confidence level is 0.7, and for the low-priority queue (10% time slice), the baseline confidence level is 0.6. The system dynamically evaluates the performance of the scheduling plan based on performance monitoring data. The baseline confidence level is maintained when response time is less than 50ms and resource utilization is between 60% and 80%. For deviations, the confidence level is reduced by 0.1 for every 10% increase. For task migration, the system evaluates execution stability before and after the migration: If performance fluctuations do not exceed 10%, the original confidence level is maintained; if fluctuations are between 10% and 20%, the confidence level is reduced by 0.2; and if fluctuations exceed 20%, the confidence level is reduced by 0.3. The system assesses recovery capabilities based on exception handling records, using recovery time as a key metric: Exceptions recovered within 100ms receive only a slight decrease in confidence (-0.1), those between 100-500ms receive a moderate decrease (-0.2), and those exceeding 500ms receive a significant decrease (-0.3). The system uses a sliding window approach to account for short-term fluctuations in confidence, with the window size adjusting within the 100-500ms range based on the task type. Finally, a confidence interval is assigned to each processing period, consisting of a central confidence value and a fluctuation range calculated based on performance metrics.

[0076] Based on the labeled confidence intervals, the system screens classification results. First, the confidence values ​​within the confidence intervals are used for preliminary filtering. For real-time tasks, results with a confidence level above 0.7 and a fluctuation range of less than ±0.2 are retained. For offline tasks, the limits can be relaxed to 0.6 and ±0.25. The system establishes a dynamic threshold strategy that adaptively adjusts based on the fluctuation range of the confidence interval: a base threshold of 0.7 is maintained when the fluctuation is less than ±0.1, raised to 0.8 when the fluctuation reaches ±0.2, and to 0.85 when it exceeds ±0.2. For periods of continuous low confidence (<0.6), additional verification steps are added, including temporal consistency checks, spatial correlation verification, and logical plausibility assessment. The system employs a hierarchical evaluation strategy, first verifying state stability within a short time window (100ms), then analyzing transition plausibility within a medium time window (500ms), and finally evaluating trend consistency within a long time window (>1s). For critical tasks like attention analysis, a stricter screening standard (confidence level > 0.8) is adopted, while for non-critical tasks like environmental monitoring, a more relaxed standard (confidence level > 0.65) is used. The system implements a verification mechanism based on historical data, comparing current classification results with historical patterns. Results with deviations exceeding 30% require additional verification steps. The system ultimately outputs a set of rigorously verified classification results, each of which includes detailed verification records.

[0077] The system constructs an output sequence using the filtered classification result set. The classification results are first time-aligned and processed using multi-scale time windows. The window size is dynamically adjusted between 100ms and 1s based on task characteristics. For periods of missing data, the system interpolates data based on the distribution characteristics of nearby valid data. Short-term missing data (<200ms) are interpolated using linear interpolation, while long-term missing data (>200ms) use spline interpolation to maintain curve smoothness. The system establishes a sequence integrity verification mechanism that verifies timestamp continuity and data integrity to ensure that classification results at key time points are accurately preserved. Sequence abrupt changes are smoothed using adaptive filtering to eliminate transient fluctuations. The filter parameters are dynamically adjusted based on the signal change rate. A hierarchical data compression strategy is designed, using a high compression ratio (10:1) for stable periods and a low compression ratio (2:1) for changing periods to preserve details. A version control mechanism is implemented to store the complete history of the last 24 hours, enabling backtracking and comparison of results. The system ultimately generates a normalized output sequence containing time-aligned classification labels, confidence metrics, and state transition records.

[0078] In some embodiments, generating classification information based on the output sequence includes: performing multidimensional EEG signal pattern analysis based on the classification label data in the output sequence; establishing an EEG feature reliability mapping relationship based on the confidence index in the credible interval; constructing cognitive state transition features based on the stability assessment results in the credible interval; generating EEG frequency band activity features based on the classification label data and the credible interval parameters; integrating state transition features and EEG activity features to establish cognitive state recognition results; and generating the classification information based on the analysis results of the pattern analysis and the cognitive state recognition results.

[0079] Based on the output sequence, the system generates final classification information. First, it processes the classification label data in the sequence and performs multi-dimensional EEG signal pattern analysis, calculating the temporal distribution, energy ratio, and state correlation of activity in the α, β, θ, and δ bands. The system quantifies the EEG feature combinations and their persistence characteristics of different cognitive states (such as "focus," "relaxation," and "fatigue"), providing a basic pattern library for online recognition by AI glasses. Second, the system analyzes confidence metrics in the sequence, establishing a mapping between EEG signal fluctuations and classification reliability, and identifying high-confidence EEG features (such as stable prefrontal β / θ ratio and significant parietal α wave suppression) and low-confidence features (such as multi-channel signal asynchrony and rapid spectral changes). Based on this information, the AI ​​glasses dynamically adjust the classification threshold to maintain recognition accuracy under varying signal quality conditions. Third, the system uses state transition records in the sequence to construct a cognitive state flow graph, capturing typical EEG signal pattern transition chains. In reading assistance scenarios, AI glasses analyze state transitions to identify the user's reading cognitive cycle, such as "browsing (theta wave enhancement) - focusing (beta wave dominance) - understanding (alpha / gamma wave synergy) - reviewing (alpha wave enhancement)." Based on this, they predict the next possible state, enabling proactive adjustments to intelligent interactions. The system integrates these three types of information (EEG classification labels, confidence changes, and state transitions) into the AI ​​glasses' online adaptive engine, enabling real-time adjustments to interface prompts, information presentation, and interactive response strategies based on the user's current EEG state. For example, when high beta wave activity (deep concentration) is detected, the AI ​​glasses suppress non-essential notifications; when enhanced theta waves (diminished attention) are identified, the salience of visual cues is enhanced. The system implements a scenario-based classification application mechanism for the AI ​​glasses, with pre-defined EEG pattern recognition strategies for different scenarios, such as learning, work, and leisure. This allows the glasses to complete state recognition and trigger corresponding interactive adjustments within 10ms. The final output classification information includes EEG frequency band activity characteristics, cognitive state recognition results and state transition predictions, which directly drives the adaptive human-computer interaction system of AI glasses and realizes real-time intelligent response based on EEG signals.

[0080] The provided method has the following beneficial effects:

[0081] 1. By establishing a multi-level signal processing and feature extraction mechanism, a complete processing flow from original signal quality control to feature domain construction is realized, and data streams are generated based on the feature domain for classification, which significantly improves the classification accuracy and system stability.

[0082] 2. An innovative sample processing and adaptive balance mechanism is proposed. Through key point extraction, novelty analysis and equilibrium point marking, the problems of dynamic feature changes and category balance are effectively solved, the system's adaptability to environmental changes is enhanced, and the stability of classification performance is ensured.

[0083] 3. A complete solution of feature combination, resource scheduling and confidence analysis was designed. Through classification chain construction, resource status analysis and trust interval marking, the system resource utilization was optimized while ensuring classification performance, thereby improving the application effect of AI glasses in actual scenarios.

[0084] In order to implement the online adaptive classification method of EEG signals in AI glasses corresponding to the above method embodiment, to achieve the corresponding functions and technical effects. Figure 2 , Figure 2 The following is a block diagram of a device 200 for online adaptive classification of EEG signals in AI glasses provided by an embodiment of the present application. For ease of illustration, only the parts relevant to this embodiment are shown. The device 200 for online adaptive classification of EEG signals in AI glasses provided by an embodiment of the present application includes:

[0085] The signal acquisition unit 201 is configured to acquire real-time EEG signals from the AI ​​glasses as raw data, perform quality testing on the raw data, generate a basic sequence based on the quality testing results, and generate a data stream corresponding to the basic sequence; perform category analysis on the data stream to obtain a classification benchmark, and generate a classification sequence based on the classification benchmark;

[0086] A sample processing unit 202 is configured to perform sample processing on the classification sequence and mark key points; perform feature extraction on the key points, establish a storage sequence based on the features obtained by the feature extraction, generate a replay data sequence based on the storage sequence, and generate an adaptive sequence based on the replay data sequence;

[0087] A category statistics unit 203 is configured to perform category statistics based on the classification sequence, obtain distribution characteristics, mark equilibrium points based on the distribution characteristics, construct selection rules based on the equilibrium points, generate a balanced sequence based on the selection rules, and generate a classification index based on the adaptive sequence and the balanced sequence;

[0088] The classification completion unit 204 is used to perform resource analysis based on the classification indicators, obtain resource status, divide tasks based on the resource status to build a processing queue, and generate a scheduling plan based on the processing queue; perform confidence analysis on the scheduling plan to obtain a credible interval, filter the classification results based on the credible interval to build an output sequence, generate classification information based on the output sequence, and complete the online adaptive classification of EEG signals in the AI ​​glasses.

[0089] The above-mentioned AI glasses EEG signal online adaptive classification device 200 can implement the AI ​​glasses EEG signal online adaptive classification method of the above-mentioned method embodiment. The optional options in the above-mentioned method embodiment also apply to this embodiment and will not be described in detail here. The remaining contents of the embodiment of this application can refer to the contents of the above-mentioned method embodiment and will not be repeated in this embodiment.

[0090] Figure 3 This is a schematic diagram of the structure of a computer device provided in one embodiment of the present application. Figure 3 As shown, the computer device 3 of this embodiment includes: at least one processor 30 ( Figure 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.

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

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

[0093] In some embodiments, the memory 31 may be an internal storage unit of the computer device 3, such as a hard drive 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 drive, a Smart Media Card (SMC), a Secure Digital (SD) card, a flash memory card, etc. equipped on the computer device 3. Furthermore, the memory 31 may include both an internal storage unit of the computer device 3 and an external storage device. The memory 31 is used to store an operating system, application programs, a boot loader, 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 about to be output.

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

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

[0096] In several embodiments provided in the present application, it is understood that each box in the flow chart or block diagram can represent a part of a module, program segment or code, and the part of the module, program segment or 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 an order different 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, which depends on the functions involved.

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

[0098] The specific embodiments described above further illustrate the objectives, technical solutions, and beneficial effects of this application. It should be understood that the above descriptions are merely specific embodiments of this application and are not intended to limit the scope of protection of this application. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of this application by those skilled in the art should be included within the scope of protection of this application.

Claims

1. An online adaptive classification method for EEG signals in AI glasses, characterized by: include: Acquire real-time EEG signals from the AI ​​glasses as raw data, perform quality testing on the raw data, generate a basic sequence based on the quality testing results, and generate a data stream corresponding to the basic sequence; Performing category analysis on the data stream to obtain a classification benchmark, and generating a classification sequence according to the classification benchmark; Performing sample processing on the classification sequence and marking key points; The key points are subjected to feature extraction, a storage sequence is established based on the features obtained by feature extraction, a replay data sequence is generated based on the storage sequence, and an adaptive sequence is generated based on the replay data sequence, including: performing novelty analysis on the replay data sequence to obtain change features; marking an adjustment interval based on the change features; determining an update step size based on the adjustment interval; and generating an adaptive sequence based on the update step size; the replay process corresponding to the replay data sequence first performs data interpolation on the storage sequence, reconstructs a continuous signal using a cubic spline function, ensures that the restored data is consistent with the original features in the storage sequence, and reproduces the cognitive state changes recorded in the storage sequence at different time scales through a multi-scale replay mechanism, and the replay data sequence retains the key features and category information in the storage sequence; the novelty analysis includes establishing a novelty evaluation system including temporal stability, spatial consistency and spectral characteristics, the change features are a set of change features that describe the novel characteristics of the replay data, and each change feature includes quantitative indicators of significance level, occurrence frequency, spatial distribution characteristics and task relevance; performing category statistics according to the classification sequence to obtain distribution characteristics, marking equilibrium points based on the distribution characteristics, constructing selection rules according to the equilibrium points, generating equilibrium sequences according to the selection rules, and generating classification indicators according to the adaptive sequence and the equilibrium sequence; Resource analysis is performed according to the classification indicators to obtain resource status, and tasks are divided based on the resource status to build a processing queue. A scheduling plan is generated based on the processing queue. Confidence analysis is performed on the scheduling plan to obtain a credible interval. Classification results are filtered based on the credible interval to construct an output sequence. Classification information is generated based on the output sequence to complete the online adaptive classification of EEG signals in AI glasses.

2. The method according to claim 1, characterized in that Generating the data stream corresponding to the basic sequence includes: Performing feature division on the basic sequence and marking feature points; constructing a feature domain based on the feature points; The data stream is generated according to a feature domain.

3. The method according to claim 1, characterized in that Generating a classification sequence according to the classification benchmark includes: constructing a feature space based on the classification benchmark; Class attributes are marked according to the feature space, and the classification sequence is generated according to the class attributes.

4. The method according to claim 1, wherein Generating a classification index according to the adaptive sequence and the balanced sequence includes: Combining features according to the adaptive sequence and the balanced sequence to construct a classification chain; performing attribute extraction based on the classification chain; A discrimination rule is established based on the attributes obtained by the attribute extraction, and a classification index is generated based on the discrimination rule.

5. The method according to claim 1, characterized in that The performing of resource analysis according to the classification indicators to obtain resource status includes: Determine the computing resource requirement type according to the category label in the classification indicator; Obtain the fluctuation range of resource demand based on the probability distribution characteristics of the classification indicators; Determining reliability requirements for resource allocation based on confidence evaluation results of the classification indicators; Determine the processing priority based on the quality score characteristics in the classification indicators; The resource status is generated according to the demand type, fluctuation range, reliability requirement and processing priority.

6. The method according to claim 5, characterized in that Generating a scheduling plan according to the processing queue includes: dividing the processing queue into parallel task units according to the demand type in the resource status; Establish a mapping relationship between task units and computing nodes based on the results of resource utilization efficiency evaluation; Setting the execution order of the task units according to processing priorities; Set a dynamically adjusted time window based on the fluctuation range of resource demand; Configure verification mechanisms for task units based on reliability requirements and establish resource transfer channels between task units based on advance scheduling features; The scheduling scheme is generated according to the execution sequence, time window, verification mechanism and resource transfer channel.

7. The method according to claim 1, characterized in that Generating classification information according to the output sequence includes: performing multidimensional EEG signal pattern analysis based on the classification label data in the output sequence; Establishing an EEG feature reliability mapping relationship based on the confidence index in the credible interval; constructing a cognitive state transition feature according to the stability assessment results in the confidence interval; Generate EEG frequency band activity features based on classification label data and credible interval parameters; integrate cognitive state transition features with EEG frequency band activity features to establish cognitive state recognition results; The classification information is generated based on the analysis result of the pattern analysis and the cognitive state recognition result.

8. An online adaptive classification device for EEG signals in AI glasses, characterized by: include: A signal acquisition unit, configured to acquire real-time EEG signals from the AI ​​glasses as raw data, perform quality detection on the raw data, generate a basic sequence based on the quality detection results, and generate a data stream corresponding to the basic sequence; Performing category analysis on the data stream to obtain a classification benchmark, and generating a classification sequence according to the classification benchmark; A sample processing unit, configured to perform sample processing on the classification sequence and mark key points; The key points are subjected to feature extraction, a storage sequence is established based on the features obtained by feature extraction, a replay data sequence is generated based on the storage sequence, and an adaptive sequence is generated based on the replay data sequence, including: performing novelty analysis on the replay data sequence to obtain change features; marking an adjustment interval based on the change features; determining an update step size based on the adjustment interval; and generating an adaptive sequence based on the update step size; the replay process corresponding to the replay data sequence first performs data interpolation on the storage sequence, reconstructs a continuous signal using a cubic spline function, ensures that the restored data is consistent with the original features in the storage sequence, and reproduces the cognitive state changes recorded in the storage sequence at different time scales through a multi-scale replay mechanism, and the replay data sequence retains the key features and category information in the storage sequence; the novelty analysis includes establishing a novelty evaluation system including temporal stability, spatial consistency and spectral characteristics, the change features are a set of change features that describe the novel characteristics of the replay data, and each change feature includes quantitative indicators of significance level, occurrence frequency, spatial distribution characteristics and task relevance; a category statistics unit, configured to perform category statistics based on the classification sequence, obtain distribution characteristics, mark equilibrium points based on the distribution characteristics, construct selection rules based on the equilibrium points, generate equilibrium sequences based on the selection rules, and generate classification indicators based on the adaptive sequence and the equilibrium sequence; The classification completion unit is used to perform resource analysis based on the classification indicators, obtain resource status, divide tasks based on the resource status to build a processing queue, and generate a scheduling plan based on the processing queue; perform confidence analysis on the scheduling plan to obtain a credible interval, filter the classification results based on the credible interval to build an output sequence, generate classification information based on the output sequence, and complete the online adaptive classification of EEG signals in the AI ​​glasses.

9. 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 7 when executing the computer program.

Images