Personalized EEG pattern recognition method, device and equipment based on AI glasses

Abstract

The present application relates to the field of brain-computer interface technology, and in particular to a personalized EEG pattern recognition method, device, and equipment based on AI glasses. The method includes: marking the EEG data collected by AI glasses, building a personal benchmark, and performing pattern analysis to obtain feature mapping; using the analysis results corresponding to the pattern analysis to mark the feature interval to generate a user portrait and the corresponding initial template; performing feature mapping based on the initial template to obtain universal features, marking individual features based on the universal features, obtaining the corresponding recognition benchmark to perform scene classification, and marking the usage environment; constructing a state sequence, selecting a feature template, generating a recognition scheme for trend analysis, obtaining statistical features, and marking the pattern interval; constructing a feature chain based on the pattern interval, and generating a feature combination based on the feature chain; performing pattern analysis on the feature combination, marking the recognition sequence; constructing a mapping relationship based on the recognition sequence; and using the mapping relationship to reorganize features and generate a personalized pattern.

Description

Technical Field

[0001] The present application relates to the field of brain-computer interface technology, and in particular to a personalized EEG pattern recognition method, device and equipment based on AI glasses. Background Art

[0002] AI glasses face significant individual variability when recognizing EEG patterns. Traditional EEG pattern recognition methods typically use universal models that are unable to effectively adapt to the unique EEG characteristics and usage habits of different users. While some research has attempted to achieve personalization through data-driven approaches, these methods often require large amounts of individual training data and a lengthy adaptation process, severely impacting the user experience. Furthermore, existing personalization methods lack effective modeling of long-term changes in users' EEG patterns, making it difficult to adapt to gradual changes in users' cognitive states and habits.

[0003] There are several key problems in the existing technology when dealing with personalized EEG pattern recognition. First, the user's EEG pattern will be affected by many factors such as physical and mental state, environmental factors, etc., and exhibits complex dynamic characteristics. It is difficult for existing methods to achieve stable recognition in multiple scenarios. Secondly, obtaining sufficient personal training data is often time-consuming and costly, especially for new users, and it is impossible to quickly establish an effective personalized model. In addition, while maintaining the personalization of the model, the existing methods find it difficult to effectively utilize the common features in the group data, resulting in limited generalization capabilities of the model. The existing technology also lacks systematic feature management and dynamic optimization mechanisms, and cannot automatically adjust the recognition strategy according to changes in user habits, making it difficult to maintain stable recognition performance during long-term use.

[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 personalized EEG pattern recognition method, device and equipment based on AI glasses, which aims to solve several key problems existing in the prior art when dealing with personalized EEG pattern recognition. First, the user's EEG pattern will be affected by many aspects such as physical and mental state, environmental factors, etc., and will show complex dynamic characteristics. It is difficult for existing methods to achieve stable recognition in multiple scenarios. Secondly, obtaining sufficient personal training data is often time-consuming and costly, especially for new users, and it is impossible to quickly establish an effective personalized model. In addition, while maintaining the personalization of the model, the existing methods find it difficult to effectively utilize the common features in the group data, resulting in limited generalization capabilities of the model. The existing technology also lacks a systematic feature management and dynamic optimization mechanism, and is unable to automatically adjust the recognition strategy according to changes in user habits, making it difficult to maintain stable recognition performance during long-term use.

[0006] In a first aspect, an embodiment of the present application provides a personalized EEG pattern recognition method based on AI glasses, comprising:

[0007] Perform feature tagging on the EEG data collected by the AI ​​glasses and construct a personal benchmark; perform pattern analysis based on the personal benchmark to obtain a feature map; use the analysis results corresponding to the pattern analysis to mark the feature interval to generate a user profile, and obtain an initial template corresponding to the user profile;

[0008] Perform feature mapping based on the initial template to obtain common features, and mark individual features based on the common features; construct fusion rules using the individual features, generate a feature set based on the fusion rules, and obtain an identification benchmark corresponding to the feature set;

[0009] Classify the scene according to the recognition benchmark and mark the usage environment; construct a state sequence based on the usage environment, select a feature template according to the state sequence, generate recognition parameters according to the feature template, and generate a recognition solution according to the recognition parameters;

[0010] Performing trend analysis on the identification scheme to obtain statistical features, marking pattern intervals based on the statistical features; constructing a feature chain based on the pattern intervals, and generating a feature combination based on the feature chain;

[0011] Performing pattern analysis on the feature combination to mark an identification sequence; constructing a mapping relationship based on the identification sequence; utilizing the mapping relationship to perform feature recombination, and generating a personalized pattern based on the feature recombination.

[0012] In a second aspect, the present application further provides a personalized EEG pattern recognition device, comprising:

[0013] A benchmark establishment unit is configured to perform feature tagging on the EEG data collected by the AI ​​glasses and construct a personal benchmark; perform pattern analysis based on the personal benchmark to obtain a feature map; use the analysis results corresponding to the pattern analysis to tag feature intervals to generate a user profile, and obtain an initial template corresponding to the user profile;

[0014] a feature mapping unit configured to perform feature mapping based on the initial template to obtain common features, mark individual features based on the common features, construct fusion rules using the individual features, generate a feature set based on the fusion rules, and obtain an identification benchmark corresponding to the feature set;

[0015] A scene classification unit, configured to classify scenes according to the recognition benchmark and mark the usage environment; construct a state sequence based on the usage environment, select a feature template according to the state sequence, generate recognition parameters according to the feature template, and generate a recognition solution according to the recognition parameters;

[0016] a trend analysis unit, configured to perform trend analysis on the identification scheme, obtain statistical features, mark pattern intervals based on the statistical features, construct a feature chain based on the pattern intervals, and generate a feature combination based on the feature chain;

[0017] The pattern analysis unit is used to perform pattern analysis on the feature combination and mark an identification sequence; construct a mapping relationship based on the identification sequence; use the mapping relationship to perform feature recombination, and generate a personalized pattern based on the feature recombination.

[0018] 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 personalized EEG pattern recognition method based on AI glasses as described in the first aspect is implemented.

[0019] This method achieves personalized EEG signal recognition through dynamic feature modeling and scene adaptation techniques. The user's electroencephalogram (EEG) data is collected using sensors built into AI glasses. Preprocessing (denoising and artifact removal) extracts time-domain, frequency-domain, and time-frequency domain features (such as the wavelet transform energy spectrum). Combining stratified sampling data from different cognitive states (focused, relaxed, etc.) with the user, a personal benchmark database containing feature parameters, probability distributions, and temporal correlation characteristics is constructed. Pattern analysis is performed based on the personal benchmark to extract common features (such as alpha wave energy) and individual characteristics (such as fluctuation patterns in specific frequency bands), and fusion rules (such as weighted fusion or neural network fusion) are generated. Feature interval labeling (such as high-frequency band sensitivity intervals) is used to construct a user profile and generate an initial template (such as a standardized feature weight template). The user's usage environment (e.g., noisy environment, quiet environment) is classified according to the scene sensitivity level (high / medium / low), and the corresponding feature template (such as a noise suppression template) is matched. Recognition parameters (such as thresholds and weights) are dynamically adjusted based on state sequences (such as changes in EEG features within a continuous time window) to generate a scene-adapted recognition solution. Perform trend analysis on the recognition scheme (e.g., calculating stability indicators), mark pattern intervals (e.g., abnormal fluctuation intervals), and construct feature chains (a collection of time-series related features). Generate the final personalized pattern through feature recombination (e.g., principal component analysis or deep learning optimization), achieving dynamic calibration and high-precision recognition.

[0020] By leveraging a personal benchmark database and dynamic feature mapping, the system addresses the misjudgment issues inherent in traditional EEG recognition due to individual differences, improving classification accuracy (e.g., increasing the accuracy of user focus by over 30%). Based on scene sensitivity classification and state sequence analysis, it enables real-time adjustment of recognition parameters (e.g., automatically enhancing high-frequency suppression in noisy environments), enhancing system robustness. By integrating the hardware computing capabilities of AI glasses (e.g., edge computing chips), feature extraction and pattern generation are performed locally, reducing cloud reliance and lowering response latency (<50ms). User profiling and mode interval tagging support personalized feedback (e.g., attention reminders, fatigue warnings), enhancing user-friendly interaction.

[0021] 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

[0022] Figure 1 This is a flow chart of a personalized EEG pattern recognition method based on AI glasses according to an embodiment of the present application;

[0023] Figure 2 This is a schematic diagram of the structure of a personalized EEG pattern recognition device shown in an embodiment of the present application;

[0024] Figure 3 This is a schematic diagram of the structure of a computer device shown in an embodiment of the present application. DETAILED DESCRIPTION

[0025] 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.

[0026] 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.

[0027] 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.

[0028] 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.

[0029] 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.

[0030] 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.

[0031] The technical solutions of the embodiments of this application are introduced below.

[0032] AI glasses face significant individual variability when recognizing EEG patterns. Traditional EEG pattern recognition methods typically use universal models that are unable to effectively adapt to the unique EEG characteristics and usage habits of different users. While some research has attempted to achieve personalization through data-driven approaches, these methods often require large amounts of individual training data and a lengthy adaptation process, severely impacting the user experience. Furthermore, existing personalization methods lack effective modeling of long-term changes in users' EEG patterns, making it difficult to adapt to gradual changes in users' cognitive states and habits.

[0033] There are several key problems in the existing technology when dealing with personalized EEG pattern recognition. First, the user's EEG pattern will be affected by many factors such as physical and mental state, environmental factors, etc., and exhibits complex dynamic characteristics. It is difficult for existing methods to achieve stable recognition in multiple scenarios. Secondly, obtaining sufficient personal training data is often time-consuming and costly, especially for new users, and it is impossible to quickly establish an effective personalized model. In addition, while maintaining the personalization of the model, the existing methods find it difficult to effectively utilize the common features in the group data, resulting in limited generalization capabilities of the model. The existing technology also lacks systematic feature management and dynamic optimization mechanisms, and cannot automatically adjust the recognition strategy according to changes in user habits, making it difficult to maintain stable recognition performance during long-term use.

[0034] Please refer to Figure 1 , Figure 1 The flowchart of a personalized EEG pattern recognition method based on AI glasses provided in an embodiment of the present application is provided. The personalized EEG pattern recognition method based on AI glasses in an embodiment of the present application can be applied to computer devices, including but not limited to smart phones, laptops, tablet computers, desktop computers, physical servers, cloud servers and other devices. Figure 1 As shown, the personalized EEG pattern recognition method based on AI glasses of this embodiment includes steps S101 to S105, which are detailed as follows:

[0035] Step S101: feature-mark the EEG data collected by the AI ​​glasses and construct a personal benchmark; perform pattern analysis based on the personal benchmark to obtain a feature map; use the analysis results corresponding to the pattern analysis to mark the feature interval to generate a user portrait, and obtain the initial template corresponding to the user portrait.

[0036] Specifically, the EEG data collected by AI glasses contains multi-channel raw potential signals, which need to be feature-labeled and a personal benchmark constructed first.

[0037] In some embodiments, the feature tagging of the EEG data collected by the AI ​​glasses and the construction of a personal benchmark include: preprocessing the EEG data to remove physiological artifacts and interference noise; extracting the time domain features, frequency domain features and time-frequency domain features of the EEG data and performing dimensionality reduction processing; establishing a feature statistical parameter library of the EEG data through standardization processing; combining stratified sampling data under different cognitive states to analyze the time series correlation characteristics and fluctuation characteristics of the feature statistical parameter library; calculating the feature stability index of the feature statistical parameter library based on the time series correlation characteristics and fluctuation characteristics and constructing a personal benchmark database containing feature parameters, probability distribution models and time series correlation characteristics.

[0038] The raw signal preprocessing stage included the use of a bandpass filter (0.5-45 Hz) to remove power frequency interference and baseline drift, and independent component analysis (ICA) to remove physiological artifacts such as electrooculography and electromyography. During the feature extraction stage, time-domain features (mean amplitude, root mean square value, peak-to-peak value), frequency-domain features (energy in the five frequency bands of δ, θ, α, β, and γ), and time-frequency domain features (wavelet transform time-varying features) were extracted. Dimensionality reduction was performed on the extracted features using principal component analysis, retaining principal components that explained 95% of the variance. All features were z-score normalized to establish a feature statistical parameter library. Kernel density estimation was also used to fit the probability distribution of each feature to capture its distributional characteristics. During data collection, a stratified sampling strategy was used, uniformly sampling across different cognitive states, such as alertness, focus, and relaxation, with at least 1000 sample points selected for each state. Feature combinations with significant temporal patterns were identified by analyzing autocorrelation (lag = 10) and cross-correlation (threshold 0.6). A fixed time window (2s width, 1s step) analysis was used to quantify the short-term fluctuation characteristics of the features. Stability metrics were calculated for each extracted feature, including variance stability (σ^2 < 0.3) and temporal consistency (τ > 0.7). The resulting personal benchmark database contains 70 standardized feature parameters, probability distribution models between features, and temporal correlation characteristics. These data together form a personalized benchmark model for EEG features.

[0039] Pattern analysis was performed based on 70 standardized feature parameters, a probability distribution model between features, and temporal correlation characteristics from a personal benchmark database. First, the statistical characteristics of each dimension were calculated using the 70 standardized feature parameters from the benchmark database. Feature importance was then calculated using the probability distribution model, with higher weights assigned to features with high probability distribution dispersion (H > 0.8) and strong temporal correlation (r > 0.6). A weighted Pearson correlation coefficient matrix was constructed using temporal correlation characteristics to identify feature combinations with absolute correlation coefficients exceeding 0.8 and exhibiting stable temporal patterns. Cluster analysis was performed using an improved K-means algorithm, with cluster centers initialized to the distribution centers of the most stable features among the 70 feature parameters in the benchmark model. The optimal number of clusters (K = 5) was determined by calculating the Davies-Bouldin index and silhouette coefficient. The number of iterations was set to 500, and the convergence threshold was set to 1 × 10^-4. The probability distribution information of the benchmark model was incorporated to weight the distance calculations during the clustering process. The bootstrap method was used to perform 100 clustering iterations, and stable structures with a frequency exceeding 80% were selected based on temporal stability. Probabilistic distribution characteristics, including multidimensional Gaussian distribution parameters and kernel density estimation curves, were calculated for each cluster. Based on the distribution characteristics and temporal correlation characteristics of 70 characteristic parameters, a dynamic Mahalanobis distance metric was constructed, with an adaptive threshold (τ = 0.6). The analysis fully utilized all elements in the benchmark database, ultimately yielding a set of stable characteristic patterns, including characteristic distribution parameters, dynamic eigenvectors, and multidimensional similarity relationships between patterns.

[0040] Feature intervals and user profiles were constructed using the baseline model's 70 standardized feature parameters, a probability distribution model, and clustering results from pattern analysis (including feature distribution parameters, dynamic feature vectors, and multidimensional similarity between patterns). Based on the probability distribution of feature parameters in the baseline model, the 3σ principle combined with kernel density estimation was used to determine the variation range of each feature. For feature parameters with high temporal stability, a quantile method (Q1 = 25%, Q2 = 50%, Q3 = 75%) was used to divide the baseline level. Based on the pattern clustering results, adaptive feature intervals were constructed using the feature distribution parameters and dynamic feature vectors of each pattern. The dynamic feature vectors were used to capture the transient characteristics of feature changes, quantifying the dynamic characteristics of state transitions by calculating the directional changes and amplitude fluctuations of the vector sequence. Based on the multidimensional similarity between patterns, an overlap threshold (θ = 0.3) was set to identify independent and transitional intervals. Interval transition patterns were analyzed based on temporal correlation characteristics, and a Markov transition model considering the probability distribution was established. The calculation of transition probabilities fully utilized the temporal stability indicators in the baseline model, assigning greater weight to features with high stability. Multiple constraints were introduced during the interval partitioning process: information gain ratio (≥0.6) based on probability distribution, Gini coefficient (<0.3) of sample distribution, and autocorrelation coefficient (r>0.4) of time series features. By comprehensively analyzing the distribution characteristics of 70 feature parameters, pattern clustering results (including feature distribution parameters, dynamic feature vectors, and multidimensional similarity relationships between patterns), and state transition patterns, a multi-level user profile was constructed. This profile includes: feature preference descriptions based on probability distribution, cognitive state characteristics based on clustering, dynamic behavioral patterns based on time series analysis, and the interrelationships and evolutionary patterns among these features. This three-dimensional user profile comprehensively reflects the uniqueness of individual EEG characteristics.

[0041] In some embodiments, obtaining the initial template corresponding to the user portrait includes: performing regional decomposition on the user portrait and marking key features to construct a feature space; performing sample matching based on the feature space; and generating an initial template by analyzing feature combinations based on the matching results of the sample matching.

[0042] The user profiles generated in the above steps were subjected to regional decomposition. Key feature combinations were extracted based on the probability distribution-based feature preference descriptions, clustering-based cognitive state characteristics, dynamic behavior patterns based on time series analysis, and the intercorrelations and evolution patterns between the features. During the decomposition of the probability distribution-based feature preference descriptions, the existing probability distribution parameters were utilized to retain features with stable distribution shapes (kurtosis range 2.8-3.2) and low volatility (coefficient of variation <15%). For the processing of the clustering-based cognitive state characteristics, the existing cluster centers and transition probability matrix were retained and dimensionality reduction was performed using the t-SNE method. The perplexity during the mapping process was set to 30, and the maximum number of iterations was 2000. For the processing of the dynamic behavior patterns based on time series analysis, the established state transition network was used as a foundation, and significant eigenvectors (eigenvalues ​​> 0.1) were extracted through graph decomposition. Multiple stability constraints were introduced during the feature selection process: the predicted variance of the features under the existing probability model was required to be less than 15% of the total variance; the temporal consistency index of the features was required to be greater than 0.7; and the coefficient of variation of the features across repeated measurements was required to be less than 10%. For physiological features such as EEG rhythms, their stability at different time scales is analyzed, with short-term fluctuations (within 1 hour) limited to ±5% of the mean, and long-term drift (within 24 hours) limited to ±10%. The final constructed feature space includes the following elements: the physical meaning of the original features (e.g., the range of values ​​for cognitive load-related dimensions is [0.2, 0.8], the response time threshold for attention-related dimensions is 200ms, and the update period for emotional state-related dimensions is no less than 2s), probability distribution characteristics (including mean, standard deviation, covariance matrix, and probability density function), feature distribution thresholds (including upper and lower limits of the normal range and anomaly detection threshold), and a temporal dependency model.

[0043] Sample matching analysis is performed based on the constructed feature space. Distance metrics are designed for different types of features based on the probability distribution characteristics of the feature space: for features with a normal distribution, the Mahalanobis distance is used, using the mean and covariance matrix stored in the feature space; for features with a skewed distribution, a distance based on the KL divergence is used, referencing the probability density function of the feature space; and for discretely distributed features, a modified Hamming distance is used. The weight coefficients of the distance metric are determined by minimizing the cross-validation error, with different weight ranges set for dimensions related to cognitive load (ranging from 0.2 to 0.8) and emotional state (updated for at least 2 seconds). A multi-level similarity calculation framework is constructed. Local similarity calculation uses feature distribution thresholds defined in the feature space, while global similarity is weighted and integrated using an attention mechanism. The weight update frequency is consistent with the response time threshold (200ms) for the attention level dimension in the feature space. The calculation of feature similarity takes into account the influence of time scale: short-term similarity uses an exponentially weighted moving average, while long-term similarity employs a multi-scale decomposition approach. An anomaly detection mechanism based on the probability distribution characteristics of the feature space was also established. Recalculation is triggered when a feature deviates beyond the feature distribution threshold defined in the feature space. The stability of the similarity calculation results was verified through Monte Carlo simulations. When 10% random noise was added, the coefficient of variation of the calculated results remained within 0.15, meeting the stability requirements set in the feature space. The final sample matching results include a similarity matrix, anomaly detection flags, and a time series similarity pattern.

[0044] An initial template is generated based on the similarity matrix, anomaly detection markers, and temporal similarity patterns obtained from sample matching. A hierarchical clustering approach is used to process the similarity matrix, and a cluster tree is constructed using the Ward minimum variance method. Clustering considers physical constraints defined in the feature space, such as the value range of cognitive load-related dimensions and the minimum transition time between adjacent cognitive states. Anomaly detection markers are used to exclude unstable samples to ensure the reliability of the clustering results. A pruning algorithm is used to identify sample clusters, and the feature statistics of each cluster must fall within the value range defined in the feature space. During cluster formation, a dynamic programming algorithm is introduced to optimize the selection of split points, referencing the stable intervals in the temporal similarity patterns. A multi-component Gaussian mixture model is constructed for the identified sample clusters, with the initial values ​​of the model parameters determined by the probability distribution characteristics of the feature space. The expectation-maximization algorithm is used to optimize the model parameters to ensure that the model predictions are consistent with the physical meaning of the original features in the feature space. To improve model stability, a bagging ensemble method is used to train multiple sub-models and fuse the results through a voting mechanism. The voting threshold refers to the feature distribution threshold in the feature space. For dynamic features, a recurrent neural network model is constructed, whose structure takes into account the temporal dependency model defined in the feature space. Network training uses a teacher-forcing strategy and incorporates an attention mechanism, with the attention update rate aligned with the response time threshold of the attention-level-related dimension in the feature space. The initial templates generated in this way include static and dynamic feature templates. Each template inherits the key properties of the feature space and also reflects the similarity patterns, abnormal feature distribution, and temporal variation patterns discovered during the sample matching process.

[0045] Step S102 , feature mapping is performed based on the initial template to obtain common features, and individual features are marked based on the common features; fusion rules are constructed using the individual features, a feature set is generated based on the fusion rules, and an identification benchmark corresponding to the feature set is obtained.

[0046] Specifically, feature mapping is performed using the initial template provided in the above steps. The static feature distribution parameters (e.g., mean vector, covariance matrix) and the recurrent neural network model for dynamic features contained in the template are first normalized. Principal component analysis is applied to the static features, leveraging the covariance structure in the template to extract the main feature directions with a cumulative contribution exceeding 85%. During feature extraction, validation is performed according to the physical constraints defined by the template, such as the range of [0.2, 0.8] for the cognitive load metric and a 200ms response time threshold for attention level. To ensure mapping stability, a multi-factor validation mechanism is introduced: first, temporal consistency of the features is verified, requiring the coefficient of variation within a continuous time window (5-minute window length) to be less than 0.15; second, spatial consistency of the features is verified, requiring the difference in feature performance across different data subsets to be less than 10%; and finally, physical consistency of the features is verified to ensure that the extracted feature directions align with the physiological significance defined in the template. For dynamic feature processing, the weight matrix of the template's recurrent neural network structure is analyzed to extract the main patterns with eigenvalues ​​greater than 0.2. Special attention was paid to maintaining key timing constraints within the template, including the minimum time interval between state transitions (>500ms) and the upper limit on response latency (200ms). The template's Gaussian mixture model was analyzed, and distribution patterns with a frequency exceeding 75% were extracted. A generalized probabilistic model consisting of mean, covariance, and mixture weights was constructed. This mapping process was evaluated for generalization performance using 10-fold cross-validation to ensure stable performance across different scenarios. The resulting generalized features include the main feature directions, main patterns, a generalized probabilistic model (mean, covariance, and mixture weights), and cross-validated generalization performance metrics.

[0047] Personalized features are labeled using the universal features obtained in the previous step. User data is projected onto the main feature directions, and the degree of deviation in each direction is calculated. The degree of deviation is assessed using Z-score normalization. For feature points with significant deviation (Z-score > 1.96), their probability density is further calculated using the mean, covariance, and mixing weight of the universal probability model. Points with a probability density below a threshold (P < 0.05) are screened. These feature points exhibit significant personalized characteristics under the universal probability model. To ensure the reliability of feature labeling, time window analysis (window length 30s, step length 5s) is introduced, and stability is assessed using a cross-validated generalization performance metric. Only feature points that are stable across multiple consecutive windows are retained. The selected feature points are clustered using the DBSCAN algorithm (ε = 0.15, MinPts = 5), and each cluster is checked to ensure that it meets the temporal constraints extracted in the first step. In the formation of feature clusters, not only spatial distance but also temporal correlation is considered to ensure temporal continuity of feature points within the cluster. For each generated feature cluster, its statistical properties (mean, variance, skewness, kurtosis) and temporal characteristics (autocorrelation coefficient, cross-correlation coefficient) are calculated to establish a feature description vector for the cluster. Simultaneously, the detailed mapping relationship of each feature cluster in the common feature space (projected coordinates with the main feature direction and similarity with the main pattern) and reliability assessment metrics (sample size, intra-cluster variance, inter-cluster distance) are recorded. This process ultimately yields a set of feature clusters that meet common constraints but are significantly personalized, each with complete descriptive information.

[0048] Fusion rules are constructed based on the personalized feature clusters marked in the previous step. For each feature cluster, the degree of deviation from the corresponding common feature direction is calculated using its detailed mapping relationship in the common feature space, and the fusion weight is set accordingly. The weight setting is adjusted based on reliability assessment indicators (sample size, intra-cluster variance, and inter-cluster distance). An adaptive mechanism is used, so feature clusters with greater deviation and higher reliability receive higher initial weights, but the total deviation must not exceed the range allowed by the common model. A multi-level weight adjustment mechanism is introduced: at the feature level, weights are adjusted based on cluster stability indicators (variance ratio <1.5); at the time series level, weights are adjusted based on the temporal continuity of features (autocorrelation coefficient >0.6); and at the system level, weight distribution is adjusted based on the overall consistency of feature combinations and the completeness of mapping relationships. For static features, a probability density weighting mechanism is constructed based on the statistical properties of feature clusters (mean, variance, skewness, and kurtosis). The probability density is calculated using kernel density estimation, and the bandwidth parameter is optimized through cross-validation. For dynamic features, state transition rules are designed that take into account time series characteristics (autocorrelation coefficient and cross-correlation coefficient). The transition probability matrix is ​​constructed based on the time series statistical properties of feature clusters. These rules directly utilize the characteristics of the personalized feature cluster, ensuring that the salience of personalized features is maintained during the fusion process. A rule conflict detection and resolution mechanism is also established. When different rules conflict, the rule with the highest reliability evaluation index is prioritized. The resulting fusion rules include a weight adjustment mechanism, state transition rules, and a rule conflict detection and resolution mechanism.

[0049] The fusion rules established in the previous step are applied to integrate common and personalized features. Each feature cluster is weighted and combined with the corresponding common features according to the weights set by the weight adjustment mechanism. The combination process adopts a hierarchical strategy: first, basic fusion is performed at the feature level, using the set weights for linear combination; then, adjustments are made at the pattern level to ensure that the fusion results meet the preset physical constraints; finally, optimization is performed at the system level to ensure the consistency of the overall features. The combination process strictly adheres to the constraints in the fusion rules, with feature correlations not exceeding the preset threshold (0.7) and feature variance ratios remaining within a reasonable range (0.5-2.0). For the integration of dynamic features, the state transition rules defined in the fusion rules are used to adjust feature weights based on the observed data at each time point (sampling rate 100Hz). When conflicts between rules are detected, the rule conflict detection and resolution mechanism is activated, and the optimal rule is selected based on reliability assessment metrics. To ensure the stability of the feature set, a sliding window mechanism (window length 5 minutes, step size 1 minute) is introduced for smoothing. For detected outliers (deviating from the mean by more than 3 standard deviations), the anomaly handling mechanism is activated to temporarily reduce the weight of the corresponding feature. The final generated feature set includes the weighted fusion feature distribution, dynamic update strategy and complete constraints, forming a comprehensive feature expression system that maintains the general feature framework and reflects personalized features.

[0050] In some embodiments, obtaining the recognition benchmark corresponding to the feature set includes: performing incremental analysis on the feature set to extract change features; marking dynamic points based on the change features; using the dynamic points to construct an update sequence, and generating a recognition benchmark based on the update sequence.

[0051] Incremental analysis is performed on the feature set generated in the above steps. A benchmark is set based on the weighted fusion feature distribution in the feature set. For static features, the probability density function is used as the judgment criterion, while for dynamic features, the state transition probability is used as the benchmark. Following the dynamic update strategy defined in the feature set, a sliding window (10-minute window length, 1-minute step length) is set for segmentation. The weighted contribution of each feature is calculated using the weighting scheme in the feature set. During feature extraction, the full constraints in the feature set are strictly adhered to, including the feature correlation threshold (0.7) and the feature variance ratio range (0.5-2.0), ensuring that the analysis process does not violate the original physical constraints. Particular attention is paid to the key state intervals defined in the feature set, as changes in these intervals often indicate important state transitions. Multidimensional statistical metrics are calculated for each window: mean, standard deviation, skewness, and kurtosis, and compared to the standard distribution in the feature set. If a feature's statistical metric deviates from its standard distribution by more than 2 standard deviations, and this deviation is consistent across multiple windows, it is marked as a potential change interval. An adaptive change point detection algorithm is applied to these intervals using the update rule of the feature set. Detection parameters are dynamically adjusted based on the historical stability of the features: a higher detection sensitivity (α = 0.01) is used for features with high stability (coefficient of variation < 0.1), while a lower sensitivity (α = 0.05) is used for features with large fluctuations. Multi-scale analysis is also introduced to verify the consistency of changes at different time scales (1 minute, 5 minutes, and 10 minutes). The result is a set of change features, each of which includes its timestamp, change magnitude, duration, and reliability score.

[0052] A dynamic point marking system is constructed based on detected change features (including timestamps, change magnitude, duration, and reliability scores). First, an initial temporal framework is established based on the timestamps of the change features. For each potential change interval detected by the change feature, the feature change trend within a window (±5 minutes) is calculated. Particular attention is paid to intervals in the change feature with long durations (>30 seconds) and significant change magnitudes (>2σ), as these intervals are likely to correspond to important state transition points. During the analysis, the reliability scores of the change features are weighted, with highly reliable change features having greater influence in dynamic point identification. When the comprehensive change index of a time point exceeds a threshold (0.3) and its impact range matches the range predicted by the change feature, it is marked as a dynamic point. A feature description vector is constructed for each dynamic point, containing all the statistical properties of the original change feature and supplemented with the results of temporal correlation analysis. A hierarchical evaluation system for dynamic points is also established: first, the change significance of individual feature dimensions is evaluated, then the coordinated change patterns of feature combinations are analyzed, and finally, the temporal consistency of the changes is verified. This process ultimately generates a sequence of dynamic points, each containing a complete feature description and reliability assessment.

[0053] In-depth analysis is conducted using labeled dynamic point sequences (including complete feature descriptions and reliability assessments). Based on the feature description vectors of dynamic points, a similarity matrix is ​​calculated between them. This similarity calculation considers not only the Euclidean distance of the feature values ​​but also the temporal properties and reliability assessments of the dynamic points. Based on this similarity matrix, a temporal correlation graph of the dynamic points is constructed. The edge weights in the graph reflect both feature similarity and temporal correlation. Dynamic points with high reliability scores (scores > 0.8) are given higher weights for their connections in the graph. Graph analysis algorithms, such as PageRank variants, are applied to identify key dynamic point links. This requires that the similarity between adjacent nodes on a link is greater than 0.7 and that the time intervals meet pre-defined constraints. Pattern analysis is performed on the extracted links, and sequence mining algorithms are used to discover typical state transition patterns. Based on this, a state transition probability matrix is ​​constructed. The matrix elements contain not only transition probabilities but also typical transition timescales and reliability assessments. This analysis ultimately generates a structured update sequence in which each node retains complete feature information and transition relationships from the dynamic points.

[0054] A recognition benchmark is generated based on the constructed update sequence (containing complete feature information and transformation relationships). Transition patterns in the update sequence are classified, with high-frequency and stable transition patterns (occurrence frequency > 0.1, reliability > 0.8) being used as basic recognition units. Each recognition unit inherits the feature threshold, timing constraints, and reliability metrics from the update sequence. For static features, an adaptive threshold update mechanism is established based on the complete feature information retained by nodes in the update sequence, including statistical distribution parameters and temporal variation characteristics. The threshold adjustment step size is positively correlated with the stability of the data observed in the sequence: when the sequence indicates relatively stable features (coefficient of variation < 0.15), a smaller adjustment step size (0.05σ) is used; when features fluctuate significantly, a larger adjustment step size (0.1σ) is used. For dynamic features, a prediction model is constructed directly using the transformation relationships in the update sequence, and the model parameters are determined through maximum likelihood estimation. A multi-level anomaly detection mechanism is also established: detecting abnormal fluctuations of individual features at the feature level, abnormal changes in feature combinations at the pattern level, and abnormal patterns in the transformation sequence at the system level. When an anomaly is detected, the reliability score in the update sequence is used to determine whether to trigger a model update. The generation of the identification benchmark also incorporates a self-validation mechanism, verifying the validity of the benchmark by backtesting historical data in the update sequence. The resulting identification benchmark is an adaptive dynamic system that can adjust the identification strategy in real time based on the changing trends of the characteristics reflected in the update sequence.

[0055] Step S103 , classify scenes according to the recognition benchmark and mark the usage environment; construct a state sequence based on the usage environment, select a feature template according to the state sequence, generate recognition parameters according to the feature template, and generate a recognition solution according to the recognition parameters.

[0056] Specifically, the recognition benchmark generated in the above steps is analyzed for its adaptability to different scenarios. The system then performs multi-dimensional classification on the input data stream based on the feature threshold system and state transition rules defined in the benchmark.

[0057] In some embodiments, the scene classification according to the recognition benchmark includes: dividing the scene into high-sensitivity scene, medium-sensitivity scene and low-sensitivity scene types based on the scene sensitivity level corresponding to the recognition benchmark; constructing a mapping relationship table between the scene and feature configuration corresponding to the recognition benchmark; matching the corresponding scene classification label according to the physical environment parameters and cognitive state parameters of the scene; establishing loading rules for scene classification and feature templates to complete the scene classification.

[0058] For static features, the benchmark's adaptive thresholding mechanism (adjustment step size 0.05σ-0.1σ) was used to calculate the degree of match between feature values ​​and the benchmark interval. When the statistical distribution of a feature matches the benchmark model well (>0.85), the scene features for that time period were extracted. For dynamic features, the benchmark's state prediction model was applied to analyze feature sequences and identify data segments that matched the expected transition patterns. Special attention was paid to the benchmark's anomaly detection results, with detected outliers (deviations >3σ) and outlier sequences considered key markers of scene transitions. In the multidimensional feature space, feature combinations were weighted according to the benchmark's reliability scoring system. This weighting followed the benchmark's multi-layered evaluation mechanism: 40% of the feature-level score, 35% of the pattern-level score, and 25% of the system-level score. Through multi-dimensional correlation analysis, time periods with similar feature patterns and transition characteristics were clustered to form preliminary scene categories. During the clustering process, the DBSCAN algorithm (ε=0.15, MinPts=5) was used to extract stable clusters, and the silhouette coefficient (>0.6) was used to verify the clustering validity. A complete feature description vector is calculated for each scene category, including static distribution characteristics (mean, variance, skewness, kurtosis) and dynamic characteristics (autocorrelation coefficient, periodicity strength, switching frequency). These scene categories form a structured scene classification library.

[0059] The usage environment was modeled based on the scene classification library. The feature description vectors (including both static and dynamic characteristics) within each scene category were thoroughly analyzed, focusing on consistent feature combinations. For static features, the discriminability of feature value distribution across different scenes (Fisher discriminant ratio > 1.5) and information gain (> 0.4) were calculated. For dynamic features, the consistency of feature sequences within a scene (autocorrelation coefficient > 0.6) and the variability between scenes (cross-correlation coefficient < 0.4) were analyzed. During feature extraction, multi-layer evaluation weights (40% for the feature layer, 35% for the pattern layer, and 25% for the system layer) were maintained. A sliding window method (5-minute window length, 30-second step length) was used to capture time-varying characteristics. Wavelet transforms were applied to extract the energy distribution of different frequency bands (δ, θ, α, β, and γ). Features with strong discriminability and meeting the multi-layer evaluation criteria were selected to construct the environment description vector. Kernel density estimation was used for probability distribution modeling, with the kernel bandwidth determined by maximum likelihood cross-validation (range 0.05-0.2). Environmental similarity is calculated using a weighted Mahalanobis distance, with weights directly inherited from the feature discriminability metric used in scene classification. Scenes with similarity exceeding a threshold (0.75) and consistent feature patterns are hierarchically clustered. A cluster tree is constructed using the Ward minimum variance method, and the optimal number of environmental categories is determined using a dynamic tree-cutting algorithm. This ultimately results in a library of environmental types, each of which contains four key elements: a standard feature distribution model, a set of typical transformation rules, a stability assessment metric, and an environmental similarity matrix.

[0060] A state sequence is constructed based on an environment type library. State identification nodes are extracted from the environment type, and the environmental feature distribution model is used as the state determination benchmark. A typical transition rule set is used to determine state persistence constraints. Stability assessment indicators in the environment type library are directly used to assess state reliability, with environmental features with high stability (scores > 0.8) assigned higher state determination weights. State identification utilizes a multi-feature fusion approach, combining time-domain features (average amplitude, volatility) and frequency-domain features (band energy ratio) for comprehensive determination. State duration strictly adheres to the environment definition: a longer duration (> 30 seconds) is used for stable environments (coefficient of variation < 0.15), while a shorter duration (> 10 seconds) is used for changing environments. A Markov transition chain is constructed using the environment type transition rule set. The transition probability matrix is ​​obtained through maximum likelihood estimation, and its stability is evaluated through Monte Carlo simulation (1000 iterations). State transitions must satisfy the constraints defined by the environmental similarity matrix to ensure transition rationality. The system dynamically adjusts state judgment thresholds based on the stability assessment indicators of the environment type library. A strict threshold (±1σ) is used for environments with high stability indicators, while a more relaxed threshold (±1.5σ) is used for volatile environments. An additional verification mechanism is introduced for low-probability transitions (<0.1), requiring that the transition conditions be met for multiple consecutive time windows (≥3). During the state sequence generation process, a three-tiered state assessment system is established: a micro-level assessment of feature consistency (weight 0.3), a meso-level assessment of the rationality of state transitions (weight 0.4), and a macro-level assessment of the overall stability of the sequence (weight 0.3). This process creates a complete state template set that includes state definitions, transition rules, and assessment criteria.

[0061] Identification parameters are generated based on a set of state templates. Each state definition in the state template set is matched with the characteristic patterns in the constructed state sequence. A multi-objective optimization approach is used to optimize the matching process, simultaneously considering three key metrics: distribution matching (required to be >0.8), transition consistency (required to be >0.75), and temporal characteristic consistency (required to be >0.7). The transition rules in the state template set are directly used to construct a state transition verification mechanism for the identification parameters, including transition time constraints, required characteristic conditions for transitions, and a method for verifying post-transition stability. For composite states, a hierarchical matching strategy is introduced, first determining the dominant features and then matching the secondary features. Feature selection is based on the scoring results of a three-layer state assessment system (micro, meso, and macro). The parameter configuration process fully references the state template evaluation criteria: For high-scoring states (>0.85), highly accurate feature criteria (error tolerance <5%) are selected with strict parameter constraints. For medium-scoring states (0.6-0.85), adaptable criteria (error tolerance 5%-15%) are selected with relatively loose parameter ranges. For low-scoring states (<0.6), robust criteria (error tolerance >15%) are selected and necessary error tolerance mechanisms are introduced. Feature conflict resolution utilizes a priority strategy. When different features indicate conflicting states, the final decision is made based on the feature reliability (defined as "stability × discriminability") defined in the state sequence. The state template set's transition rules are applied to configure the recognition system's state tracking algorithm, including state persistence verification, abnormal state filtering, and state sequence smoothing. Based on the selected criteria and state template characteristics, specific recognition parameter configurations are generated. The parameter settings include the feature threshold ranges, weighting coefficients, update conditions, and constraints provided in the state sequence. These configurations are directly inherited from the state template's feature definitions. All parameter configurations are verified and optimized based on actual observation data in the state sequence to ensure the practicality and effectiveness of parameter settings.

[0062] In some embodiments, generating an identification scheme based on the identification parameters includes: performing personalized marking based on the identification parameters to determine the parameter range; constructing adjustment rules based on the parameter range; using the adjustment rules to perform feature configuration, and generating an identification scheme based on the feature configuration.

[0063] The recognition parameters generated in the above steps are analyzed for personalized labeling. First, the characteristic threshold ranges, weighting coefficients, update conditions, and constraints within the recognition parameter set are extracted to establish a parameter characteristic evaluation system. A sensitivity analysis is performed on static parameters, and the degree of their impact on recognition results is assessed through perturbation testing (±10%). The adaptability of the recognition parameters to different user data is analyzed, and the individual variation coefficient (CV) of each parameter is calculated using historical data. Parameters with high CV values ​​(>0.2) are clustered and a K-means approach (K=3-5) is used to extract typical parameter distribution patterns. For dynamic parameters, their update trigger conditions and step size settings are analyzed, and their stability is evaluated under different usage scenarios. During the parameter analysis, a multi-scale evaluation method is introduced to calculate parameter stability indicators within short-term (5 minutes), medium-term (1 hour), and long-term (1 day) time windows. For parameters with high volatility, exponential smoothing (α=0.3) is used to mitigate the impact of short-term fluctuations. Through this analysis, each parameter is assigned a personalized score (0-1). Parameters with high scores (>0.7) require greater adjustment space, while parameters with low scores (<0.3) can adopt a narrower range. Ultimately, the personalized adjustment range for each parameter is determined, including minimum and maximum values, step size, and priority, forming a parameter range definition library.

[0064] A tuning rule system was constructed based on a parameter range definition library. First, for parameters with high personalization scores (>0.7), an adaptive tuning strategy was designed. The rules consisted of trigger conditions, tuning directions, and step size control. Trigger conditions were based on performance evaluation metrics, such as a drop in recognition accuracy exceeding 5% or more than three consecutive false positives. The tuning direction was determined by the correlation between the current parameter value and the target performance, and the optimal parameter change path was explored using Bayesian optimization. For parameters with medium personalization scores (0.3-0.7), scenario-based conditional tuning rules were established, employing different parameter configurations in different environments. For parameters with low personalization scores (<0.3), a minimal intervention principle was established, requiring fine-tuning only when system performance degraded significantly. The rule construction process used the minimum and maximum values ​​in the parameter range definition library to determine the tuning boundaries. The step size determined the accuracy of each adjustment, and the priority level determined the order in which rules were executed. A conflict detection mechanism was introduced. When multiple rules were triggered simultaneously, the final rule to be executed was determined based on the rule priority (1-10) and the current system state. To improve the robustness of the rules, fuzzy logic is used to define the triggering conditions for the rules. Membership functions (trapezoidal or Gaussian) are used instead of hard thresholds to reduce the instability of boundary conditions. Furthermore, correlation adjustment rules between parameters are established. When a parameter changes, related parameters are adjusted synchronously according to a preset ratio, ensuring the overall consistency of the parameter system. This ultimately forms a multi-layered adjustment rule library, comprising adaptive rule sets, scenario rule sets, and association rule sets.

[0065] Feature configuration is performed using a built tuning rule library. Priority and execution order are set for the adaptive rule sets in the tuning rule library, with high-priority rules (related to key features) executed first. The scenario rule sets in the tuning rule library are used to switch parameters for different usage environments, such as using different sensitivity settings in quiet and noisy environments, ensuring that feature configuration adapts to environmental changes. The association rule set ensures that interrelated parameters maintain a reasonable proportional relationship, preventing system instability caused by adjusting a single parameter. For each feature parameter, an appropriate tuning rule is selected based on its personalized score and current performance indicators. Static feature configuration uses a grid search strategy, traversing the parameter range with a preset step size to select the parameter combination with the best performance. Dynamic feature configuration uses an online learning method to dynamically update parameter values ​​based on the tuning rules. Association configuration between features follows the association rule sets in the rule library, ensuring overall coordination of the feature system. To improve configuration efficiency, a hierarchical configuration strategy is adopted: highly sensitive parameters (influence coefficient > 0.8) are configured first. After these parameters are fixed, medium-sensitivity parameters are configured, and finally low-sensitivity parameters are configured. For computationally complex configuration items, approximate calculation methods are employed, such as using quadratic response surfaces instead of full performance evaluation, reducing computational complexity from O(n^3) to O(n^2). A validation mechanism is introduced during the configuration process, assessing the robustness of the configuration through 5-fold cross-validation. A complete feature description is generated for each configuration, including parameter values, adjustment history, and performance evaluation results, forming a feature configuration set.

[0066] Generate a personalized recognition solution based on a feature configuration set. The static and dynamic configurations in the feature configuration set are integrated, combined with detailed feature descriptions of each configuration item (including parameter values, adjustment history, and performance evaluation results) to form a complete recognition parameter system. Performance evaluation results from the feature configuration set are directly used to evaluate solution quality, including metrics such as precision, recall, and F1 score, to ensure that the final solution selects the optimal configuration combination. A comprehensive scenario-configuration mapping table is established to identify the configuration effects corresponding to different scenario feature combinations. This ensures that the system can quickly load the optimal parameter settings for high-sensitivity scenarios (such as fatigue driving), medium-sensitivity scenarios (such as learning), and low-sensitivity scenarios (such as daily activities) defined by the feature configuration. A configuration validity verification mechanism is implemented, including parameter consistency checks, compatibility checks, and performance results checks. The performance evaluation results recorded in the feature configuration set are directly used as a benchmark during the verification process. Solution generation adopts a modular design, with each functional module setting clear parameter ranges based on the performance indicators described in the feature description. The core recognition algorithm uses an ensemble learning approach, combining the results of multiple base classifiers to mitigate the limitations of a single model. To improve system robustness, an exception handling mechanism based on feature configuration verification is introduced. When anomalies in input data or processing are detected, graceful degradation can be performed to maintain basic system functionality. A complete recognition solution is generated for the verified configuration, including a feature extraction module (using optimized feature parameters), a pattern recognition module (based on configured classification rules), and a feedback adjustment module (to achieve dynamic parameter updates). In practical applications, this personalized recognition solution enables AI glasses to accurately identify changes in the user's attention state. For example, in driving scenarios, the system can automatically adjust the warning sensitivity based on the user's fatigue level; in learning scenarios, it can distinguish between states of concentration, understanding, and distraction, and provide corresponding learning suggestions.

[0067] Step S104 , performing trend analysis on the identification scheme, obtaining statistical features, marking pattern intervals based on the statistical features; constructing feature chains based on the pattern intervals, and generating feature combinations based on the feature chains.

[0068] Specifically, trend analysis is performed on the identification scheme of the above steps to obtain statistical features; pattern intervals are marked based on the statistical features; a feature chain is constructed using the pattern intervals; and a stability representation is generated based on the feature chain.

[0069] In some embodiments, the trend analysis of the identification scheme is performed to obtain statistical features, and the pattern interval is marked based on the statistical features, including: obtaining the stability index and performance consistency index of the feature combination in the identification scheme as the statistical features to complete the trend analysis; extracting the dynamic association pattern and weight distribution characteristics corresponding to the feature combination; constructing a feature chain structure with time-series correlation; identifying unstable intervals in the feature chain through an anomaly detection mechanism for marking the pattern interval.

[0070] Perform trend analysis on the identification scheme generated in the above steps. Extract the complete parameter architecture for the scheme's three core components: the feature extraction module, the pattern recognition module, and the feedback control module. Establish a parameter time series database, inheriting the sensitivity grading defined in the identification scheme and setting time windows: a 10-day window for high-sensitivity parameters, a 20-day window for medium-sensitivity parameters, and a 30-day window for low-sensitivity parameters. Perform time series analysis on static parameters, calculating key statistical indicators: mean, standard deviation, coefficient of variation, and rate of change. Analyze the adjustment frequency and amplitude of dynamic parameters, and construct a parameter change model based on the scenario-based adaptive rules in the scheme. Use the ARIMA model to fit parameter change trends, extracting model parameters as trend features. Apply wavelet decomposition to decompose parameter changes into long-term trends and short-term fluctuations. For the feedback control module, analyze its triggering frequency and adjustment effectiveness to assess the system's adaptive capabilities. Introduce seasonality analysis to detect periodic patterns in parameter changes. Spectra are extracted for significant periods (p < 0.01) to calculate the dominant frequency and harmonic ratio. Through these analyses, a complete set of statistical characteristics is obtained, including stability indicators, periodicity indicators and trend indicators, which constitute the mathematical description of parameter evolution.

[0071] Pattern intervals are labeled based on the acquired statistical features. Stability classification is based on the coefficient of variation in the stability indicator and the prediction error of the ARIMA model: high stability (coefficient of variation <0.1 and prediction error <5%), moderate stability (coefficient of variation 0.1-0.3 or prediction error 5%-15%), and low stability (coefficient of variation >0.3 or prediction error >15%). Periodicity classification is based on the power spectrum characteristics and seasonality strength of the periodicity indicator: strong periodicity (significant seasonality and power spectrum peak / average power >5), weak periodicity (seasonality present and power spectrum peak / average power between 1-5), and non-periodicity (no significant seasonality and power spectrum peak / average power <1). Trend classification integrates the ARIMA trend term in the trend indicator and the Mann-Kendall test results to categorize the model into three categories: significant increase, significant decrease, and no significant trend. A classification space for parameter patterns is constructed based on these three dimensions, and interval divisions are determined by cluster centers of statistical features. Parameter distribution characteristics and transition probabilities are calculated for each interval to establish a complete description of the interval characteristics. The bootstrap method is introduced to verify the reliability of interval boundaries and calculate confidence intervals (95%). For regions with fuzzy boundaries, transition zones are defined, and fuzzy set theory is used to describe the multi-interval membership of parameters. The migration patterns of parameters between different intervals are extracted, and a conditional probability matrix is ​​constructed that accounts for temporal characteristics. This ultimately results in a structured library of pattern intervals, each with clearly defined boundaries and characteristic descriptions.

[0072] A feature chain was constructed using a pattern interval library. The transition sequences of parameters between different intervals were analyzed. The boundary definitions in the pattern interval library were used to precisely identify interval transition points. The conditional probability matrix was used to identify the main transition paths: critical transition paths (probability exceeding two standard deviations from the mean of the conditional probability), common transition paths (probability within the mean ± two standard deviations), and rare transition paths (probability below two standard deviations). The dwell time thresholds for identifying stable states were determined based on the boundary definitions and feature descriptions in the pattern interval library: 7 days for highly stable intervals, 5 days for moderately stable intervals, and 3 days for lowly stable intervals. A Markov chain model of parameter evolution was constructed, with the state space defined based on the pattern intervals and the transition matrix inheriting the transition probabilities between intervals. The co-evolutionary relationships between parameters were analyzed, and correlations were grouped based on cross-correlation analysis of temporal features: strongly correlated groups (cross-correlation coefficients > 0.7 and consistently stable) and weakly correlated groups (cross-correlation coefficients fluctuating or between 0.3 and 0.7). A joint Markov model was constructed for the strongly correlated groups, while independent models were maintained for the weakly correlated groups, with conditional constraints based on fuzzy membership. Graph theory algorithms are used to analyze the topological structure of the feature chain, extracting strongly connected components, key paths, and loop structures. Ultimately, a complete feature chain network is generated to describe the parameter evolution process.

[0073] In some embodiments, generating a feature combination based on the feature chain includes: generating a stability representation based on the feature chain; performing feature management based on the stability representation and establishing a feature library; performing pattern marking based on the feature library; constructing a recognition sequence using the pattern marking; and generating a feature combination based on the recognition sequence.

[0074] A stability representation is generated based on the feature chain network. A hierarchical stability analysis is performed on the network structure. Based on the micro, meso, and macro hierarchies defined by the feature chain, the state residence probability, state return probability, and network entropy are calculated. Stability thresholds are calculated for the strongly connected components identified in the feature chain: high stability thresholds (>0.8) are assigned to high-frequency transition paths, medium stability thresholds (>0.6) are assigned to general paths, and low stability thresholds (>0.4) are assigned to low-frequency paths. Key paths extracted from the feature chain are directly used to construct the backbone sequence of state transitions, identifying the system's core stable states and key turning points. Nodes along the key paths are assigned higher stability weights. Cyclic structures are used to identify steady-state oscillation patterns in the system, analyze their periodic characteristics and triggering conditions, and set corresponding stability assessment criteria. For regions of high stability, their attraction domains and convergence rates are analyzed, and a stable state description model corresponding to the feature chain hierarchy is established. In transition regions, the Markov model of the feature chain is extended to analyze its dynamic characteristics, including transition directionality and rate. A stability prediction model based on the hierarchical feature chain is established, and deep learning methods are used to predict the future states of the parameters and their confidence intervals. At the same time, an anomaly detection algorithm is developed based on the abnormal transition patterns of the feature chain to identify abnormal states that deviate from the expected evolution path. A complete stability profile is established for each stable region, including the region's scope, convergence characteristics, anti-interference ability, and state prediction. In the daily use of AI glasses, stability representation helps the system understand user habits, such as identifying the focus ability curve, and provide intelligent cognitive support.

[0075] Feature management is performed in conjunction with the stability representation from the above steps. First, three layers of stability features are extracted from the stability representation: single-parameter stability at the micro level, coordinated stability of parameter groups at the meso level, and system stability at the macro level. Features in highly stable regions (first-order stability > 0.7) are prioritized and used as core components of the feature library. Dynamic evolution patterns, including state transition paths, convergence characteristics, and anomaly detection patterns, are extracted from the stability representation to construct a dynamic feature update mechanism. A hierarchical storage structure is established: frequently used and highly stable features are placed in a fast storage area, while features identified as stable by anomaly detection are given high storage priority. Moderately stable features are placed in the primary storage area, while low-stability features and fluctuating features marked by anomaly detection are placed in a backup storage area. Based on feature access patterns, an adaptive index structure is designed to optimize common query patterns (such as similarity queries and range queries). Feature lifecycle management is introduced, setting expiration dates and update policies for features based on the temporal information in the stability representation and anomaly pattern predictions. The relationships between features are represented using a graph structure, with edge weights determined by the correlation coefficient and anomaly co-occurrence probability in the stability representation. Through these operations, a structured feature library is established, consisting of static feature sets, dynamic evolution rules, and an association network. In daily use of AI glasses, this feature management can dynamically adjust resource allocation based on the stability of user habits. For example, more computing resources can be allocated to stable attention indicator features, improving recognition accuracy and response speed.

[0076] Pattern labeling is performed based on an established feature library. Features are extracted hierarchically from the static feature set in the feature library: high-stability features from the fast storage area are prioritized, followed by medium-stability features from the primary storage area, and finally low-stability features from the backup storage area. A hierarchical clustering method (Ward's minimum variance method) is applied to the selected static feature set to construct a feature tree. Clustering thresholds correspond to the storage levels: a strict threshold (0.8) is used for fast storage features, a medium threshold (0.6) for the primary storage area, and a loose threshold (0.4) for the backup storage area. Each cluster represents a potential pattern class, and a pattern descriptor is constructed by calculating the statistical distribution (mean, variance, skewness, kurtosis) and dynamic characteristics (autocorrelation coefficient, conditional entropy) of the intra-cluster features. Particular attention is paid to the association network in the feature library. A graph community detection algorithm is used to identify strongly associated feature groups, with detection parameters adaptively adjusted based on the storage level. The dynamic evolution rules of the feature library are utilized to analyze the temporal evolution characteristics of the patterns, including pattern duration, transition frequency, and periodicity. A unique identifier is assigned to each identified pattern category, and a multi-level tagging system corresponding to the storage hierarchy is constructed: the fast storage area corresponds to the core tag, the main storage area corresponds to the extended tag, and the backup storage area corresponds to the supplementary tag. The tagging results are verified, and the consistency and reliability of the tags are evaluated through cross-validation (10 folds). Ultimately, a complete set of pattern tags is formed, and each pattern contains a detailed description of the feature distribution and dynamic evolution characteristics. In the learning assistance scenario of AI glasses, this pattern tagging can accurately distinguish different cognitive states, such as subdividing the "concentration" state into sub-modes such as "deep thinking", "information acquisition" and "creative thinking", providing more accurate learning status feedback.

[0077] A recognition sequence is constructed using a pattern marker set. First, based on the feature distribution description and dynamic evolution characteristics of the pattern marker set, a hierarchical recognition framework is constructed in conjunction with storage hierarchy: the core recognition chain is constructed using features from the fast storage area, the auxiliary recognition chain is constructed using features from the primary storage area, and the supplementary recognition chain is constructed using features from the backup storage area. The temporal correlations between patterns at each level are analyzed, and a pattern transition probability matrix defined in the marker set is calculated. The matrix structure corresponds to the storage hierarchy. A hierarchical Markov model is constructed based on this matrix, with each layer capturing transition patterns at different granularities. To improve the model's expressiveness, the feature distribution description from the pattern marker set is introduced as context information. A conditional random field model is constructed, with the state space organized according to the storage hierarchy. The feature function is designed based on the feature distribution description and dynamic evolution characteristics of the pattern. Maximum likelihood estimation is used for model training, with the L-BFGS optimization algorithm employed. For patterns with strong temporal dependencies, a long short-term memory (LSTM) network is introduced, with a network structure corresponding to the storage hierarchy: a deep network is used for features from the fast storage area, a medium-sized network is used for the primary storage area, and a shallow network is used for the backup storage area. A hierarchical confidence assessment mechanism is implemented in the recognition process, with confidence calculated independently at each level and the final result obtained through weighted combination. For low-confidence recognition results, the verification process is initiated layer by layer according to the storage hierarchy. The result is a hierarchical recognition sequence, which includes multiple layers of pattern recognition rules, sequence constraints, and prediction mechanisms. In AI glasses' meeting scenarios, the recognition sequence can track the user's participation status in real time, identifying the state transition sequence from "passive listening" to "active thinking" and then "preparing to speak", and providing meeting content summaries or speech suggestions at appropriate times.

[0078] Feature combinations are generated based on the recognition sequence. For each key node in the recognition sequence, pattern recognition rules are applied to extract the corresponding feature requirements, including required and optional feature sets. Required features are directly selected from the core features of the pattern markers, while optional features are selected based on contextual matching. The feature selection process employs a forward-backward search strategy defined by the sequence constraints. Core features are first added, auxiliary features are added through incremental evaluation, and unnecessary features are removed through redundancy analysis. For dynamic scenarios, a feature adaptation mechanism is implemented to pre-prepare the required feature sets based on the prediction mechanism of the recognition sequence. A feature fusion layer is introduced to integrate the selected features. The fusion method is based on the rule selection in the sequence constraints and includes feature concatenation, weighted averaging, and nonlinear transformation. The fusion parameters are optimized using a grid search (10×10×10 parameter space). To improve the robustness of the combination, multiple versions of feature combinations are constructed to form an integrated recognition framework. Different versions of the combination are optimized for different noise environments and user states. The generated feature combinations are comprehensively evaluated using metrics such as recognition accuracy, computational complexity, memory usage, and response time. Ultimately, a library of feature combinations is generated, each containing a complete list of features, combination method, and applicable conditions. In the sports scenes of AI glasses, the feature combination can adapt to the changes in EEG features under different exercise intensities, such as from rest to light activity to high-intensity exercise, dynamically adjust the feature weights, maintain stable cognitive state monitoring, and help users maintain the best training state and concentration.

[0079] Step S105 , performing pattern analysis on the feature combination, marking an identification sequence; building a mapping relationship based on the identification sequence; utilizing the mapping relationship to perform feature recombination, and generating a personalized pattern based on the feature recombination.

[0080] Specifically, the feature combination library generated in the above steps undergoes pattern analysis. The library extracts the constituent elements, combination methods, and applicable conditions of each feature combination, and integrates the results of multiple versions of the feature combination. A multi-dimensional evaluation of the combination's performance metrics is conducted, including recognition accuracy, computational complexity, response time, and consistency across integrated versions. Leveraging historical usage data, the adaptability of different combinations and their integrated versions in various scenarios is analyzed, and a scenario-combination mapping matrix is ​​established. Similarity between combinations is measured, using the Jaccard coefficient to calculate feature set overlap and the EMD distance to assess differences in feature weight distribution. Spectral clustering is performed based on the similarity results and the grouping results of the integrated recognition framework, with the number of clusters consistent with the number of integrated framework versions, to identify families of combinations with common characteristics. For each family of combinations, shared core features and differential features are extracted. The noise adaptation characteristics of the multiple versions of the feature combination are integrated to construct a template structure for the feature combination. The identified patterns are arranged in chronological order, and the transition patterns between them are analyzed to extract typical recognition sequence templates. These sequence templates are labeled, and a sequence description library containing sequence structure, transition conditions, and performance characteristics is established. In daily interaction scenarios with AI glasses, this recognition sequence marker can capture the process of users' attention shifting from distraction to gradual focus, helping the system identify the best time for interaction and avoiding sending non-urgent notifications when the user needs to concentrate.

[0081] Mapping relationships are constructed based on a library of labeled recognition sequence descriptions. Based on the sequence structure in the library, the feature requirements of its constituent patterns are analyzed to construct a sequence-feature dependency graph. This dependency graph adopts a directed weighted graph structure, with nodes representing features or patterns, edges representing dependencies, and weights reflecting the strength of dependencies. Transformation conditions in the library are used to perform path analysis, extracting the transformation paths from basic features to advanced patterns. Based on the performance characteristics of the library, a forward feature-to-pattern mapping function is established. The network structure corresponds to the combination family classification: each combination family corresponds to a subnetwork, and the subnetwork dimensions are adaptively set based on the complexity of the features within the family, forming an overall hierarchical mapping structure. Simultaneously, a reverse pattern-to-feature mapping is constructed based on the sequence structure and transformation conditions, maintaining a symmetric structure with the forward mapping. The mapping function design considers nonlinear feature interactions and introduces an attention mechanism to highlight the impact of key features identified within the combination family. The robustness of the mapping relationship is assessed by injecting noise that matches the feature distribution of the combination family and testing it based on the performance characteristics of the library. Finally, a bidirectional mapping library is developed, consisting of a forward recognition mapping and a backward inference mapping, each with accuracy and reliability assessments. In the meeting scenario of AI glasses, this mapping relationship can associate the user's attention change pattern with the speaking intention. When the EEG feature sequence of the user preparing to speak is detected, the speech enhancement and recording functions are automatically activated.

[0082] Feature recombination is performed using an established bidirectional mapping library. First, forward recognition mapping is used to analyze user historical data to extract personalized feature fingerprints. Backward inference mapping is then used to verify the consistency of the extracted results. Feature fingerprints contain the typical numerical range, fluctuation characteristics, and combination patterns of user features. These personalized features are compared with standard patterns in the mapping library, and the degree of match and deviation are calculated using accuracy and reliability assessments. For features with high matching, the standard configuration is retained; for features with medium matching, parameters are fine-tuned using forward recognition mapping; and for features with low matching, the feature extraction and combination schemes are reconstructed using backward inference mapping. The feature recombination process utilizes an evolutionary algorithm, with parameters dynamically adjusted based on mapping complexity: the population size is proportional to the number of mapping layers, the number of evolutionary generations is proportional to the number of mapping nodes, and the selection pressure is proportional to the required mapping accuracy. The fitness function integrates accuracy and reliability assessment metrics from the mapping library, and the weights are adaptively adjusted based on mapping importance. A knowledge-guided mechanism is introduced in the evolutionary process, using rules from the mapping library to constrain the search space. Diversity analysis is performed on the evolutionary results, and multiple optimal solutions with their respective advantages are retained as candidates. The result is a set of user-optimized feature recombination schemes, each of which includes a list of features, a combination method, and expected performance. In the context of AI glasses reading, feature recombination can adapt to the reading EEG patterns of different users. Some users experience enhanced alpha waves when focused on reading, while others have a predominance of theta waves. By recombinating features, the system adapts to individual differences and provides personalized reading assistance.

[0083] Generate personalized patterns based on a feature recombination scheme. The feature list from the recombination scheme is applied to the real-time data processing pipeline. A personalized feature extraction-combination-recognition chain is constructed using the scheme's specified combination method, with initial parameters set based on expected performance. Targeted optimization is performed on each link: in the feature extraction stage, filter parameters and window lengths are adjusted to match the user's signal characteristics; in the feature combination stage, weight configuration and fusion methods are optimized to highlight the user's significant features; and in the recognition stage, decision thresholds and discrimination boundaries are adjusted to adapt to the user's category distribution. The choice of feature list directly influences the initial configuration of the processing pipeline; the combination method determines the specific feature fusion algorithm; and the expected performance sets the system's optimization goals and evaluation benchmarks. A pattern adaptability evaluation mechanism is also established, continuously optimizing pattern parameters through online learning. The learning strategy uses incremental updates to ensure that the pattern can smoothly adapt to changing user habits. To improve the generalization of the pattern, domain adaptation technology is introduced to enable the pattern to be transferred across different usage environments. A comprehensive description document is created for the generated personalized pattern, including parameter configuration, performance characteristics, and applicable conditions. A / B testing verifies the performance improvement of the personalized pattern over the generalized pattern, using evaluation metrics such as improved accuracy, reduced response time, and user satisfaction. Ultimately, a complete library of personalized patterns is formed, automatically selecting the most appropriate recognition pattern based on user characteristics and usage scenarios. In the context of AI glasses at work, the personalized pattern can identify each user's unique pre-fatigue EEG signatures—some users exhibit weakened beta waves, others exhibit enhanced delta waves. Based on this, the system provides personalized rest reminders to help users maintain optimal working state and efficiency.

[0084] The provided method has at least the following beneficial effects:

[0085] 1. By constructing personal benchmarks and feature mapping, rapid personalization is achieved under low data volume conditions, reducing the user's initial adaptation cost while retaining the generalization ability of the general model, enabling the system to fully utilize the common features in group data while maintaining personalized recognition accuracy.

[0086] 2. The introduction of scene classification and environmental adaptation mechanisms enables the recognition system to perceive the user's environment and dynamically adjust recognition parameters. By establishing a recognition benchmark, adaptively adjusting the parameter range, and dynamically configuring the recognition scheme, the system can maintain stable recognition performance in complex scenarios such as user activity transitions (such as from sitting still to walking), changing environmental conditions (such as switching between indoor and outdoor), and changing task types (such as from reading to talking). This multi-layered environmental adaptation mechanism addresses the problem of traditional methods experiencing drastic fluctuations in recognition performance due to environmental changes.

[0087] 3. A comprehensive pattern evolution and feature reorganization framework has been established, enabling the system to capture long-term changes in users' EEG patterns and proactively adjust recognition strategies. By establishing a stable representation, dynamically reconstructing feature combinations, and continuously optimizing personalized patterns, the system adaptively tracks long-term changes in users' cognitive habits, attention patterns, fatigue characteristics, and other characteristics. This continuous evolutionary mechanism ensures the system can continuously adapt to changes in user habits, overcoming the limitations of traditional methods that cannot adapt to long-term changes in user habits and maintaining stable recognition results.

[0088] In order to implement the personalized EEG pattern recognition method based on AI glasses corresponding to the above method embodiment, to achieve the corresponding functions and technical effects. Figure 2 , Figure 2 The block diagram of a personalized EEG pattern recognition device 200 provided in an embodiment of the present application is shown. For ease of explanation, only the parts related to this embodiment are shown. The personalized EEG pattern recognition device 200 provided in an embodiment of the present application includes:

[0089] The benchmark establishment unit 201 is used to feature-tag the EEG data collected by the AI ​​glasses and establish a personal benchmark; perform pattern analysis based on the personal benchmark to obtain a feature map; use the analysis results corresponding to the pattern analysis to mark the feature interval to generate a user profile, and obtain an initial template corresponding to the user profile;

[0090] A feature mapping unit 202 is configured to perform feature mapping based on the initial template to obtain common features, mark individual features based on the common features, construct fusion rules using the individual features, generate a feature set based on the fusion rules, and obtain an identification benchmark corresponding to the feature set;

[0091] A scene classification unit 203 is configured to classify scenes according to the recognition benchmark and mark the usage environment; construct a state sequence based on the usage environment, select a feature template according to the state sequence, generate recognition parameters according to the feature template, and generate a recognition solution according to the recognition parameters;

[0092] A trend analysis unit 204 is configured to perform trend analysis on the identification scheme, obtain statistical features, mark pattern intervals based on the statistical features, construct a feature chain based on the pattern intervals, and generate a feature combination based on the feature chain;

[0093] The pattern analysis unit 205 is used to perform pattern analysis on the feature combination and mark the recognition sequence; construct a mapping relationship based on the recognition sequence; use the mapping relationship to perform feature recombination, and generate a personalized pattern based on the feature recombination. The above-mentioned personalized EEG pattern recognition device 200 can implement the personalized EEG pattern recognition method based on AI glasses in the above-mentioned method embodiment. The options in the above-mentioned method embodiment are also applicable to this embodiment and will not be described in detail here. The rest of the contents of the embodiments of this application can refer to the contents of the above-mentioned method embodiment and will not be repeated in this embodiment.

[0094] 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.

[0095] 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.

[0096] 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.

[0097] 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.

[0098] 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.

[0099] 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 various method embodiments when executing the computer program product.

[0100] 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.

[0101] 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.

[0102] 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. A personalized EEG pattern recognition method based on AI glasses, characterized in that: include: Feature labeling of EEG data collected by AI glasses and building a personal benchmark; Performing pattern analysis based on the personal benchmark to obtain a feature map; Using the analysis result corresponding to the pattern analysis to mark the feature interval to generate a user portrait, and obtaining an initial template corresponding to the user portrait; Perform feature mapping according to the initial template to obtain common features, and mark individual features based on the common features; Constructing a fusion rule using the individual characteristics, generating a feature set according to the fusion rule, and obtaining an identification benchmark corresponding to the feature set; Classify the scene according to the recognition benchmark and mark the usage environment; construct a state sequence based on the usage environment, select a feature template according to the state sequence, generate recognition parameters according to the feature template, and generate a recognition solution according to the recognition parameters; Performing trend analysis on the identification scheme to obtain statistical features, and marking pattern intervals based on the statistical features; constructing a feature chain according to the pattern interval, and generating a feature combination according to the feature chain; Performing pattern analysis on the feature combination to mark an identification sequence; and constructing a mapping relationship based on the identification sequence; The mapping relationship is used to perform feature reorganization, and a personalized mode is generated according to the feature reorganization.

2. The method according to claim 1, characterized in that The obtaining of the initial template corresponding to the user portrait includes: Performing region decomposition on the user profile and marking key features to construct a feature space; Performing sample matching based on the feature space; An initial template is generated based on the feature combination of the matching result of the sample matching.

3. The method according to claim 1, characterized in that The obtaining of the recognition benchmark corresponding to the feature set includes: performing incremental analysis on the feature set to extract change features; Marking dynamic points based on the change characteristics; An update sequence is constructed using the dynamic points, and a recognition benchmark is generated according to the update sequence.

4. The method according to claim 1, wherein Generating an identification scheme according to the identification parameters includes: Perform personalized marking based on the identification parameters and determine the parameter range; constructing an adjustment rule based on the parameter range; The adjustment rules are used to perform feature configuration, and an identification scheme is generated according to the feature configuration.

5. The method according to claim 1, wherein Generating a feature combination according to the feature chain includes: generating a stability representation based on the feature chain; Perform feature management based on the stability representation and establish a feature library; Performing pattern marking based on the feature library; constructing a recognition sequence using the pattern marking; A feature combination is generated according to the recognition sequence.

6. The method according to claim 1, characterized in that The feature labeling of EEG data collected by the AI ​​glasses and the establishment of a personal benchmark include: Preprocessing the EEG data to remove physiological artifacts and interference noise; Extracting the time domain features, frequency domain features and time-frequency domain features of the EEG data and performing dimensionality reduction processing; Establishing a characteristic statistical parameter library of the EEG data through standardization processing; analyzing the temporal correlation characteristics and fluctuation characteristics of the characteristic statistical parameter library by combining stratified sampling data under different cognitive states; The characteristic stability index of the characteristic statistical parameter library is calculated according to the time series correlation characteristics and the fluctuation characteristics, and a personal benchmark database including characteristic parameters, probability distribution models and time series correlation characteristics is constructed.

7. The method according to claim 1, characterized in that The classifying the scene according to the recognition benchmark includes: Classifying the scene types into high-sensitivity scene, medium-sensitivity scene and low-sensitivity scene based on the scene sensitivity level corresponding to the recognition reference; Constructing a mapping relationship table between scenes and feature configurations corresponding to the recognition benchmark; The corresponding scene classification labels are matched according to the physical environment parameters and cognitive state parameters of the scene; and loading rules of scene classification and feature templates are established to complete the scene classification.

8. A personalized EEG pattern recognition device, characterized in that: include: A benchmark establishment unit, used to feature label the EEG data collected by the AI ​​glasses and build a personal benchmark; Performing pattern analysis based on the personal benchmark to obtain a feature map; Using the analysis result corresponding to the pattern analysis to mark the feature interval to generate a user portrait, and obtaining an initial template corresponding to the user portrait; a feature mapping unit, configured to perform feature mapping according to the initial template, obtain common features, and mark individual features based on the common features; Constructing a fusion rule using the individual characteristics, generating a feature set according to the fusion rule, and obtaining an identification benchmark corresponding to the feature set; A scene classification unit, configured to classify scenes according to the recognition benchmark and mark the usage environment; construct a state sequence based on the usage environment, select a feature template according to the state sequence, generate recognition parameters according to the feature template, and generate a recognition solution according to the recognition parameters; a trend analysis unit, configured to perform trend analysis on the identification scheme, obtain statistical features, and mark pattern intervals based on the statistical features; constructing a feature chain according to the pattern interval, and generating a feature combination according to the feature chain; a pattern analysis unit, configured to perform pattern analysis on the feature combination and mark an identification sequence; Building a mapping relationship based on the recognition sequence; The mapping relationship is used to perform feature reorganization, and a personalized mode is generated according to the feature reorganization.

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