Electroencephalogram-based user behavior prediction method, device and equipment in intelligent glasses

Abstract

The invention relates to the technical field of brain-computer interfaces, in particular to an electroencephalogram-based user behavior prediction method, device and equipment in intelligent glasses. The method comprises the steps that multi-channel electroencephalogram data of the intelligent glasses are collected, and original signals are obtained; performing data segmentation based on the original signal, obtaining behavior characteristics for sequence analysis, and obtaining behavior samples to construct a behavior library; performing time sequence mapping on the behavior library to obtain a behavior path extraction dependency relationship construction sequence rule; generating a prediction mark according to a sequence rule to perform environment analysis, obtaining scene features to perform association matching to construct a fusion sequence, generating a prediction result to perform feature decomposition, marking a prediction interval to perform rule construction to determine a confidence coefficient, and generating a judgment sequence extraction behavior pattern; obtaining a prediction parameter based on the behavior pattern; and constructing an updating rule according to the prediction parameters, generating a prediction strategy according to the updating rule, and outputting a behavior prediction result according to the prediction strategy. Accurate extraction of behavior intention information in the electroencephalogram signals is achieved.

Description

Technical Field

[0001] This application relates to the technical field of brain-computer interfaces, and particularly to a method, device, and equipment for predicting user behavior based on electroencephalogram in smart glasses. Background Art

[0002] Smart glasses show great application potential in user behavior understanding and intelligent interaction. Traditional behavior prediction methods mainly rely on users' historical behavior data and explicit operation records, and cannot predict users' behavior intentions in advance. Although electroencephalogram signals contain rich behavior intention information, current analysis methods mainly focus on simple intention recognition and are difficult to achieve the prediction of complex behaviors. These problems urgently need to be broken through from multiple technical levels such as signal processing, feature extraction, and prediction models. At the same time, the existing technologies have obvious deficiencies in dealing with the real-time performance, accuracy, and generalization of behavior prediction.

[0003] In practical applications, users' behavior intentions are often affected by various factors, including cognitive state, environmental conditions, and task requirements, etc. Existing prediction models are difficult to effectively integrate these influencing factors, resulting in unreliable prediction results. In addition, the quality of electroencephalogram signals collected by smart glasses is limited by the characteristics of portable devices, with problems such as low signal-to-noise ratio and poor stability, which further increases the difficulty of behavior prediction. Moreover, the individual differences and time-varying characteristics of user behavior patterns also pose great challenges to the design of prediction models. These technical difficulties need to be overcome through systematic solutions.

[0004] Therefore, there is an urgent need for a method to solve at least one of the above problems. Summary of the Invention

[0005] The embodiments of this application provide a method, device, and equipment for predicting user behavior based on electroencephalogram in smart glasses. The method aims to solve the problems that in practical applications, users' behavior intentions are often affected by various factors, including cognitive state, environmental conditions, and task requirements, etc. Existing prediction models are difficult to effectively integrate these influencing factors, resulting in unreliable prediction results. In addition, the quality of electroencephalogram signals collected by smart glasses is limited by the characteristics of portable devices, with problems such as low signal-to-noise ratio and poor stability, which further increases the difficulty of behavior prediction. Moreover, the individual differences and time-varying characteristics of user behavior patterns also pose great challenges to the design of prediction models. These technical difficulties need to be overcome through systematic solutions.

[0006] In a first aspect, the embodiments of this application provide a method for predicting user behavior based on electroencephalogram in smart glasses, including:

[0007] Collect multi-channel EEG data of smart glasses to obtain the original signal; segment the data based on the original signal, generate a signal sequence according to the segmentation result corresponding to the data segmentation, obtain the feature sequence corresponding to the signal sequence, and obtain the behavior characteristics corresponding to the feature sequence;

[0008] Perform sequence analysis based on the behavior characteristics to obtain behavior samples, and construct a behavior library according to the behavior samples; perform temporal mapping on the behavior library to obtain a behavior path; extract dependency relationships based on the behavior path; use the dependency relationships to construct sequence rules; generate prediction markers according to the sequence rules;

[0009] Perform environmental analysis according to the prediction markers to obtain scene characteristics; perform association matching on the scene characteristics, construct a fusion sequence based on the matching result of the association matching, and generate a prediction result according to the fusion sequence; perform feature decomposition on the prediction result, mark the prediction interval; perform rule construction based on the prediction interval; determine the confidence level according to the rule, and generate a decision sequence according to the confidence level;

[0010] Perform behavior analysis according to the decision sequence to extract behavior patterns; obtain prediction parameters based on the behavior patterns; construct an update rule according to the prediction parameters, generate a prediction strategy according to the update rule, and output a behavior prediction result according to the prediction strategy.

[0011] In a second aspect, the present application also provides a user behavior prediction device, including:

[0012] A data acquisition unit, configured to collect multi-channel EEG data of smart glasses to obtain the original signal; segment the data based on the original signal, generate a signal sequence according to the segmentation result corresponding to the data segmentation, obtain the feature sequence corresponding to the signal sequence, and obtain the behavior characteristics corresponding to the feature sequence;

[0013] A sequence analysis unit, configured to perform sequence analysis based on the behavior characteristics to obtain behavior samples, and construct a behavior library according to the behavior samples; perform temporal mapping on the behavior library to obtain a behavior path; extract dependency relationships based on the behavior path; use the dependency relationships to construct sequence rules; generate prediction markers according to the sequence rules;

[0014] An environmental analysis unit, configured to perform environmental analysis according to the prediction markers to obtain scene characteristics; perform association matching on the scene characteristics, construct a fusion sequence based on the matching result of the association matching, and generate a prediction result according to the fusion sequence; perform feature decomposition on the prediction result, mark the prediction interval; perform rule construction based on the prediction interval; determine the confidence level according to the rule, and generate a decision sequence according to the confidence level;

[0015] A behavior analysis unit is configured to perform behavior analysis according to the determination sequence, extract behavior patterns, obtain prediction parameters based on the behavior patterns, construct update rules according to the prediction parameters, generate prediction strategies according to the update rules, and output behavior prediction results according to the prediction strategies.

[0016] In a third aspect, the present application further provides a computer device, including a processor and a memory. The memory is used to store a computer program, and when the computer program is executed by the processor, it implements the method for predicting user behavior based on electroencephalogram in the smart glasses as described in the first aspect.

[0017] This method collects raw electroencephalogram signals through a multi-channel electrode array of a smart glasses, performs segmentation processing on the signals and generates a signal sequence. A high-dimensional time-series feature matrix is extracted from the signal sequence, and a structured feature sequence is constructed through behavior indication markers (such as attention, cognitive load state). The feature sequence is processed in layers (fast response layer, medium-term change layer, long-term trend layer), independent components (such as attention transfer component, cognitive load component) are extracted, and a feature hierarchy is constructed to generate behavior features. Behavior samples are generated based on behavior feature analysis, and a behavior library containing multi-dimensional feature templates is constructed. Temporal mapping is performed on the behavior library, the micro-dependency relationship, meso-combination pattern and macro-trend of the behavior path are analyzed, and temporal rules (such as conditional transition probability) are extracted to generate prediction markers. Environmental analysis is combined with the prediction markers, scene features (such as task requirements, environmental constraint conditions) are extracted, and a fusion sequence is generated through association matching. Feature decomposition is performed on the prediction results, the prediction interval (immediate, medium-term, long-term) is marked, a confidence rule is constructed and a determination sequence is generated. Finally, prediction parameters are extracted through behavior pattern analysis (such as individual difference adaptation mechanism), a dynamically updated prediction strategy is generated, and behavior prediction results are output.

[0018] Through multi-channel signal segmentation, hierarchical processing and component extraction techniques (such as ICA independent component analysis), the problem of unstable signal quality of portable devices is overcome, and the time resolution and expression ability of features are improved. The systematic expression of behavior indication markers (such as attention concentration, cognitive fatigue) and feature hierarchy realizes the accurate capture of complex behavior intentions. The construction of a behavior library based on feature templates and multi-dimensional temporal rules (such as conditional transition probability) enhances the temporal logic and adaptability of prediction. The association matching between environmental features and behavior paths (such as scene constraint condition analysis) improves the reliability and real-time performance of prediction results. Through the feature decomposition of the prediction interval and dynamic update rules (such as confidence determination sequence), the challenges of individual differences and time-varying characteristics of user behavior are solved. Behavior pattern analysis (such as pattern conversion rules) and prediction strategy generation mechanism ensure the stability and generalization ability of the system in complex scenarios.

[0019] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and should not limit this application. Description of the Drawings

[0020] Figure 1 It is a schematic flowchart of the method for predicting user behavior based on electroencephalogram in the smart glasses shown in the embodiments of this application;

[0021] Figure 2 It is a schematic structural diagram of the user behavior prediction device shown in the embodiments of this application;

[0022] Figure 3 It is a schematic structural diagram of the computer device shown in the embodiments of this application. Detailed Embodiments

[0023] In the following description, specific details such as specific system architectures and technologies are presented for the purpose of illustration rather than limitation, so as to thoroughly understand the embodiments of this application. However, those skilled in the art should clearly understand that this application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid unnecessary details from interfering with the description of this application.

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

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

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

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

[0028] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that specific features, structures, or characteristics described in connection with that embodiment are included in one or more embodiments of this application. Thus, statements such as "in one embodiment", "in some embodiments", "in other some embodiments", "in still other embodiments", etc. that appear in different places in this specification do not necessarily all refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized. The terms "comprising", "including", "having" and their variants all mean "including but not limited to", unless otherwise specifically emphasized.

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

[0030] Smart glasses show great application potential in user behavior understanding and intelligent interaction. Traditional behavior prediction methods mainly rely on users' historical behavior data and explicit operation records, and cannot predict users' behavior intentions in advance. Although electroencephalogram (EEG) signals contain rich behavior intention information, current analysis methods mainly focus on simple intention recognition and are difficult to achieve the prediction of complex behaviors. These problems urgently need to be broken through from multiple technical levels such as signal processing, feature extraction, and prediction models. At the same time, there are obvious deficiencies in the existing technologies in terms of the real-time performance, accuracy, and generalization of behavior prediction.

[0031] In practical applications, users' behavior intentions are often affected by various factors, including cognitive states, environmental conditions, and task requirements, etc. Existing prediction models are difficult to effectively integrate these influencing factors, resulting in unreliable prediction results. In addition, the quality of EEG signals collected by smart glasses is limited by the characteristics of portable devices, and there are problems such as low signal-to-noise ratio and poor stability, which further increase the difficulty of behavior prediction. Moreover, the individual differences and time-varying characteristics of user behavior patterns also pose great challenges to the design of prediction models. These technical difficulties need to be overcome through systematic solutions.

[0032] Please refer to Figure 1 , Figure 1 , which is a schematic flowchart of a method for predicting user behavior based on EEG in a smart glass provided by an embodiment of this application. The method for predicting user behavior based on EEG in the smart glass of the embodiment of this application can be applied to computer devices, and these computer devices include but are not limited to devices such as smart phones, laptop computers, tablet computers, desktop computers, physical servers, and cloud servers. As Figure 1 shown, the method for predicting user behavior based on EEG in the smart glass of this embodiment includes steps S101 to S104, which are described in detail as follows:

[0033] Step S101, collect multi-channel EEG data of the smart glasses to obtain the original signal; segment the data based on the original signal, generate a signal sequence according to the segmentation result corresponding to the data segmentation, obtain the feature sequence corresponding to the signal sequence, and obtain the behavioral characteristics corresponding to the feature sequence.

[0034] Specifically, when the smart glasses collect the user's multi-channel EEG data, it is mainly carried out through dry electrode arrays set at key positions such as the temples, forehead, and behind the ears. During the collection process, a differential amplifier circuit is used to preliminarily process the signal, and the sampling frequency is set at 1024 Hz to ensure the time-frequency characteristics of the signal are not distorted. Considering the characteristics of portable devices, an adaptive gain control technology is adopted to dynamically adjust the signal amplitude to keep it within the range of ±100 μV. At the same time, an independent reference electrode technology is introduced to eliminate common-mode interference, and an impedance self-check circuit is used to monitor the contact impedance between the electrode and the skin in real time. When the impedance exceeds 50 kΩ, it is automatically marked as invalid data. In practical applications, the duration of the original signal collection is usually set to 5 - 10 minutes, which can cover the EEG activity characteristics of users in different cognitive states. Through the above collection scheme, a multi-channel EEG original data stream containing complete timestamp, electrode position, gain coefficient, etc. information can be obtained, where the average signal-to-noise ratio of the effective data segment can reach more than 15 dB, the common-mode rejection ratio exceeds 100 dB, and the electrode contact impedance is stable within the range of 20 - 30 kΩ. These original data are continuously stored at a frequency of 1024 Hz, and each data point contains the voltage values of 16 channels and the corresponding status marks.

[0035] Based on the obtained high-quality original data stream, the system starts to perform adaptive data segmentation processing. First, the sliding window method is used to preliminarily segment the signal, and the window length is set to 2 seconds with an overlap rate of 50%, which can ensure the continuity of the signal. On this basis, a multi-scale entropy value analysis method is introduced to evaluate the complexity change of the signal. When the entropy change rate exceeds 20%, it is marked as a segmentation point. At the same time, the energy distribution of each frequency band is calculated through wavelet packet decomposition, focusing on the energy changes of typical EEG rhythms such as δ (1 - 4 Hz), θ (4 - 8 Hz), α (8 - 13 Hz), and β (13 - 30 Hz). When significant energy changes occur simultaneously in multiple frequency bands (the change amplitude exceeds 2 times the standard deviation of the mean), the effectiveness of the segmentation position is further confirmed. During the segmentation process, the system establishes a complete anomaly detection mechanism to identify various interference types including electromyogram artifacts, electrooculogram artifacts, and device shaking. When an anomaly is detected, the affected data segment will be automatically marked and isolated. After this segmentation processing flow, the original data stream is segmented into a series of data segments with uniform length and relatively stable content. Each data segment carries detailed energy distribution characteristics, entropy value characteristics, and anomaly marking information, with an average segmentation length of 2 seconds, an overlapping interval of 1 second between adjacent segments, and the signal stability index within a single segment remaining above 0.85.

[0036] For these normalized data segments, the system further performs signal sequence generation processing. First, through precise time alignment technology, it ensures that all sequences have a unified time reference and the number of sampling points. Then, five-layer wavelet decomposition based on the sym4 wavelet basis function is used for noise reduction processing, and the soft threshold method is adopted to optimize the wavelet coefficients, which can effectively suppress more than 90% of the high-frequency interference components. On this basis, the system introduces the Hilbert transform to calculate the instantaneous phase and instantaneous frequency characteristics of the signal, obtaining a complete description of the time-varying characteristics. Considering individual differences, Z-score normalization processing is performed on the signal amplitudes of all channels to make their mean value 0 and standard deviation 1, thereby establishing a unified amplitude measurement standard. In the generated standard signal sequence, in addition to the normalized electroencephalogram data of 16 channels, it also contains rich auxiliary features, such as phase synchronization index, band energy ratio, channel coupling strength, etc. The system calculates a series of quality assessment indicators for each sequence, including signal-to-noise ratio (SNR), signal smoothness index, and artifact contamination rate. Among them, the SNR threshold is set to 10 dB, and the upper limit of the artifact contamination rate is 30%. Sequences exceeding these thresholds will be marked as low-quality sequences. All sequences are numbered and stored in chronological order, forming a structured signal sequence database. After complete processing, about 1024×16×4 bytes of high-quality signal sequences can be generated per second of electroencephalogram data, which contains information at three levels: original data, feature data, and quality index data. The time resolution of the data reaches the millisecond level, and the spatial resolution covers the entire scalp area, effectively meeting the accuracy requirements for subsequent analysis.

[0037] In some embodiments, obtaining the feature sequence corresponding to the signal sequence includes: dividing the signal sequence into windows, extracting temporal features; performing behavior indication marking based on the temporal features; and generating the feature sequence according to the marking results corresponding to the behavior indication marking.

[0038] The system receives the standard signal sequence database generated in the above steps, which includes auxiliary features such as 16-channel normalized EEG data, phase synchronization index, band energy ratio, channel coupling strength, etc., as well as signal quality evaluation indicators such as SNR, signal stationarity index, and artifact contamination rate. When performing window partitioning on these signal sequences, an adaptive multi-scale window strategy is adopted. The system first screens high-quality data segments based on SNR and artifact contamination rate. Data segments with SNR lower than 8 dB or artifact contamination rate exceeding 30% are marked as low confidence. The basic window length is set to 500 ms, the sliding step is 100 ms, and the Hanning window function is used to reduce edge effects. At the same time, a dynamic window adjustment mechanism is introduced to automatically adjust the window size based on the local stability of the signal. Within each window, the system calculates time series feature parameters, including frequency domain features such as power spectral density, sub-band energy ratio, spectral entropy, etc., and statistical features such as mean, variance, kurtosis, skewness, etc. The system directly integrates the phase synchronization index provided in the above steps to evaluate the information flow between channels, uses the existing band energy ratio as the basis for cognitive state evaluation, and analyzes the cooperative pattern of brain region activities based on the channel coupling strength. The time-frequency diagram is obtained through short-time Fourier transform, and the energy change trends of four typical frequency bands, namely δ, θ, α, and β, are extracted by combining wavelet packet decomposition. For example, in the scenario where the user is concentrating on reading an e-book, the system can accurately capture the continuous enhancement process of the β-band energy, and the window length is automatically adjusted to better retain the EEG features related to attention. The processing results of each window form a feature vector, including frequency domain features and statistical features. After complete window partitioning and feature extraction, the system obtains a high-dimensional time series feature matrix, and the time resolution of the features reaches 100 ms.

[0039] Based on the high-dimensional time-series feature matrix obtained in the previous stage, the system begins to execute the behavior indication marking process. First, the principal component analysis (PCA) method is used to reduce the dimension of these high-dimensional time-series features, retaining the principal components that explain the main variance. Then, a time-series pattern mining algorithm is adopted to identify the key behavior indications in the feature sequence, including typical patterns such as concentration, cognitive load changes, and emotional fluctuations. The system designs fuzzy inference rules to establish a mapping relationship between time-series features such as the energy ratio of different frequency bands and the change of spectral entropy and specific behavior indications. For example, when the β / α energy ratio continues to increase and the energy in the θ band is relatively stable, it is marked as the concentration state; when the energy in the α band drops sharply and is accompanied by an increase in the energy in the θ band, it is marked as the state of increased cognitive load. In actual application scenarios, such as when users are dealing with complex data analysis tasks, the system can identify when users enter the deep thinking state (obvious α wave suppression) and when cognitive fatigue occurs (significant enhancement of θ waves) by monitoring the dynamic changes of these EEG features, so as to provide user state feedback for smart glasses. Through this marking mechanism, the system adds corresponding behavior indication labels to each window feature vector, including basic behavior states (such as focused, relaxed, fatigued, etc.) and transition states (such as attention transfer, task switching, etc.). The accuracy of the marking is directly related to the EEG change patterns captured in the high-dimensional time-series feature matrix. The richer the feature expression, the higher the accuracy of state recognition. Finally, a feature matrix containing behavior indication markings is output, where each time window is associated with corresponding basic behavior state and transition state labels.

[0040] Based on obtaining a feature matrix containing behavior indication markers, the system begins to construct a structured feature sequence. First, perform temporal correlation analysis on the feature vectors to calculate the feature change rate and behavior indication transition probability between adjacent windows. The system pays particular attention to the basic behavior states and transition state labels in the matrix, and analyzes the temporal distribution laws and transition characteristics of these states. Use the dynamic time warping (DTW) algorithm to align feature patterns at different time scales, while retaining the label information of the basic behavior states and transition states. Use a sequence encoding method to convert the features and these state labels into a unified sequence representation. During the encoding process, introduce a hierarchical feature organizational structure, stratify the principal component features according to the information contribution degree, and embed behavior indication labels in each layer, including the complete information of the basic state and transition state. This hierarchical structure shows good adaptability in practical applications. For example, in the scenario where a user participates in a video conference, the system can flexibly switch the focus of attention on feature levels: when the user is in the listening state (basic behavior state), it gives priority to the feature level related to attention maintenance; when the user needs to switch from listening to speaking (transition state), it then turns to the feature level related to cognitive processing to achieve the capture of the user's behavior intention. At the same time, establish a sequence indexing mechanism to record the temporal dependence relationship between features and the evolution law of behavior indications. For example, the system can track the gradual change process (transition state) of the user from a focused working state (basic state) to a fatigued state (basic state), and capture features such as the change trend of alpha wave energy and the change of beta wave energy. The finally generated feature sequence is stored in a sparse matrix format, and each sequence element contains three parts: a feature vector, a behavior indication probability distribution (reflecting the possibility of the basic state and transition state), and temporal correlation information, forming a structured feature representation.

[0041] In some embodiments, obtaining the behavior features corresponding to the feature sequence includes: performing hierarchical processing on the feature sequence to obtain basic features; performing component extraction based on the basic features; constructing a feature hierarchy using the components obtained by the component extraction; and generating the behavior features according to the feature hierarchy.

[0042] The system receives the feature sequence generated by the above steps, which contains feature vectors, probability distribution of behavioral indicators, and temporal association information. Based on this sequence information, the system first performs hierarchical processing to obtain basic features. Specifically, the temporal change characteristics of the feature vectors are used for hierarchical division, and the probability distribution of behavioral indicators is used as the basis for stratification. The feature sequence is divided into three time scale layers using an adaptive hierarchical strategy: the rapid response layer processes instantaneous feature changes within 100ms, such as EEG oscillation mode conversion; the medium-term change layer focuses on feature evolution within 1-5 seconds, such as changes in attention level; and the long-term trend layer tracks continuous changes over 10 seconds, such as fatigue growth. In the video learning scenario, the system uses the rapid response layer to capture the user's instantaneous attention changes on important knowledge points, evaluates the progress of knowledge understanding through the medium-term change layer, and monitors the overall learning status with the long-term trend layer. The system also establishes an inter-layer information transmission mechanism to realize the linkage update of features at different time scales. Based on this hierarchical processing, the system obtains a set of hierarchical basic features containing temporal associations.

[0043] Based on the above three layers of basic features, the system performs component extraction on each layer independently. For the rapid response layer features, the independent component analysis (ICA) method is used to separate the independent components related to instantaneous behaviors such as attention transfer component and decision control component; for the medium-term change layer features, the components related to continuous behaviors such as cognitive load change component and task engagement component are extracted; for the long-term trend layer features, the components related to state evolution such as fatigue accumulation component and emotional tone component are obtained. In the office scene, the system can extract the attention component (rapid layer) of text scanning, the memory load component (medium-term layer) of text comprehension, and the continuous component (long-term layer) of cognitive engagement from the process of users processing multi-document tasks. At the same time, the system calculates the correlation coefficient matrix between components at different levels, quantifies the correlation strength between components, and provides a data basis for the subsequent feature level construction. Through this component extraction process, the system obtains a set of independent components and their correlation relationships that describe the behavioral characteristics of different time scales.

[0044] Based on the set of extracted independent components and their associated strength matrix, the system begins to construct the hierarchical relationship of features. Using the hierarchical clustering algorithm, the components with associated strength higher than the threshold are organized into feature clusters, and each feature cluster contains a combination of components from different time scales but with related behaviors. The system calculates the hierarchical distance between components according to the correlation coefficient matrix and uses a tree structure to save the clustering results. For example, based on the high correlation (correlation coefficient > 0.8) between the attention transfer component of the fast layer and the attention persistence component of the medium term layer, the system constructs an attention feature cluster; similarly, based on the association between the decision control component and the task engagement component (correlation coefficient > 0.75), a task execution feature cluster is constructed. For each feature cluster, the system calculates its internal average association strength and external discrimination for subsequent feature fusion. For example, in an online meeting scenario, the system uses the association relationship between components to construct a complete hierarchical structure of interactive behavior features: organizing the voice attention component, the expression change component (fast layer, correlation coefficient 0.82), the dialogue participation component (medium term layer, correlation coefficient 0.78), and the meeting engagement component (long term layer, correlation coefficient 0.71) together to form a feature cluster reflecting meeting participation. The system records the hierarchical position, the list of components included, and the associated strength matrix of each feature cluster, providing structured input for the construction of the feature fusion network.

[0045] Using the above - constructed feature clusters and their complete hierarchical organization relationships, the system begins to generate behavioral features. First, a local fusion network is established according to the internal structure of each feature cluster: The system uses the list of components recorded in the feature cluster as network nodes and sets the connection weights between nodes using the coefficients in the association strength matrix. For example, for the attention feature cluster, the connection weight between the attention - transfer component in the fast layer and the attention - persistence component in the mid - term layer is set to 0.8 (corresponding to its correlation coefficient). Then, based on the hierarchical position of the feature cluster, cross - cluster connections are constructed: When two feature clusters are close in the hierarchical structure (hierarchical distance < 2) and have common components, a connection path is established between them, and the connection strength is determined by the correlation coefficient of the common components. The system designs an adaptive weight mechanism that can dynamically adjust the connection strength in the network according to the current task scenario. For example, in a reading - comprehension task, the system increases the connection weight between the attention feature cluster and the cognitive - processing feature cluster (from the baseline value of 0.6 to 0.8); in a visual - search task, it highlights the connection between the attention feature cluster and the visual - processing feature cluster (weight increased from 0.5 to 0.7). In a two - hour programming course, the system can real - time capture the changes in students' concentration when writing code (attention feature cluster), the mental activity when understanding algorithm difficulties (cognitive - processing feature cluster), and the learning enthusiasm for the entire course (emotional - state feature cluster). When students are debugging complex programs, the system increases the weight of the cognitive - processing feature cluster; when students participate in classroom interactions, it highlights the role of the emotional - state feature cluster, so as to accurately evaluate students' learning status and participation. At the same time, the system introduces a context - awareness mechanism to analyze the dynamic changes of the feature - fusion network in a continuous time window and capture the evolution trend of behavioral features. Through this multi - level fusion and dynamic analysis based on feature clusters, the system finally generates a complete set of behavioral features including behavioral - state descriptions and trend predictions.

[0046] Step S102: Conduct sequence analysis based on behavioral features to obtain behavioral samples, construct a behavior library according to the behavioral samples; perform temporal mapping on the behavior library to obtain a behavior path; extract dependency relationships based on the behavior path; construct sequence rules using the dependency relationships; generate prediction markers according to the sequence rules.

[0047] Specifically, the system receives the set of behavioral features generated in the above steps, determines the initial behavioral types using behavioral state descriptions, providing basic labels for sequence analysis. At the same time, it integrates trend prediction information to guide the selection of window lengths, using shorter windows for regions with obvious predicted trend changes to capture subtle variations. Based on these multi-level features, the system starts to perform sequence analysis to label behavioral samples. It uses multi-scale sliding windows for feature segmentation, with the window length adaptively adjusted between 2 seconds and 10 seconds according to the behavioral type. For the feature sequences within each window, the system calculates the stability index and jump features of the feature clusters respectively to identify the key transition points of the behavioral patterns. Meanwhile, a multi-dimensional analysis method is introduced, combining the feature changes in three dimensions of attention, cognition, and emotion to determine the boundaries of behavioral samples. In an actual teaching scenario, such as when students are watching an online video course, the system can accurately capture typical segments of learning behaviors: when students encounter difficulties and repeatedly watch, it can identify a characteristic combination where the attention feature cluster shows a highly concentrated state, the cognitive processing feature cluster shows an increased load, and the emotional state feature cluster shows an enhanced learning engagement; when students understand a knowledge point, it can capture characteristic changes where the attention feature cluster relaxes moderately, the cognitive processing feature cluster has a decreased load, and the emotional state feature cluster remains stable. The system also establishes a behavioral continuity verification mechanism to ensure the temporal integrity of the labeled samples by analyzing the feature change trends between adjacent samples. Through this multi-dimensional sequence analysis, the system obtains a series of behavioral sample units with complete feature labels.

[0048] In some embodiments, constructing the behavior library according to the behavioral samples includes: constructing a feature template based on the behavioral samples; performing data matching using the feature template; and generating the behavior library according to the matching results corresponding to the data matching.

[0049] Based on the above-mentioned complete behavioral sample units with tags, the system begins to construct feature templates. First, the hierarchical clustering method is used to analyze the behavioral samples, and the samples are clustered into different behavioral prototypes according to the similarity of feature combinations. Each behavioral prototype contains the typical feature sequences and variation rules of this type of behavior in three dimensions: attention, cognition, and emotion. Then, the system analyzes the temporal dependence relationship of each feature cluster in the behavioral prototype, extracts the key feature combinations that can characterize the behavioral characteristics, and constructs the template feature vector. In the telecommuting scenario, the system extracts multiple feature templates from the user's daily work behaviors: such as the focused work template (characterized by the continuous concentration of the attention feature cluster, the stable load of the cognitive processing feature cluster, and the positive engagement shown by the emotional state feature cluster), the task switching template (characterized by the brief dispersion of the attention feature cluster, the fluctuating load of the cognitive processing feature cluster, and the slight fluctuation of the emotional state feature cluster), the rest and recovery template (characterized by the relaxation of the attention feature cluster, the reduction of the load of the cognitive processing feature cluster, and the tendency of the emotional state feature cluster to be peaceful), etc. The system also establishes a template optimization mechanism, and continuously adjusts and refines the template features by analyzing the matching degree of different behavioral samples to the same template. Through this iterative optimization, a set of feature template libraries that can accurately describe different behavioral states is formed.

[0050] Using the constructed feature template library, the system performs the data matching process. A multi-dimensional feature matching algorithm based on dynamic time warping (DTW) is designed to compare the real-time collected behavioral feature sequences with the feature templates in the template library. The matching process not only considers the similarity of feature values, but also introduces the matching of temporal variation patterns, and determines the best matching result by calculating the alignment degree of the feature sequence and the template in the time dimension. At the same time, a fuzzy matching mechanism is established, and the situation of partial missing or uncertainty of features is processed in a confidence-weighted manner. For example, in the intelligent classroom environment, when students are performing experimental operations, the system can perform real-time matching of their current multi-dimensional feature sequences with the existing experimental operation templates: by comparing the attention allocation feature sequence (reflecting the operation concentration), the cognitive processing feature sequence (reflecting the understanding degree), and the emotional state feature sequence (reflecting the operation confidence), it is judged whether the students correctly master the experimental essentials and whether they encounter operation difficulties. The system also establishes an early warning mechanism for behavioral patterns based on the temporal variation of the matching results. When it is detected that the deviation of the feature sequence from the standard template exceeds the threshold, the potential problem areas are marked in time. Finally, the system generates a set of real-time matching results with confidence evaluation.

[0051] Based on the above real-time matching results, the system begins to generate a complete behavior library. First, hierarchical processing is performed according to the confidence evaluation of the matching results. The matching results with high confidence (>0.8) are directly incorporated into the core patterns of the behavior library. The results with medium confidence (0.6 - 0.8) are candidate patterns that need to be further verified. The results with low confidence (<0.6) are only used as references and are not directly incorporated into the behavior library. Then, statistical analysis is carried out on the matched behavior patterns, calculating the occurrence frequency, duration, and transition probability of each pattern, and establishing a multi-dimensional index structure for the behavior patterns. At the same time, the temporal correlation and causal relationship between different behavior patterns are analyzed, and a behavior state transition network is constructed to describe the evolution law of the behavior patterns. In the university classroom scenario, the system has accumulated rich behavior pattern data through long-term observation of students' learning behaviors, including active participation patterns (characterized by highly concentrated attention, active cognitive processing, and positive emotional states), deep thinking patterns (characterized by continuous and stable attention, high cognitive processing load, and focused emotional states), attention dispersion patterns (characterized by frequent jumps in attention, shallow cognitive processing, and fluctuating emotional states), etc. The system systematically organizes these behavior patterns and establishes a multi-level behavior knowledge structure: the bottom layer stores the original feature templates and matching rules, the middle layer records the conversion rules and triggering conditions of the behavior patterns, and the top layer summarizes the long-term evolution trends and influencing factors of the behavior patterns. This hierarchical behavior library can not only support real-time behavior recognition but also provide knowledge support for long-term behavior analysis and prediction.

[0052] In some embodiments, the temporal mapping of the behavior library to obtain a behavior path includes: constructing a multi-dimensional temporal correlation map based on the behavior library for performing micro-dependency analysis, meso-combination pattern recognition, and macro-trend tracking of the behavior path; establishing a temporal causal network based on the multi-dimensional temporal correlation map for calculating the conditional transition probability and temporal correlation coefficient of adjacent behavior patterns; identifying typical combination patterns and variant forms in the behavior path according to the conditional transition probability and temporal correlation coefficient, and generating a behavior path including direct dependency relationships and long-term evolution laws.

[0053] The system receives the multi-level behavior library generated in the above steps, and the system starts to execute temporal mapping to obtain the behavior path. First, it uses the underlying feature templates and matching rules to identify the current behavior pattern and mark its position in the behavior sequence; then it analyzes the temporal transition characteristics between patterns through the behavior pattern conversion rules and triggering conditions recorded in the middle layer; finally, it combines the long-term evolution trend and influencing factors at the high level to predict the overall development direction of the behavior sequence. Using time series analysis methods, the system calculates the transition probability matrix and time interval distribution of behavior patterns, and establishes a dynamic behavior index based on time windows. In the remote learning scenario, the system conducts a refined temporal analysis of a student's daily learning behavior: for example, the learning behavior path in the morning usually shows "preview preparation (15 - 20 minutes) - in-depth learning (45 - 60 minutes) - short break (5 - 10 minutes) - knowledge consolidation (30 minutes)"; in the afternoon, the learning behavior often appears as "shallow learning (20 - 30 minutes) - attention dispersion (5 - 10 minutes) - forced concentration (15 - 20 minutes)". The system also combines the knowledge structure in the behavior library to analyze the behavior distribution at different time scales: from the hourly change in learning concentration, to the daily learning efficiency fluctuation, and then to the weekly formation of learning habits. By establishing a multi-dimensional temporal correlation map of behaviors, the system not only captures the immediate conversion characteristics of behavior patterns, but also discovers potential periodic laws, such as the high-efficiency learning window that appears at specific times of each day, or the attention fluctuation pattern that appears on specific days of each week.

[0054] Based on the obtained behavior path and its multi-dimensional temporal correlation map, the system begins to deeply extract dependency relationships. A three-layer dependency analysis framework is constructed, targeting the behavior associations at the micro, meso, and macro levels respectively. At the micro level, the system applies association rule mining algorithms to analyze the direct dependency relationships between adjacent patterns in the behavior path. For example, after the focused learning pattern, the system calculates the conditional probabilities of different subsequent behaviors such as short breaks, knowledge consolidation, and attention dispersion, and considers the influence weight of the time interval. At the meso level, the system uses sequence pattern mining methods to identify typical combined patterns in the behavior path. Through sliding time window analysis, it discovers the repeated patterns and variant forms of the behavior sequence. For example, the standard pattern of "preview - learning - review" may vary due to fatigue to "preview - learning interruption - forced review". At the macro level, the system establishes a temporal causal network to track the evolution law of long-term behavior trends. In the intelligent office environment, the system has discovered a series of key dependency chains through long-term observation: for example, high-quality morning meeting discussions (8:30 - 9:00) can improve the work concentration in the morning (9:00 - 11:30), which in turn promotes the task completion efficiency in the afternoon (14:00 - 17:00), and this chain reaction has been verified in 80% of working days. For each dependency relationship, the system calculates the temporal correlation coefficient and conditional transition probability, and establishes a quantitative dependency strength evaluation system.

[0055] Using the extracted multi-level dependency relationships, the system begins to construct sequence rules. First, the direct dependency relationships at the micro level are transformed into a basic rule set. The typical combination patterns at the meso level are used to construct composite rules, and the temporal causal network at the macro level forms long-term prediction rules. The system assigns initial weights to each rule based on the temporal correlation coefficient and conditional transition probability. Strong dependency relationships with a correlation coefficient higher than 0.7 are preferentially converted into core rules. Each rule contains three core elements: a trigger condition (current behavior state and environmental factors), a time constraint (effective action time window), and an expected result (subsequent behavior probability distribution). Then, through a rule optimization algorithm, the rule set is refined and enhanced: rules with similar trigger conditions and expected results are merged, redundant rules with a confidence level lower than the threshold are deleted, and necessary environmental constraint conditions are added. In the classroom teaching scenario, the system summarizes a series of high-confidence behavior sequence rules based on a large amount of teaching observation data. For example, "When a student has been continuously concentrating on learning for 45 ± 5 minutes and has had no short breaks recently, the probability of a decrease in attention will increase to 80% within the next 10 minutes. It is recommended to immediately arrange a 5- to 10-minute relaxation activity." and "After 20 minutes of theoretical knowledge explanation, if relevant practical operations are immediately arranged, the probability of an increase in the depth of students' knowledge understanding will reach 75%, and this improvement effect will remain valid within the next 2 hours." The system establishes a complete evaluation file for each rule, recording the number of times the rule is triggered, the success rate, the applicable conditions, and failure cases. Through continuous data accumulation and feedback analysis, the accuracy and practicality of the rules are continuously optimized.

[0056] Based on the constructed sequence rule library, the system starts to generate prediction tags. During the rule selection process, the system determines the reliability weight according to the trigger times of the rules, evaluates the prediction accuracy based on the historical success rate, filters the rules relevant to the current scenario through the applicable conditions, and analyzes the failure cases to avoid potential misjudgments. The system designs a multi-layer prediction evaluation framework, combining rule matching with historical data analysis. At the bottom layer, the system monitors the current behavior state in real time and calculates the matching degree with each rule; at the middle layer, the system combines historical behavior data and uses the dynamic programming algorithm to search for the optimal prediction path among multiple possible behavior paths; at the top layer, the system considers environmental factors and individual differences and makes adaptive adjustments to the prediction results. In practical teaching applications, such as online programming courses, the system can perform multi-dimensional behavior predictions. For example, when it is detected that a student's attention begins to disperse while dealing with complex algorithm problems (rule matching degree > 90%), the system will predict the possible behavior sequence within the next 30 minutes: first is the thinking block stage (occurrence probability 85%, duration 10 - 15 minutes), then is the code debugging difficulty period (occurrence probability 75%, duration 15 - 20 minutes), and finally may lead to learning frustration (occurrence probability 65%). Each prediction tag contains four basic attributes: behavior type, occurrence probability, expected duration, and prediction confidence. Through this rule-based prediction mechanism, the system can generate structured prediction tags for each time window, providing basic data support for subsequent environmental analysis and behavior prediction.

[0057] Step S103, perform environmental analysis according to the prediction tags to obtain scenario features; perform correlation matching on the scenario features, construct a fusion sequence based on the matching results of the correlation matching, generate a prediction result according to the fusion sequence; perform feature decomposition on the prediction result, mark the prediction interval; construct rules based on the prediction interval; determine the confidence according to the rules, and generate a decision sequence according to the confidence.

[0058] Specifically, the system receives the prediction tag information generated in the above steps, including behavior type, occurrence probability, expected duration, and prediction confidence. Based on these prediction tags, the system starts to perform environmental analysis to obtain scenario features.

[0059] In some embodiments, the performing environmental analysis according to the prediction tags to obtain scenario features includes: constructing an environmental feature library, extracting the correlation features between environmental conditions and task requirements; establishing a mapping relationship between environmental parameters and behavior indication labels, and analyzing the influence weight of scenario elements on the behavior path; matching the real-time environmental data with the historical scenario patterns in the feature template library through a correlation matching algorithm to generate a scenario feature vector including environmental constraint conditions and behavior trigger mechanisms.

[0060] A multi-dimensional environmental perception framework is established, including physical environment monitoring (light, temperature, noise, etc.), social environment analysis (personnel distribution, interaction intensity, etc.) and task environment assessment (difficulty, urgency, importance, etc.). The system adopts a stratified sampling strategy to dynamically adjust the collection frequency of environmental information of different dimensions according to the behavior type and duration in the prediction tag. An adaptive sampling algorithm is used for physical environment parameters to increase the sampling frequency at key time points (such as reducing the sampling interval to 100ms before the predicted behavior occurs) and appropriately reduce the frequency in the stable stage (sampling interval increased to 1s); event-triggered sampling is used for social environment characteristics, and data is collected immediately when key events such as personnel interaction and position change are detected; periodic sampling is used for task environment characteristics, and task progress and difficulty assessment is performed every 5 minutes. The system also designs a data quality control mechanism to screen out reliable environmental features through temporal correlation analysis and spatial consistency test of sensor data. The final constructed scene feature vector adopts a multi-level coding structure, with the bottom layer recording the original sensor data, the middle layer storing feature statistics, and the top layer saving the trend of environmental changes, forming a comprehensive scene description.

[0061] Based on the acquired scene feature vector and prediction tag information, the system starts to perform association matching. The system designs a hierarchical processing strategy for the multi-level coding structure of scene features: noise filtering and anomaly detection algorithms are applied to the bottom-level raw sensor data to ensure data quality; normalization and dimension reduction are performed on the middle-level feature statistics to extract core features; and time series pattern recognition is applied to the top-level environmental change trend to capture the dynamic characteristics of the environment. First, a scene-behavior mapping model is constructed, and the improved Kalman filter algorithm is used to update the environmental state estimation in real time for the bottom-level sensor data, and the particle filter method is used to predict the evolution trend of environmental parameters using the middle-level statistical features and top-level trend data. The system designs an adaptive weight function based on the confidence of the prediction tag, giving a larger weight (0.6-0.8) to the prediction tag with a confidence higher than 0.8, and a smaller weight (0.2-0.4) to the prediction tag with a confidence lower than 0.5. In the fuzzy reasoning process, the system uses a multi-level fuzzy rule base, including environmental threshold rules (such as temperature suitable range, noise tolerance threshold), combined effect rules (multi-factor interaction) and time evolution rules (the lag effect of environmental changes on behavior). For each environmental factor, the system calculates its direct and indirect influence coefficients on the predicted behavior and builds a complete influence propagation network. The scene adaptability assessment uses a deep learning model based on the attention mechanism. By analyzing the scene-behavior correspondence pattern in historical data, learning the mapping relationship between environmental characteristics and behavior prediction, it can achieve rapid adaptation to new scenes.

[0062] Using the scenario-behavior mapping results constructed in the second stage, the system begins to construct the fusion sequence. The system uses the environmental state estimation and evolution trend prediction results in the scenario-behavior mapping model as the basic data for sequence construction, and determines the fusion weights of different features according to the output of the adaptive weight function: features corresponding to high-confidence prediction markers (weights 0.6 - 0.8) obtain higher fusion priorities. At the same time, the system applies the rule set in the multi-level fuzzy rule base to the sequence construction process, uses environmental threshold rules to screen effective features, processes feature interactions through combination effect rules, and controls the time evolution pattern of the sequence based on temporal evolution rules. In addition, the system integrates the direct and indirect influence coefficients in the influence propagation network, establishes a hierarchical fusion relationship between features, and preferentially fuses features with high influence coefficients. The results of scenario adaptability assessment are used to guide the dynamic adjustment of the fusion strategy, and the most suitable fusion parameters are selected for different scenario features. On this basis, the system designs a hierarchical feature fusion framework to integrate environmental features and predicted behaviors at different time scales and spatial dimensions. The analysis in the time dimension uses multi-scale wavelet transform to decompose the environmental feature sequence into different frequency components, and extracts fast-changing features (second level), medium-term evolution features (minute level), and long-term trend features (hour level) respectively. In the spatial dimension, a spatial interpolation algorithm is used to establish a continuous distribution model of environmental parameters and analyze the spatial propagation characteristics of environmental factors. During the feature fusion process, the system uses an improved LSTM network to process temporal features and controls the fusion ratio of features at different time scales through a gating mechanism. At the same time, a graph neural network is introduced to process spatial correlation features and establish a spatial dependence relationship between environmental factors. The system also develops a feature importance assessment module to analyze the contribution of different features to behavior prediction through the XGBoost algorithm and dynamically adjust the feature fusion weights.

[0063] Based on the constructed fusion sequence, the system generates the final prediction result. A multi-dimensional prediction evaluation matrix is established, and the matrix elements include three dimensions: feature contribution degree, time correlation, and spatial correlation degree. A hierarchical progressive prediction strategy is adopted: short-term prediction (5 - 15 minutes) mainly relies on physical environment features and instant behavior data, and uses an ensemble learning method (a combination of random forest and gradient boosting tree) to generate high-precision predictions; medium-term prediction (15 - 60 minutes) comprehensively considers environmental trends and behavior patterns, and uses a recurrent neural network to predict the evolution of the behavior sequence; long-term prediction (1 - 4 hours) combines historical data patterns and uses a Transformer model with an attention mechanism to capture long-term dependencies. For different types of prediction results, the system uses different evaluation metrics: for categorical predictions, the confusion matrix is used to evaluate accuracy; for continuous predictions, the mean squared error and mean absolute error are used to evaluate precision; for probabilistic predictions, the log-likelihood loss is used to evaluate reliability. Through this multi-level prediction mechanism, the system can generate accurate and practical behavior prediction results, providing strong support for decision-making in intelligent environments. In an intelligent classroom environment, the system can combine classroom physical parameter states (temperature, humidity, light, noise), student group characteristics (location distribution, activity intensity), and learning task attributes (difficulty level, completion progress) to generate accurate predictions at multiple time scales and equip each prediction result with a reliability evaluation metric.

[0064] The system receives the short-term, medium-term, and long-term prediction results generated in the above steps, which contain behavior prediction probabilities and environmental impact weights. Based on these hierarchical prediction results, the system begins to perform feature decomposition to mark the prediction intervals. A multi-scale analysis method is used to decompose the prediction results in the time dimension into a fast response interval (5 - 15 minutes), a medium-term change interval (15 - 60 minutes), and a long-term trend interval (1 - 4 hours). For each time interval, the system extracts the main behavior patterns and environmental impact features respectively. During the decomposition process, the system designs an adaptive threshold mechanism that can dynamically adjust the decomposition parameters according to the prediction reliability at different time scales. For example, for the fast response interval, the system focuses on the instantaneous changes of behavior features; for the medium-term change interval, it focuses on analyzing the gradual change features of behavior patterns; for the long-term trend interval, it mainly extracts the stable features of behavior evolution. Through this multi-level decomposition strategy, the system finally obtains a sequence of prediction intervals containing time marks and reliability metrics.

[0065] Based on the obtained prediction interval sequence and its change patterns, the system begins to construct a rule system. The system uses the time markers in the prediction interval sequence to organize the rules in chronological order to ensure that the rules can be accurately applied to the corresponding time windows. At the same time, high-quality features are screened according to the reliability index, and the prediction intervals with high reliability provide stronger supporting evidence for rule construction. First, interval mapping rules are established to establish the corresponding relationship between the key features (duration, fluctuation range, change trend, etc.) of the prediction interval and the behavior patterns. In the rapid response interval, the rules mainly describe the immediate change features of the behavior. In the medium-term change interval, the rules focus on the evolution law of the behavior pattern. In the long-term trend interval, the rules focus on analyzing the cumulative effect of the behavior. The system adopts a fuzzy inference mechanism to handle rule judgment, which can effectively handle the uncertainty in the prediction process. At the same time, a rule conflict handling mechanism is established. When rules at different time scales conflict, the rule priorities are dynamically adjusted by analyzing the applicable conditions and confidence levels of the rules. Each rule in the rule library is equipped with a complete set of conditional attributes and rule strength indicators, forming an adaptive behavior judgment rule system.

[0066] Using the established multi-level rule system, the system begins to determine the confidence level of the prediction. A multi-dimensional confidence level evaluation framework is designed to construct an evaluation system from three dimensions: time correlation, spatial consistency, and rule matching degree. For time correlation, the prediction reliability is evaluated by analyzing the feature stability within the prediction interval. For spatial consistency, the distribution characteristics of environmental features are examined. For rule matching degree, the confidence level is calculated based on the trigger strength of the rules. The system adopts a comprehensive evaluation strategy to weight and integrate the evaluation results of multiple dimensions to generate the final confidence level indicator. In the intelligent classroom scenario, when the system predicts that students may be in a distracted state, it will comprehensively evaluate multiple factors: check the stability of the prediction within the recent time window (time correlation), analyze whether the distribution of classroom environmental features supports the prediction (spatial consistency), and the matching degree of this type of behavior pattern in the historical rule library (rule matching degree). For example, if it is detected that students keep their heads down continuously, the environmental noise increases, and the historical rules show that this combination is highly correlated with distraction, the system will give a high confidence level evaluation. At the same time, a dynamic learning mechanism is introduced to adjust the weight coefficients of each dimension according to the accuracy of historical predictions. The timeliness factor is also considered in the confidence level calculation process, and the weight of earlier judgment results is appropriately reduced.

[0067] Based on the determined multi-dimensional confidence index, the system begins to generate a decision sequence. A time-series decision framework is constructed to organize the behavior states and confidence indices in the prediction interval into a continuous decision sequence. The generation of the decision sequence adopts a multi-scale sliding window strategy, and the window size is adaptively adjusted according to the time span of the prediction interval. The system designs a sequence smoothing algorithm that can effectively eliminate mutations and noises in the decision results. For the decision at each time point, in addition to giving the prediction of the main behavior type, it also includes the corresponding confidence interval. In the college programming experiment course, the system can generate a complete learning behavior decision sequence: when it recognizes that a student is debugging a complex program, the decision sequence will reflect the changes in the entire problem-solving process. For example, in the early stage, it may show "deep thinking state (confidence interval 0.85 - 0.95)", in the middle stage, "problem troubleshooting state (confidence interval 0.75 - 0.85)" appears, and if the problem remains unresolved, it may turn into "attention fatigue state (confidence interval 0.80 - 0.90)" in the later stage. The system establishes a time-series correction mechanism for the decision results by analyzing the decision consistency of adjacent time windows, and can timely identify and correct possible decision biases. For example, when it detects a sudden change in the decision result, the system will comprehensively analyze the characteristics of the front and back time windows to confirm whether this change is reasonable. This structured decision sequence not only maintains time continuity but also reflects the reliability of the prediction.

[0068] Step S104, perform behavior analysis according to the decision sequence, extract behavior patterns; obtain prediction parameters based on the behavior patterns; construct update rules according to the prediction parameters, generate prediction strategies according to the update rules, and output behavior prediction results according to the prediction strategies.

[0069] Specifically, the system receives the decision sequence generated in the above steps, which includes the types, time windows, and confidence intervals of behavior predictions. Based on this decision information, the system begins to perform behavior analysis to extract behavior patterns.

[0070] In some embodiments, the performing behavior analysis according to the decision sequence and extracting behavior patterns includes: hierarchically decomposing the decision sequence, identifying basic behavior units and their combination rules; obtaining individual behavior characteristics and time-varying pattern characteristics through pattern matching methods; constructing a pattern rule library according to the individual behavior characteristics and time-varying pattern characteristics, extracting behavior pattern conversion rules and stability indicators corresponding to the prediction parameters, and generating the behavior patterns including an individual difference adaptation mechanism.

[0071] First, use the time window division in the judgment sequence to determine the time granularity of behavior analysis; then, based on the distribution of prediction types and confidence intervals, identify high-confidence behavior characteristics. Adopt a multi-level analysis framework to extract behavior characteristics from three levels: micro, meso, and macro. At the micro level, the system focuses on the behavior characteristics within a single time window, such as immediate states like attention level and cognitive load; at the meso level, analyze the behavior transition patterns between adjacent time windows to identify typical behavior evolution paths; at the macro level, extract the periodic and trend characteristics of behaviors in a long time series. In the online learning scenario, the system can extract a complete behavior pattern from a student's learning process in a class: such as identifying a cyclic pattern of "focus - distraction - adjustment" or a progressive pattern of "understanding - practice - improvement". Through this multi-level behavior analysis, the system obtains a set of structured behavior pattern descriptions.

[0072] Based on the extracted behavior patterns, the system begins to mark prediction parameters. An adaptive parameter marking framework is designed to assign corresponding prediction parameter sets to different types of behavior patterns. The marking process considers three key dimensions: time dimension parameters (such as behavior duration, transition time points, cycle length, etc.), state dimension parameters (such as behavior intensity, stability, change rate, etc.), and environmental dimension parameters (such as the influence weight of environmental factors, trigger thresholds, etc.). In the remote programming teaching scenario, the system marks the parameters of the students' programming behaviors: record the continuous coding duration, debugging frequency, and time interval for submitting code in the time dimension; mark the code modification rate, compilation error rate, and problem-solving efficiency in the state dimension; mark the development tool switching frequency, number of times of querying reference materials, etc. in the environmental dimension. When it is observed that the programming efficiency of students significantly improves during a specific period, the system will accordingly adjust the weights of these parameters to strengthen the recognition ability of effective behavior patterns. For each dimension of parameters, the system has established a dynamic update mechanism that can adjust the parameter values according to the actual observation results. In addition, the system also configures an effective range and update step for each parameter to ensure the rationality of parameter adjustment. Through this systematic parameter marking, a complete prediction parameter system is established for each behavior pattern.

[0073] Using the labeled prediction parameter system, the system begins to construct update rules. First, it establishes the mapping relationship between parameters and the prediction effect of behaviors, and analyzes the impact of different parameter combinations on prediction accuracy. According to the analysis results, a set of parameter update rules is designed, including trigger conditions (when to start the update), update directions (how to adjust the parameters), and update amplitudes (the magnitude of the adjustment). The system adopts a hierarchical update strategy, giving more frequent update opportunities to the core parameters that have a significant impact on the prediction effect, while using a lower update frequency for secondary parameters. For example, in an online learning scenario, if it is found that the attention prediction parameters of students perform poorly in the afternoon period, the system will first analyze the deviation pattern between the actual attention state and the prediction result. When the deviation mainly appears in the attention duration, the system preferentially updates the relevant parameters in the time dimension; when the deviation is reflected in the attention switching frequency, it focuses on adjusting the state transition parameters. For environmental impact parameters, the system will perform periodic updates after accumulating enough observation samples. The system also establishes a parameter linkage mechanism. When a significant adjustment occurs to a certain core parameter, the relevant subordinate parameters will also be updated synchronously according to the preset rules to ensure the overall consistency of the parameter system. Through this multi-level update rule system, the system can continuously optimize the accuracy of prediction parameters.

[0074] Based on the constructed update rules, the system begins to generate prediction strategies. It adopts a hierarchical strategy generation framework to integrate the prediction requirements of different time scales and behavior types into a unified strategy system. During the strategy generation process, the system comprehensively considers multiple factors: the historical effects of parameter updates, the characteristics of the current behavior patterns, changes in environmental conditions, etc. At the same time, a strategy evaluation mechanism is established. By tracking the execution effects of strategies, the priorities and execution methods of strategies are dynamically adjusted. In an intelligent classroom environment, when the system is applied to a 180-minute programming course, it can generate refined prediction strategies for different stages: in the knowledge explanation stage (0 - 45 minutes), the strategy focuses on the prediction of the understanding level. By analyzing the code following speed of students and the behavior of marking key knowledge, it predicts the depth of understanding; in the example analysis stage (45 - 90 minutes), the strategy turns to the prediction of problem analysis ability. By monitoring the code debugging process and problem-solving paths of students, it evaluates the analysis ability; in the practical programming stage (90 - 180 minutes), the strategy focuses on predicting the improvement of programming proficiency. By analyzing indicators such as code quality, completion speed, and error handling ability, it predicts the learning effect. The system will also dynamically adjust the parameter thresholds and update frequencies of the prediction strategies according to the programming levels and learning characteristics of different students. For example, for students with relatively weak foundations, the system will reduce the evaluation weight of programming speed and increase the proportion of code correctness inspection; for advanced students, it pays more attention to the prediction of code optimization ability and problem-solving efficiency. This adaptive prediction strategy system can accurately capture the behavioral characteristics of students at different learning stages and provide timely and effective decision-making support for teaching intervention.

[0075] In some embodiments, outputting a behavior prediction result according to the prediction strategy includes: performing multi-dimensional verification on the prediction strategy to obtain verification features; marking behavior attributes based on the verification features, and constructing a prediction sequence by using the behavior attributes; and outputting the behavior prediction result according to the prediction sequence.

[0076] The system receives the prediction strategy generated in the above steps and its hierarchical strategy system. Based on this strategy information, the system starts to perform multi-dimensional verification to obtain verification features. A verification feature framework is constructed, including three dimensions: temporal consistency verification, spatial correlation verification, and logical integrity verification. In temporal verification, the system constructs a time series similarity matrix to compare historical data with predicted data, calculates the short-term fluctuation index and long-term trend consistency, and at the same time decomposes the prediction sequence into periodic, trend, and random components through time series decomposition technology to evaluate reliability. The multi-window analysis technology is introduced in the verification process to evaluate prediction consistency on different time scales, and abnormal prediction points are identified through fluctuation threshold detection. In spatial verification, an environmental feature map is established, the prediction result is projected into a multi-dimensional environmental space to calculate vector similarity, and a spatial association heat map is constructed to identify significant association regions and abnormal points. The system implements a feature space partitioning algorithm to divide the environmental feature space into multiple functional regions, calculates the distribution probability of the prediction result in each region, and forms a spatial distribution feature vector. In logical verification, a behavior rule inference engine is implemented to compare predefined rules with the current prediction, a state transition graph is constructed to verify the rationality of the behavior path, and the consistency index of the transition probability matrix is calculated. At the same time, a rule conflict detection module is developed to identify and quantify the degree of logical violation in the prediction result, and a violation index scoring system is constructed. In the distance education scenario, the system verifies the duration distribution of the attention index, the support degree of the feature combination of the learning environment, the integrity of the previous behavior sequence, and the attention resource allocation rule, so as to obtain multi-dimensional verification features.

[0077] Based on the obtained verification features, the system begins to label behavioral attributes. An adaptive attribute labeling framework is designed to extract three types of key attributes: basic attributes, associated attributes, and evolutionary attributes. In the basic attribute labeling, a mapping graph from features to behaviors is constructed, a duration prediction module is developed to analyze historical distribution data, and a multi-factor probability calculator is implemented to integrate environmental conditions and individual differences. The system introduces an attribute differentiation processing mechanism, sets a differentiated attribute weight system for different types of behaviors, and automatically adjusts the accuracy and granularity of attribute extraction according to the quality of the verification features. In the associated attribute labeling, a pre-behavior graph is constructed to record the trigger chain and intensity, a trigger condition extractor is developed to analyze the environmental variable threshold range, and an influence factor weight matrix is generated to quantify the influence degree of each factor. A factor interaction effect analysis module is implemented to evaluate the combined impact of multi-factor collaboration on behaviors and identify the key factor combination patterns and their trigger thresholds. In the evolutionary attribute labeling, a trend fitting tool is developed to perform polynomial fitting on behavioral indicators, a state transition network is constructed to calculate transition probabilities and times, and a stability assessment module is implemented to quantify the fluctuation range and frequency. An attribute evolution trajectory tracking system is designed to record the change sequence of attribute values over time, identify the characteristic patterns in the stable period, transition period, and mutation period, and construct a complete attribute evolution profile. The system defines priorities and relationships through an attribute dependency graph, calculates the historical data support degree to generate an accurate credibility score. In the programming learning scenario, the system labels the exact numerical range and fluctuation curve of focus, records the operation sequence and error handling strategies during the problem-solving process, and analyzes the skill improvement rate and the programming style evolution trajectory.

[0078] Using the labeled behavioral attributes, the system constructs a prediction chain. Analyze the temporal correlations and conditional dependencies among the attributes, construct a directed influence graph to represent causal relationships, quantify the influence strength between different attributes, and achieve attribute conflict detection to resolve contradictory dependencies. The system develops an attribute propagation model to simulate how the changes in attribute values spread and evolve in the prediction network, evaluate the sensitivity of the overall prediction result to the change of a single attribute, and identify key decision nodes and stability bottlenecks. The prediction chain adopts a three-layer architecture: the bottom layer integrates behavioral attributes to generate immediate feature vectors, constructs a feature mapping relationship to predict the probability distribution of short-term behaviors; the middle layer develops a state sequence generator to predict the state sequence based on the current state and historical patterns, constructs a trigger condition monitor to evaluate the satisfaction level in real time, and calculates the exact time point of the behavioral state transition; the top layer designs a trend simulation engine to separate short-term fluctuations from long-term trends, identify periodic patterns in behavioral changes, and evaluate trend stability and potential change points. The system implements a multi-path prediction generation mechanism, maintains multiple possible prediction paths simultaneously, and adjusts the path probability weights according to real-time feedback to ensure that the prediction system has sufficient adaptability and robustness to environmental changes. The system configures a dynamic weight mechanism for each node, adjusts the prediction path weights according to historical accuracy, environmental matching degree, and time span; establishes a temporal correlation mechanism to construct a time dependence graph, and realizes a time decay function to quantify the change of prediction accuracy over time. In a classroom environment, the prediction chain can predict the recent change curve and focus peak from the current attention state, calculate the possible turning points of cognitive fatigue and trigger conditions, and analyze the learning effect improvement trend and understanding obstacle points throughout the course.

[0079] Based on the constructed prediction chain, the system outputs behavior predictions. It realizes the cross-validation and comparison consistency of multi-layer prediction results, and automatically adjusts the parts with logical contradictions to ensure coherence. The system designs a prediction result fusion engine, which adopts a hierarchical integration strategy to integrate prediction outputs at different levels, and generates a consistent and robust prediction result set through techniques such as weighted voting, Bayesian averaging, and sequence alignment. The bottom layer outputs immediate behavior predictions, including behavior types, occurrence probabilities, and durations; the middle layer outputs state transition predictions, describing the evolution path, trigger conditions, and key time points; the top layer outputs trend predictions, indicating the development direction, change rate, and stability. In the bottom layer prediction, a behavior classification engine is implemented to map complex patterns to standard types, generate confidence intervals of probabilities, and predict the duration of behaviors. The system develops a multi-granularity time window analysis tool to evaluate behavior characteristics at different time scales and capture the complete behavior spectrum from second-level micro-changes to hour-level macro-trends. In the middle layer prediction, an evolution sequence of behavior states over time is generated, the threshold values of trigger conditions for each state transition are defined, and the occurrence time of key transitions is accurately located. A state transition warning system is implemented to send warning signals before the state is about to transition, and provide a quantitative assessment of the transition probability and expected impact. In the top layer prediction, a trend direction indicator is constructed to predict the long-term development direction and inflection points, quantify the speed and acceleration of behavior changes, and evaluate the stability degree and fluctuation range of trends. In intelligent teaching applications, the system outputs "current attention focus > 0.8, continuous probability 85% ± 5%, expected 15 minutes, low fixation point conversion frequency", "it is expected that a state transition will be triggered by cognitive load accumulation 10 minutes later, probability 75% ± 8%, completion time 5 minutes, manifested as a change in eye movement pattern", "learning efficiency improvement rate 0.05 / hour, fluctuation ±0.02, trend maintenance probability 80%, it is expected to show an improvement in programming speed". The system integrates the results of each layer through a multi-level prediction fusion matrix, dynamically adjusts the weight coefficients, and converts the prediction results into a standard data structure for output.

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

[0081] 1. The present invention realizes the accurate extraction of behavior intention information in EEG signals through multi-scale feature analysis and adaptive hierarchical processing. Specifically, by using the multi-channel signal segmentation and time series feature extraction method, the time resolution of the feature sequence is improved; by using the hierarchical processing and component extraction technology, the expression ability of features is enhanced; through the construction of feature levels, the systematic expression of behavior features is realized, effectively overcoming the problem of unstable signal quality collected by smart glasses.

[0082] The present invention proposes a method for constructing a behavior library based on feature templates, and establishes a complete behavior prediction framework by combining multi-dimensional prediction markers and environmental feature analysis. Through the marking of behavior samples and the construction of feature templates, a rich behavior knowledge base is formed; by using the temporal mapping and dependency analysis of behavior paths, the temporal characteristics of prediction are enhanced; by combining the associated matching of environmental features, the reliability and real-time performance of prediction results are improved.

[0083] The present invention designs a systematic prediction verification and strategy generation mechanism, and realizes accurate prediction of user behavior through the construction of multi-dimensional verification features and prediction chains. By using the feature decomposition and rule construction of the prediction interval, the adaptability of prediction is enhanced; by using behavior pattern analysis and prediction parameter marking, the accuracy of prediction is improved; by extracting verification features and constructing prediction chains, the prediction challenges brought by individual differences and time-varying characteristics are solved.

[0084] In order to execute the electroencephalogram-based user behavior prediction method in the smart glasses corresponding to the above method embodiments to achieve the corresponding functions and technical effects. Refer to Figure 2 , Figure 2 The block diagram of a user behavior prediction device 200 provided by an embodiment of the present application is shown. For the sake of convenience of description, only the parts related to this embodiment are shown. The user behavior prediction device 200 provided by the embodiment of the present application includes:

[0085] A data acquisition unit 201, configured to acquire multi-channel electroencephalogram data of the smart glasses to obtain an original signal; perform data segmentation based on the original signal, generate a signal sequence according to the segmentation result corresponding to the data segmentation, obtain a feature sequence corresponding to the signal sequence, and obtain a behavior feature corresponding to the feature sequence;

[0086] A sequence analysis unit 202, configured to perform sequence analysis according to the behavior feature to obtain a behavior sample, construct a behavior library according to the behavior sample; perform temporal mapping on the behavior library to obtain a behavior path; extract a dependency relationship based on the behavior path; construct a sequence rule by using the dependency relationship; generate a prediction marker according to the sequence rule;

[0087] An environment analysis unit 203, configured to perform environment analysis according to the prediction marker to obtain a scene feature; perform associated matching on the scene feature, construct a fusion sequence based on the matching result of the associated matching, generate a prediction result according to the fusion sequence; perform feature decomposition on the prediction result, mark a prediction interval; construct a rule based on the prediction interval; determine a confidence level according to the rule, and generate a determination sequence according to the confidence level;

[0088] A behavior analysis unit 204 is configured to perform behavior analysis according to the determination sequence, extract behavior patterns, obtain prediction parameters based on the behavior patterns, construct an update rule according to the prediction parameters, generate a prediction strategy according to the update rule, and output a behavior prediction result according to the prediction strategy.

[0089] The above-mentioned user behavior prediction device 200 can implement the EEG-based user behavior prediction method in the smart glasses of the above method embodiment. The optional items in the above method embodiment are also applicable to this embodiment and will not be elaborated here. The remaining content of this application embodiment can refer to the content of the above method embodiment and will not be repeated in this embodiment.

[0090] Figure 3 It is a schematic structural diagram of a computer device provided in an embodiment of the present application. As Figure 3 shown, the computer device 3 in this embodiment includes: at least one processor 30 ( Figure 3 only one is shown), a memory 31, and a computer program 32 stored in the memory 31 and executable on the at least one processor 30. When the processor 30 executes the computer program 32, the steps in any of the above method embodiments are implemented.

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

[0092] The so-called processor 30 may be a central processing unit (CPU), and the processor 30 may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

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

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

[0095] An embodiment of the present application provides a computer program product, and when the computer program product runs on a computer device, the computer device is caused to implement the steps in each of the above method embodiments when executed.

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

[0097] When the above-mentioned 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 this application, in essence, or the part that contributes to the prior art or a part of this technical solution can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device to execute all or part of the steps of the methods described in various embodiments of this application. The aforementioned storage medium includes: various media that can store program codes such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs.

[0098] The specific embodiments described above further elaborate on the purpose, technical solution, and beneficial effects of this application. It should be understood that the above description is only specific embodiments of this application and is not used to limit the protection scope of this application. It is particularly pointed out that for those skilled in the art, any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of this application should be included in the protection scope of this application.

Claims

1. A method for predicting user behavior based on EEG in smart glasses, characterized in that: include: Collect multi-channel EEG data from smart glasses and obtain original signals; Performing data segmentation based on the original signal, generating a signal sequence according to the segmentation results corresponding to the data segmentation, obtaining a feature sequence corresponding to the signal sequence, and obtaining a behavior feature corresponding to the feature sequence; Perform sequence analysis according to the behavior characteristics to obtain behavior samples, and build a behavior library according to the behavior samples; perform time sequence mapping on the behavior library to obtain a behavior path; and extract dependency relationships based on the behavior path; constructing a sequence rule using the dependency relationship; generating a prediction tag according to the sequence rule; Performing environmental analysis based on the prediction markers to obtain scene features; Performing association matching on the scene features, constructing a fusion sequence based on the matching results of the association matching, and generating a prediction result according to the fusion sequence; performing feature decomposition on the prediction result, and marking a prediction interval; constructing rules based on the prediction interval; Determining a confidence level according to the rule, and generating a decision sequence according to the confidence level; Behavior analysis is performed according to the determination sequence to extract behavior patterns; prediction parameters are obtained based on the behavior patterns; update rules are constructed according to the prediction parameters, prediction strategies are generated according to the update rules, and behavior prediction results are output according to the prediction strategies.

2. The method according to claim 1, characterized in that The acquiring the characteristic sequence corresponding to the signal sequence comprises: Dividing the signal sequence into windows to extract time series features; Marking behavioral indicators based on the temporal characteristics; The feature sequence is generated according to the marking result corresponding to the behavior indication mark.

3. The method according to claim 1, characterized in that The obtaining of the behavior feature corresponding to the feature sequence includes: Performing layered processing on the feature sequence to obtain basic features; performing component extraction based on the basic features; constructing a feature hierarchy using components obtained by the component extraction; The behavior feature is generated according to the feature hierarchy.

4. The method according to claim 1, characterized in that The step of constructing a behavior library according to the behavior samples comprises: Building a feature template based on the behavior sample; Using the feature template to perform data matching; The behavior library is generated according to the matching result corresponding to the data matching.

5. The method according to claim 1, characterized in that The performing time sequence mapping on the behavior library to obtain the behavior path includes: Constructing a multi-dimensional temporal correlation map based on the behavior library to perform micro-dependency analysis, meso-combination pattern recognition and macro-trend tracking of behavior paths; Establishing a temporal causal network based on the multi-dimensional temporal association graph to calculate the conditional transition probability and temporal correlation coefficient of adjacent behavior patterns; Typical combination patterns and variation forms in the behavior path are identified according to the conditional transfer probability and the time series correlation coefficient, and a behavior path containing direct dependency and long-term evolution law is generated.

6. The method according to claim 1, characterized in that The performing environmental analysis according to the prediction mark to obtain scene features includes: Build an environmental feature library to extract the correlation features between environmental conditions and task requirements; Establish the mapping relationship between environmental parameters and behavioral indicator labels, and analyze the impact weight of scene elements on behavioral paths; The real-time environmental data is matched with the historical scene patterns in the feature template library through the association matching algorithm to generate a scene feature vector containing environmental constraints and behavior triggering mechanisms.

7. The method according to claim 1, characterized in that The performing behavior analysis according to the determination sequence and extracting the behavior pattern comprises: Decomposing the judgment sequence hierarchically to identify basic behavior units and their combination rules; The individual behavior characteristics and time-varying pattern features are obtained through pattern matching methods; A style rule library is constructed according to the individual behavior characteristics and time-varying pattern features, the behavior pattern conversion rules and stability indicators corresponding to the prediction parameters are extracted, and the behavior style containing the individual difference adaptation mechanism is generated.

8. The method according to claim 1, characterized in that: Outputting the behavior prediction result according to the prediction strategy includes: Performing multi-dimensional verification on the prediction strategy to obtain verification features; Marking behavior attributes based on the verification features, and constructing a prediction sequence using the behavior attributes; The behavior prediction result is output according to the prediction sequence.

9. A user behavior prediction device, characterized in that: include: A data acquisition unit, used to collect multi-channel EEG data of smart glasses and obtain original signals; Performing data segmentation based on the original signal, generating a signal sequence according to the segmentation results corresponding to the data segmentation, obtaining a feature sequence corresponding to the signal sequence, and obtaining a behavior feature corresponding to the feature sequence; A sequence analysis unit, configured to perform sequence analysis according to the behavior characteristics, obtain behavior samples, and construct a behavior library according to the behavior samples; perform time sequence mapping on the behavior library to obtain a behavior path; and extract dependency relationships based on the behavior path; constructing a sequence rule using the dependency relationship; generating a prediction tag according to the sequence rule; An environment analysis unit, used to perform environment analysis based on the prediction mark to obtain scene features; Performing association matching on the scene features, constructing a fusion sequence based on the matching results of the association matching, and generating a prediction result based on the fusion sequence; performing feature decomposition on the prediction result, marking a prediction interval; and constructing rules based on the prediction interval; Determining a confidence level according to the rule, and generating a decision sequence according to the confidence level; A behavior analysis unit is used to perform behavior analysis according to the judgment sequence and extract behavior patterns; obtain prediction parameters based on the behavior patterns; construct update rules according to the prediction parameters, generate prediction strategies according to the update rules, and output behavior prediction results according to the prediction strategies.

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

Images