Remote education data processing system

Abstract

The invention relates to a remote education data processing system which comprises the following steps: under a remote teaching task, pre-defining a task intention and a data expectation point; a semantic timestamp and a task binding label are printed on each data fragment; mapping the collected confusion data including silence, eye movement drift and prediction into a unified learning semantic vector; a micro-expression + interactive behavior + time sequence decision path ternary modeling mode is introduced, and a potential cognitive intention corresponding to the feature combination is recognized; teaching context information is fused; constructing a cognitive state mapping model; reasoning a current cognitive state label from multi-modal sensing data; searching intervention track VS effect feedback data in a historical database; generating a predicted intervention behavior sequence by using a sequence modeling algorithm; a dynamic combination suggestion chain including light prompt, content reconstruction, personalized practice and tutoring invitation is adopted; and superposing the cognitive state sequences of all students into a group cognitive trajectory map.

Description

Technical Field

[0001] The present invention relates to a remote education data processing system. Background Art

[0002] While data processing currently used in distance education scenarios supports the basic needs of online teaching to a certain extent, such as video playback, answer statistics, and student activity records, it still has many deficiencies and systemic drawbacks in terms of higher-level intelligent teaching support, deep cognitive analysis, and real-time teaching feedback. This seriously restricts the development of distance education towards precision, personalization, and efficiency. First, from the perspective of data collection, most existing distance education systems adopt a single-device + full-time collection model, usually relying only on a single or static channel such as a camera, microphone, or click record to collect information. This lacks dynamic perception of teaching tasks and coordinated control of devices, resulting in redundant and inefficient collection processes. A large amount of invalid data squeezes storage and bandwidth resources, and also causes students to experience data fatigue and privacy anxiety during long teaching processes. The system is unable to intelligently adjust the collection strategy according to the specific teaching stage, which neither effectively obtains valuable information nor easily misses key cognitive moments. Secondly, at the data structure level, current systems mostly store student behavior data in the form of event records or operation logs, lacking semantic layer structure extraction and cognitive association modeling of the data. In other words, the system can often only record what students are doing, but cannot understand why students do so, nor can it judge the psychological state and cognitive motivation behind the behavior. The behavioral data is isolated and fragmented, which seriously affects the quality of subsequent analysis.

[0003] Furthermore, current systems often rely on static rules and preset thresholds for behavioral analysis and teaching feedback, lacking model adaptability and contextual learning capabilities. This makes them inadequate for complex, dynamic, and personalized learning behavior patterns, prone to misjudgment, over-intervention, or unresponsiveness, resulting in a failure to accurately capture students' true learning state. Furthermore, existing systems almost completely ignore the crucial role of teaching context. For example, the impact of variables such as explanation difficulty, the pacing of knowledge point transitions, and video speech speed on student behavior is generally overlooked. This results in the system's inability to correctly understand the varying meanings of the same behavior in different teaching contexts, leading to biased interpretations of the same behavior. More critically, when it comes to cognitive state analysis, most existing systems rely solely on superficial metrics such as active-inactive or completed-incomplete. They lack the ability to identify students' complex cognitive states, such as superficial memory errors and misconfidence, or deeper engagement. Furthermore, they lack effective mathematical models for mapping data to cognitive labels. This hinders the system's ability to conduct in-depth teaching interventions and personalized recommendations, and prevents teachers from obtaining high-quality cognitive feedback to support their teaching decisions. Furthermore, current systems typically analyze individual behaviors in isolation and lack the ability to collaboratively model group cognition. This makes it impossible to identify systemic teaching issues such as collective comprehension barriers and mismatched teaching rhythms, making it difficult to support quality optimization for large-scale remote learning. Finally, most platforms lack the ability to predict data value trends. When resources are limited or the collection load is heavy, they still use a fixed-frequency collection strategy. This wastes computing power and may miss key behavioral segments. This makes it impossible to achieve true simultaneous collection, prediction, and screening. This not only reduces the system's intelligence but also limits the sustainable use of teaching data. Summary of the Invention

[0004] The purpose of the present invention is to provide a remote education data processing system to solve some of the drawbacks and shortcomings pointed out in the background technology.

[0005] The remote education data processing system is characterized by comprising the following steps:

[0006] S1. Data collection collaboration mechanism using cognitive scenario modeling:

[0007] S1.1. Predefine task intent and data expectations for each distance learning task, including lectures, answering questions, and discussions.

[0008] S1.2. All terminal acquisition devices, including microphones, cameras, and mice / touch screens, are coordinated to align with the current task objectives. A simultaneous acquisition and screening model is adopted to retain data segments that are strongly related to cognitive tasks. Semantic timestamps and task-binding tags are added to each data segment to create a traceable data layer structure.

[0009] S2. Using semantic layer data fusion and interpretation mechanism:

[0010] S2.1. Map the collected data, including silence + eye movement drift → prediction confusion, into a unified learning semantic vector; introduce a ternary modeling approach of micro-expression + interactive behavior + temporal decision path to identify the potential cognitive intention corresponding to the feature combination;

[0011] S2.2. Integrate teaching context information, including the difficulty of teaching content, explanation speed, and knowledge point switching rhythm; build a cognitive state mapping model: infer the current cognitive state label from multimodal perception data;

[0012] S3. Adopting the teaching intervention decision chain generation mechanism:

[0013] S3.1 Based on the student's current cognitive state label, search the historical database for intervention trajectory VS effect feedback data; use the sequence modeling algorithm to generate a predicted intervention behavior sequence;

[0014] S3.2. Adopt a dynamic combined suggestion chain including gentle prompts → content reconstruction → personalized practice → tutoring invitation; choose whether to intervene in any of these nodes, so that the system and teachers can form a co-governance intervention mechanism;

[0015] S4. Multi-user cognitive trajectory collaborative analysis mechanism:

[0016] S4.1. Superimpose all students’ cognitive state sequences into a group cognitive trajectory map; identify collective cognitive breakpoints;

[0017] S4.2. Create a map of hot misunderstanding areas to prompt teachers to repeat, change the lecture, or insert problem sections.

[0018] Furthermore, the method for constructing a data collection collaborative mechanism for cognitive scenario modeling includes:

[0019] Through the teaching platform, the current teaching activities are semantically recognized and the teaching intention identifier T is generated. i , marking the teaching target and carrying the expected point D of a single set of data i , predicting the data type and manifestations of facial micro-movements, gaze drift, and interaction delay in the current teaching context; forming a dynamic mapping model for each type of data expectation based on historical data and teacher settings to guide the scheduling of subsequent collection behaviors; and controlling whether different collection devices participate in collaborative collection through dynamic mapping functions:

[0020] Φ(T i )→{E j ∣μ j (T i )>β}

[0021] in:

[0022] Φ(T i) represents the matching mapping between the current task intention and the acquisition device; E j represents the jth candidate acquisition device including camera, microphone, and touch device; μ j (T i ) is the device j for task T i The semantic fitness function is used to measure whether the device can effectively perceive the data required for the task; β is the acquisition activation threshold. When the device fitness exceeds this value, it is activated by the system to participate in the acquisition.

[0023] Furthermore, the method for constructing a data collection collaborative mechanism for cognitive scenario modeling includes:

[0024] Real-time evaluation of whether each frame or set of data is consistent with the expected behavior of the task; using a cognitive relevance evaluation function, running locally on the terminal, and providing instant feedback on the value index of the data; the relevance function comprehensively considers the degree of match between behavioral characteristics and predefined cognitive templates, while also weighing signal interference and response sensitivity:

[0025]

[0026] in:

[0027] R(t) is the cognitive relevance score of the data at time t; α is the interference tolerance constant preset by the system, which is used to limit the impact of low-quality data on the score; A(x) is the cognitive behavior matching function, and its derivative is represents the rate at which the degree of match between the data behavior and the target cognitive template changes over time; Γ(x) represents the signal confidence weight function currently sampled by the data acquisition device, which is used to dynamically adjust the impact of data quality; the integration interval [t0, t] represents the continuous process from the start of the current task phase to the current one; the cognitive strength of the data is evaluated in real time, and the threshold θ is set to determine whether the current segment should be retained or discarded.

[0028] Furthermore, the method for constructing a data collection collaborative mechanism for cognitive scenario modeling includes:

[0029] Based on optimizing system resources and collection density, a future data value trend prediction mechanism is introduced to assist in determining whether high-frequency collection is needed.

[0030]

[0031] in:

[0032] V(Δt) is the total trend of data value predicted by the system from the current time t to the next Δt period; Λ(t ′ ) is the student at time t ′ The sensitivity coefficient of task response, a high value indicates that the state is volatile and the acquisition value is high; It represents the derivative of the task synchronization function with respect to time, reflecting whether the learning behavior is still closely following the task process; if V(Δt)<∈, that is, the future trend is insufficient, the system reduces the collection rate and retains the summary information.

[0033] Furthermore, the method for constructing a semantic layer data fusion and interpretation mechanism is as follows:

[0034] Based on real-time collection of multimodal behavioral data including eye movement trajectories, mouse movements, and periods of silence, the original perceptual features are encoded and compressed, and mapped into a unified learning semantic vector. The semantic vector represents the cognitive meaning of the behavioral combination. For example, sustained eye movement drift + long periods of silence are interpreted as a state of confusion and lack of active feedback; frequent answer revisions + no facial emotion fluctuations are interpreted as knowledge retrieval driven by recall. Cognitive input units are established. The semantic strength of the combination is quantified, and the following semantic combination mapping function is introduced:

[0035]

[0036] in:

[0037] a is the frequency of eye movement drift per unit time, which is used to assess the fluctuation of visual attention; b is the length of time the student remains silent within a specific task window, which expresses non-responsiveness to information input; c is the reaction delay when the student responds to task prompts or operation instructions; η is the silence behavior amplification coefficient, which is used to adjust the weight of the silence feature in different task types; γ is the normalization constant used to limit the range of variation of the overall score; ξ is the expected normal interaction reaction time in the context task.

[0038] Furthermore, the method for constructing a semantic layer data fusion and interpretation mechanism is as follows:

[0039] A modeling method based on the ternary feature combination of micro-expressions, interactive behaviors, and decision paths was developed. Micro-expression frequency bands, including eyebrow tremors and eyelid retraction, were extracted from the visual expression recognition model, and recognition operation features, including look-back behavior and mouse hovering, were extracted from the behavioral trajectory. The logic of the student's path evolution during the teaching task was simultaneously analyzed, and the three were used as a unified structure to construct a feature tensor. Time weight analysis was performed using the following combined response function:

[0040]

[0041] in:

[0042] Θ(t) is the semantic activity integral of the behavior combination at time t; τ1 and τ2 are the start and end time points of the current task phase, defining the integral time window; F1(i) is the intensity curve of the i-th micro-expression feature changing over time; is the speed of micro-expression change, indicating the frequency of emotional fluctuations; F2(j) is the behavioral trajectory function of the jth interactive behavior; is the behavioral acceleration of the behavior; ω k (t) is the importance weight of the decision path of the k-th student at time t, which is determined based on behavioral history learning.

[0043] Furthermore, the method for constructing a semantic layer data fusion and interpretation mechanism is as follows:

[0044] The generated semantic vector L v , ternary model output tensor and teaching context variables C including explanation speed, content difficulty, and knowledge point conversion frequency n The fusion forms a structured input sequence, which is then fed into the cognitive state mapping engine to generate the current student's cognitive state label through a nonlinear reasoning function. Label generation is based on the following state reasoning function:

[0045]

[0046] in:

[0047] Υ(S m ) is the student’s state label set S at the current moment m The prediction confidence value in S m It includes deep understanding, cognitive rupture, surface memory, passive acceptance of false confidence state labels; N is the dimension of cognitive state space; φ n (·) is the feature mapping function of the nth state, and its input is the learning semantic vector L v , behavioral feature tensor and context factor C n ;λ n is the sensitivity weight of the state label, which controls the response to the input feature; κ is the normalization coefficient, which is used to adjust the range of the predicted value.

[0048] The distance education data processing system achieves a deep understanding of and dynamic response to students' learning behaviors during distance education by building a closed-loop mechanism of cognitive scenario-driven data acquisition, multimodal semantic fusion, intelligent state reasoning, and adaptive teaching feedback. This system has the following significant benefits:

[0049] First, this system overcomes the limitations of traditional distance education platforms, which primarily rely on behavioral recording and lack cognitive understanding. By driving data collection through cognitive task modeling, it retains only key behavioral data closely related to teaching objectives, significantly improving data processing efficiency and collection quality while reducing bandwidth and storage burdens. Second, the system constructs a semantic-layer data fusion model based on a three-dimensional modeling of micro-expressions, interactive behaviors, and learning paths, incorporating contextual information. This allows the system to not only understand what students do but also why they do it, thus achieving a leap from perceptual behavior to reasoning. Third, the system utilizes a series of original mathematical modeling functions (such as cognitive relevance functions, semantic mapping functions, and state inference functions) to quantitatively identify and label students' current cognitive states. This allows it to accurately distinguish between cognitive discontinuities, superficial memory errors, and key learning states, providing teachers and the system with real-time, actionable information for teaching decisions. Furthermore, the system supports group cognitive trajectory analysis and prediction of future behavioral value trends, shifting distance education from individual reaction processing to systematic teaching optimization, achieving a smart teaching model that integrates human and machine co-governance. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a flow chart of the distance education data processing system of the present invention.

[0051] Figure 2 Flowchart of the method for constructing a data acquisition collaborative mechanism for cognitive scenario modeling in the present invention.

[0052] Figure 3 This is a flow chart of the method for constructing semantic layer data fusion and interpretation mechanism adopted in the present invention. DETAILED DESCRIPTION

[0053] The following is a detailed description of the specific embodiments of the present invention with reference to the accompanying drawings.

[0054] The S1 data collection collaborative mechanism, based on cognitive scenario modeling, proposed in the distance education data processing system is a dynamic data collection and screening method based on cognitive task orientation. Its core goal is to overcome the bandwidth waste, excessive data noise, and delayed teaching feedback caused by the indiscriminate, large-scale, and low-structured data collection of traditional distance education, and to achieve a data processing mindset guided by cognitive value. This mechanism is developed through three key technical actions, of which S1.1 is the semantic driving starting point of the entire mechanism. As students engage in different distance learning task contexts—including lectures, answering questions, and discussing typical learning tasks—the system predefines task intent (i.e., the cognitive goals students should achieve during that phase, such as acquiring knowledge, expressing opinions, and evaluating understanding) and corresponding data expectations (focusing on facial expressions, gaze tracking, and periods of silence during the lecture phase; mouse behavior, repeated answers, and duration of answers during the answering phase; and voice fluctuations, speaking frequency, and inflection during the discussion phase). By predefining this cognitive model, the system possesses target perception capabilities during the data collection phase. Then, in the S1.2 stage, various types of terminal acquisition devices (including but not limited to microphones, cameras, mice or touch devices) are uniformly scheduled by the system after the task is started, and the perception target is aligned according to the current task intent. That is, not all devices are started or all data are recorded. Instead, the device priority is arranged and the acquisition parameters are set according to the weight and timeliness of the behavioral characteristics required for the current task, thereby building a multi-device collaborative acquisition mode driven by cognitive goals. At the same time, during the acquisition process, the system will introduce a lightweight edge computing module to perform a processing mechanism for collecting and screening real-time data locally on the terminal. The system determines whether the data segment has cognitive relevance based on the real-time characteristics: if the behavioral characteristics are found to be consistent with the task intent matching path (such as when answering questions, there is a high frequency of answer change behavior and nervous expression), the segment is determined to be a high-value segment and is retained; on the contrary, if the data appears to be static for a long time, without interaction, and disconnected from the task, it is directly discarded or only stored in the form of a low-resolution summary to reduce the pressure of redundant data storage. Each retained data fragment is semantically timestamped, recording the semantic relationship between the behavior and the task phase and teaching node at the time of its occurrence. Task-binding tags are also added to indicate the task intent and cognitive expectations under which the data was generated. This creates a traceable data layer structure that not only facilitates subsequent cognitive analysis and teaching feedback generation, but also provides structural support for the system's closed-loop mechanism of cognitive behavior, data, and feedback. The essential innovation of this mechanism lies in moving the traditional post-processing of teaching data—what to collect, who to collect, and what to retain—to the data entry point. Using cognitive intent as the driving signal, this reconstructs the organizational logic and semantic value system of data in distance education.

[0055] Entering the S2 stage, the semantic layer data fusion and interpretation mechanism is an intelligent analysis process based on the deep semantic understanding of multimodal behavioral data. It aims to break through the limitations of traditional distance learning that can only stay at the surface behavior recognition (such as the number of clicks, answering time, and expression changes), and realize the reasoning and labeling expression of the learner's potential cognitive state. Its core idea is to transform the original perception data into an interpretable and structured cognitive semantic expression system. In step S2.1, the system first performs semantic mapping on the collected key behavioral data. Typical behavioral patterns such as sustained silence + eye drift are identified as a state of cognitive confusion. Such complex behaviors will no longer exist as isolated events, but will be compressed into a unified learning semantic vector through the encoder structure. This vector represents the cognitive meaning behind the behavior rather than the surface action itself. To improve recognition accuracy, the system further introduces a ternary modeling mechanism of micro-expressions + interactive behaviors + temporal decision paths: micro-expressions are used to capture students' potential emotional feedback signals at different task nodes. Raised eyebrows suggest thinking and lip biting represents anxiety. Interactive behavior records such as click-to-undo, content review, and sliding rate action features are used to reflect their knowledge processing methods. The temporal decision path presents the evolution of students' behavioral decisions in a task. The path from directly answering the question → revising the answer → repeatedly checking the answer corresponds to a recall-type cognitive strategy or a low-confidence state. The three types of features are fused to form a combined map, which is used to train the model to identify the relationship between feature patterns and cognitive intentions. Entering the S2.2 stage, to avoid over-interpretation or misjudgment of behavior, the system also introduces teaching context variables as key calibration criteria. These include the theoretical complexity of the current explanation content, the fluctuation range of the teacher's speaking speed and rhythm, the frequency of switching between knowledge points, and the duration of dwelling. This information is combined with student behavior analysis, allowing the system to focus not only on what the student did, but also on what teaching context they did, thereby significantly improving judgment accuracy. Ultimately, the system uses the learning semantic vector + behavioral feature map + teaching context as input to construct a cognitive state mapping model, inferring the current student's cognitive label. This label includes but is not limited to deep understanding, surface absorption, cognitive rupture, passive acceptance, error, self-confidence, and clear semantics and clear state classification results. This is used to support subsequent personalized teaching recommendations, real-time intervention, or learning path adjustment. This gives the distance education system insight similar to that of human teachers, enabling it to not only record student behavior but also understand their cognitive motivations and psychological states, thereby truly realizing data-driven intelligent teaching.

[0056] S3 adopts a teaching intervention decision chain generation mechanism, which is an action design module for cognitive feedback. Its essence is to realize the intelligent teaching control process from understanding students to making appropriate responses. In dynamic teaching scenarios, it not only identifies and labels students' status, but also further promotes the system to drive a set of forward-looking and adaptive intervention behavior paths based on labels, realizing intelligent collaboration between the system and teachers. In the S3.1 stage, the system first matches the historical records similar to the student's cognitive state labels (such as cognitive discontinuity, shallow absorption, and false confidence) output by the S2 link in the existing teaching behavior history database, and focuses on finding the associated intervention trajectories and effect feedback data. In other words, it not only retrieves the intervention methods (prompts, re-teaching, and practice push) that teachers or systems have used in the past, but also evaluates the actual learning effects of these interventions under specific conditions (such as improved mastery of knowledge points, increased accuracy of answering questions, and improved learning continuity). This part of the data constitutes the experience support layer; based on these structured historical samples, the system trains a predictive model of the cognitive state → intervention behavior sequence through a sequence modeling algorithm (a sequence generation model based on the attention mechanism), thereby not only outputting a single-point intervention, but also generating a chain of teaching intervention recommendations that changes according to time, stage, and conditions. In the S3.2 stage, the suggestion chain manifests as a dynamically combinable set of behavioral units, including subtle prompts (such as pop-up reminders and attention-stimulating actions), content reconfiguration (adjusting the explanation order and switching to more vivid video materials), personalized practice (delivering question types precisely corresponding to the student's current confusion), and tutoring invitations (informing the teacher whether to intervene in one-on-one tutoring). These four intervention strategies are hierarchical and flexible. The system generates the most appropriate intervention path recommendations based on the severity of the student's current cognitive state and behavioral fluctuations, forming a response curve from weak intervention to strong guidance. The most significant feature of this mechanism is its configurability: the system provides a complete suggestion chain but does not replace the teacher's decision-making. Instead, it opens the chain nodes as co-governance interfaces, allowing the teacher to choose whether to intervene in a specific link (such as skipping subtle prompts and directly initiating content reconfiguration). This implements a flexible intervention logic of human-computer co-governance, ensuring the immediacy of intelligent response while preserving the humanistic judgment space for educational behavior. This enables the distance education system to be both data-driven and teacher-led, effectively implementing the concept of personalized and timely teaching feedback in large-scale teaching environments.

[0057] Building on this foundation, the S4 multi-user cognitive trajectory collaborative analysis mechanism is an intelligent analysis method for classes or learning groups. Its core is to further aggregate the results of individual student cognitive state analysis to generate group-level cognitive evolution trends, thereby providing teachers with a systematic basis for instructional adjustments and improving the accuracy and responsiveness of the entire teaching activity. In the S4.1 stage, based on the cognitive state labels output by each student at different teaching time points in the previous module, the system uniformly formats and synchronizes these label sequences along the time axis. Then, using statistical models or time series superposition algorithms, it integrates all students' cognitive change sequences into a structured group cognitive trajectory map. This map represents teaching progress, knowledge point sequences, or video explanation segments on the horizontal axis, and the number of students or state frequency on the vertical axis. In this map, if a large number of students display cognitive discontinuity labels after a certain knowledge point is explained, or if attention labels decrease significantly during a certain explanation segment, this can be identified by the system as a collective cognitive discontinuity. This indicates a structural deviation or mismatch between the current teaching content and the students' general understanding, and serves as a key warning signal for teaching quality monitoring. Next, in the S4.2 stage, the system generates a visual hot spot misunderstanding area map based on the cluster analysis and frequency weight judgment of the above trajectory map. The map uses color, density, and heat range to mark which content segments or knowledge points have group understanding barriers or concentrated cognitive loads, thereby helping teachers quickly locate high-risk teaching nodes; at the same time, the system can push teaching suggestions based on these area maps, suggesting whether teachers should arrange re-teaching at this node, use different expressions to change the lecture, or temporarily insert a guiding question link to increase students' re-engagement and deepening of understanding of the content. This mechanism not only avoids blind advancement in teaching, but also realizes the data migration capability from individual cognitive perception to collective teaching decision-making. It is an important bridge from personality analysis to system optimization in distance education systems.

[0058] Example 1:

[0059] A high school remote physics course platform is currently teaching an online chapter on the law of reflection of light. The current task in the classroom is for students to watch the teaching video, reflect on the content, and then answer multiple-choice questions. After the system is started, the semantic recognition module identifies the current teaching intention as concept input + preliminary understanding verification based on the teacher's settings and task progress. This intention is identified by the system as task intention T. i =T 反射理解 The system associates this task intention with the cognitive model set in the background and generates the corresponding data expectation point set D i= {d1, d2, d3}, where d1 represents facial micro-movements (used to identify students' emotional feedback), d2 represents eye movement drift frequency (used to determine the degree of attention), and d3 represents interaction latency (such as the response time after clicking on an answer). Using historical model training data and manual annotation, the system sets reasonable fluctuation ranges and behavioral characteristic standards for these three data dimensions.

[0060] Entering the acquisition phase, the system uses the preset dynamic mapping function Φ(T i )→{E j ∣μ j (T i )>β} schedules and screens the devices. Candidate acquisition devices include E1: student front camera (for facial micro-movement and eye movement tracking), E2: microphone (for voice detection and silence recognition), and E3: touch screen recording module (for answering click trajectory and delay recording). Each device is evaluated for suitability based on the task intent by the system, that is, its semantic fitness function μ is calculated. j (T i ), specifically defined as follows:

[0061] For the front camera E1, the fitness function consists of three parts: facial micro-motion perception capability (set to 0.85), eye tracking sensitivity (0.92), and data delay control coefficient (0.96). Using the weighted average method and combining the task intention weight vector w = [0.4, 0.5, 0.1], we calculated:

[0062] μ1(T i )=0.4×0.85+0.5×0.92+0.1×0.96=0.902

[0063] For microphone E2, since the current task is not based on language expression, the system assigns a lower weight, and its semantic adaptability is:

[0064] μ2(T i )=0.2×0.7+0.3×0.6+0.5×0.4=0.52

[0065] For touch device E3, since it directly collects interaction data and the task requires completing multiple-choice questions, the weight distribution is 0.88 for operation accuracy, 0.9 for click response sensitivity, and 0.85 for historical task completion rate:

[0066] μ3(T i )=0.3×0.88+0.4×0.9+0.3×0.85=0.877

[0067] This system sets the data collection activation threshold β in the range of [0.75, 0.9]. Given the dynamic changes in real-time stress levels during the teaching phase, and the current task requiring high cognitive relevance, which involves answering questions after explaining key concepts, the system currently sets β to 0.85. This indicates that devices E1 and E3 both have a compatibility greater than the threshold, at 0.902 and 0.877, respectively, and are therefore activated by the system to participate in data collection. However, microphone E2 has a compatibility of only 0.52, which falls below the threshold and is therefore not activated.

[0068] As students enter the state of answering questions, the system starts to run the collection and screening module. A student enters the answering page at the 12th minute. When choosing the second question, the front camera captures a brief upward eyebrow raise on his face (recognized by the system as a short-term understanding block signal), and the eye movement frequency reaches 15 times within 5 seconds, exceeding the upper limit of the set steady-state interval (10 times / 5 seconds). In addition, the click time for answering this question is 8.2 seconds, while the average click time for the question is 4.3 seconds. The system immediately determines that the behavior is unstable. This fragment of data triggers the filter to pass, and its data fragment will be stamped with a semantic timestamp (task: multiple-choice question 2, stage: preliminary understanding) and bound to the task label T 反射理解 , as an input source for subsequent cognitive state judgments.

[0069] The system acquisition device has activated the front camera E1 and touch device E3 according to the semantic fitness function and entered the real-time collaborative data acquisition mode. During the answering process, the system not only completes the raw data collection and labeling, but also needs to judge in real time whether the data is closely related to the task intention and whether it should be retained for subsequent cognitive reasoning. Therefore, the cognitive relevance evaluation function module is enabled. When entering the second question, the acquisition system begins to continuously evaluate the quality of the behavioral data in the question from t0=720 seconds (i.e., the 12th minute). Its eye movement behavior is compared with the steady-state gaze model preset by the teacher. At the same time, the camera signal quality, facial key point clarity, and click response path become weighted inputs to form the data behavior signal flow for the current period. At this stage, the system uses the following cognitive relevance evaluation function for judgment:

[0070]

[0071] The system preset interference tolerance constant α ranges from [0.8, 1.2]. The current task is the question-answering stage, which requires high behavioral stability. The system sets α to 1.0. A(x) is the matching function between behavior and cognitive template. The template for this question is stable gaze + rapid judgment, while the student's performance is high-frequency eye movement + click delay. The system's instant calculation of its matching derivative function is: from the 720th second to the 728th second, the matching degree decreases, and the rate of change is This indicates that the deviation from the template is increasing. During this period, the device signal confidence weight Γ(x) fluctuates steadily between [0.9, 0.95], indicating that the camera image is clear, the system latency is low, and the signal quality is high. The integral calculation value in the interval [720, 728] seconds is approximately:

[0072]

[0073] Substituting into the function we get:

[0074]

[0075] The system's cognitive relevance score threshold θ is set in the range of [1.5, 2.2] and is dynamically adjusted according to the importance of the task stage. Currently, at the high cognitive judgment node, θ is set to 2.0. Since R(728)=2.83>θ, the system determines that although the student's behavior during this period exhibits a deviation from the template, the data exhibits obvious cognitive disturbance characteristics and has strong analytical value. Therefore, this data segment will be retained and recorded in the high cognitive fluctuation pool for subsequent model analysis. In addition, based on this judgment result, the system will also issue real-time intervention suggestions, including whether to re-examine the knowledge points involved in the question or optimize the content.

[0076] Considering the behavior of another student in the second question answering stage during the same period, the system records that his behavior from 720 seconds to 730 seconds is almost unchanged, his eye movements are very stable, the click time is standard, and the matching degree derivative is The signal confidence Γ(x) = 0.92, and the integral value is:

[0077]

[0078] After substitution, we get:

[0079]

[0080] Because the value R(730)<θ=2.0, the system believes that although the behavior segment conforms to the template, it lacks cognitive fluctuation information and does not constitute key data. Therefore, the data will be processed in summary form and only statistical features such as answering speed and click path will be stored. It will not enter the deep analysis module.

[0081] Entering the next phase of the core process: Dynamically optimizing system resources and collection density based on a mechanism that predicts future data value trends. This feature is introduced to ensure the system can continuously capture high-value cognitive data while avoiding ineffective high-frequency collection when cognitive states are stable or learning volatility is insufficient. This reduces the transmission load between terminals and servers, extends device lifespan, and improves overall processing efficiency.

[0082] At this point, having just completed the second multiple-choice question on the physics concept that the angle of reflection of light is equal to the angle of incidence, the system will evaluate whether it is worthwhile to continue collecting data at the current time window Δt = 30 seconds after the current time point t = 730 seconds. To this end, the system calculates the future data value trend prediction function:

[0083]

[0084] When the student enters the third question answering stage, the teaching task synchronization model determines that his behavior is in the task continuation period, that is, he has not entered a new knowledge point, but needs to complete the knowledge reflection and preliminary transfer stage. The system calls the learning state history feedback and compares it with the current behavior, and sets the task response sensitivity coefficient function Λ(t ′ ) fluctuates between the interval [0.2, 1.0]. Due to recent cognitive distress (manifested by high-frequency eye movements and delayed responses in the previous question), it is currently in a stable recovery phase. The system fits its state sensitivity into a gradually converging function model:

[0085] Λ(t ′ )=0.9·e -0.05(t′-730)

[0086] This function indicates that starting from the current 730 seconds, the sensitivity of the cognitive state to external task changes will gradually decrease, which means that the behavior will tend to be stable. At the same time, the system monitors the changes in its synchronization with the task in real time. The task synchronization function S(t′) records whether the student's operation rhythm is consistent with the system task guidance. If the student follows the rhythm of the question closely and does not stagnate for a long time or drift in operation, then Close to zero, reflecting a steady state of behavior; during this period, the click time for each question is relatively balanced (5.1 seconds, 4.8 seconds, 5.0 seconds), the content switches smoothly, and the system fits its synchronization derivative to a constant value:

[0087]

[0088] Substituting this into the trend prediction formula, we can calculate the following for the next 30 seconds (i.e. from 730 seconds to 760 seconds):

[0089]

[0090] The result of the integral calculation is:

[0091]

[0092] but:

[0093] V(30)=0.009·15.538≈0.1398

[0094] The system presets the lower limit threshold ∈ in the range of [0.3, 0.6], which is dynamically adjusted according to the complexity of the task and the resource load. The current task is in the steady-state stage of answering questions, and the requirements for collection accuracy are reduced. Therefore, the system sets the threshold ∈=0.4. Because the predicted value V(30)=0.1398<∈, the system intelligently judges that the behavior will remain stable in the next 30 seconds, and the collection value is low. Therefore, the camera frame rate is reduced from 30FPS to 10FPS. At the same time, the micro-expression details are no longer continuously tracked, and only summary information is retained: the start and end time of the question answering, the click location path, and whether the key elements of the answer are modified, and the summary collection mode is entered.

[0095] At this time, another student entered the image question to answer. The question was about the new knowledge point of mirror and diffuse reflection. The task synchronization fluctuated violently (start answering the question → freeze → switch to the explanation playback → return to the question → answer again). The system ′ ) The fitting result is a high-fluctuation curve with a derivative peak of 0.065. At the same time, its state sensitivity coefficient is a linear rising model Λ(t ′ )=0.6+0.01(t ′ -730), is actively trying to understand and adjust its strategy. At this time, the system calculates V(30)≈1.02>∈, which is significantly higher than the set threshold. The system determines that the behavior is in a high cognitive fluctuation area and continues to maintain a high-frequency multimodal acquisition state, retaining the full amount of facial key frames, micro-expression sequences, and click path data for subsequent reasoning of its cognitive transformation process.

[0096] Example 2:

[0097] Based on Example 1, the system has entered the semantic layer data fusion and interpretation mechanism stage. The core of this stage is to structurally fuse the multimodal behavioral data (including eye movement trajectory, mouse operation, and silent time) collected in the previous link, further encode it into a unified learning semantic vector, and quantify the cognitive strength of the semantic vector through a mathematical function model, thereby providing an accurate and computable input signal for subsequent cognitive state reasoning and feedback generation. In the process of answering the third question on the law of reflection of light, the system acquisition device recorded the behavioral characteristics between the 745th and 755th seconds, including the eye movement drift frequency per unit time of a = 6.3 times / second, the continuous silent time of b = 8.1 seconds, and the mouse click response delay of c = 3.2 seconds. Combined with the teaching task context analysis, this stage is marked as a knowledge concept recognition + comprehension task. The system queries the task context knowledge base and obtains the expected reaction time ξ = 2.5 seconds corresponding to this type of task. The normalization constant γ is set to 1.5 (the value range is [1.2, 2.0], which is used to compress the interference of abnormal fluctuations on the model output), and the silent behavior amplification coefficient η is set to 0.75 (the value range is [0.5, 1.0], the task is an input-type understanding stage, and silent behavior should be given due attention).

[0098] The system then calculates the semantic strength of the current behavior combination through the following semantic combination mapping function:

[0099]

[0100] Substitute the data into:

[0101] a 2 =(6.3) 2 =39.69;

[0102] ηb 3 =0.75×(8.1) 3 =0.75×531.44≈398.58;

[0103] The molecular part is

[0104] The denominator is γ + |c - ξ| = 1.5 + |3.2 - 2.5| = 1.5 + 0.7 = 2.2

[0105] The final semantic strength score is:

[0106]

[0107] This score was marked as a high-intensity cognitive signal by the system. According to the semantic intensity threshold rules set by the system (low: <1.8, normal: 1.8–2.8, high: >2.8), this indicates that although this behavior combination appears to be only frequent eye movements and silence, it has obvious cognitive fluctuations after fusion calculation. It is speculated that the individual is undergoing internal conceptual reflection. The system writes its learned semantic vector into the cognitive input unit and attaches a semantic label indicating potential confusion or conceptual ambiguity.

[0108] In contrast, the system also processed another student's behavior on the same question. The behavior data showed a = 3.2, b = 2.5, c = 2.6. The same calculation was performed by substituting them into the function:

[0109] a 2 =10.24;

[0110] ηb 3 =0.75×15.625=11.72;

[0111] The molecule is

[0112] The denominator is 1.5 + |2.6 - 2.5| = 1.5 + 0.1 = 1.6

[0113] but:

[0114]

[0115] Based on the score, the system determines that this behavior belongs to the low semantic intensity range, indicating that its operation is smooth and the state is normal, without significant cognitive deviation or learning challenges. Therefore, this behavior is only stored as an ordinary record in the low-priority behavior trajectory library and does not enter the deep semantic reasoning channel.

[0116] We then proceed to the core submodule of the semantic layer data fusion and interpretation mechanism, namely the three-element feature combination modeling stage of micro-expression + interactive behavior + decision path. The system has evaluated the confused behavior generated during the question-answering process through the semantic combination mapping function, and the subsequent analysis requires the system to have a deeper understanding of the underlying intentions and strategies behind its cognition. Therefore, the system activates the mechanism of this stage, structurally integrates the student's behavioral characteristics, and constructs a dynamic semantic feature tensor for cognitive state reasoning. The current teaching task is in the stage of explaining new knowledge points + answering image-type questions. The time interval is defined as τ1 = 760 seconds to τ2 = 790 seconds. The behavior of this stage is fully recorded by the system, and the following features are extracted: First, the visual expression recognition module detects two relatively obvious eyebrow tremors and one eyelid shrinkage in the interval from 764 seconds to 770 seconds, indicating emotional fluctuations or high cognitive load; the system uses this as the i = 1 micro-expression channel and constructs the feature intensity function F1 (1). The system measures the derivative of its intensity change curve over time based on the expression model. The average value in this time window is 0.06, and the unit is expression intensity change / second, which means that the facial expression is frequent and the fluctuation rate is high during this period. Second, at the level of interactive behavior trajectory, the system records that the mouse is repeatedly hovered over the image description area and repeatedly reviewed. Its behavior path function F2(1) shows an action pattern of first rising and then quickly switching in this stage. Its second-order derivative function Acceleration modeling was performed, with an average value of 0.042, expressed as the rate of change of action / second², indicating that the operation rhythm exhibited a typical reflective behavior of hesitation-switching-returning. Third, the system evaluated the position of the current path in the entire answering behavior tree, confirming that its behavior was highly consistent with the strategic understanding path (i.e., first jumping out of the explanation to review, then returning to the question to answer), and accordingly assigned it an importance weight ω for the decision path. l (t), which is learned based on historical data and has an average weight of 0.85 in the current time window (the value range is [0.4, 1.0], indicating that this type of path has a strong ability to explain the target). The system then substitutes the following combined response function:

[0117]

[0118] After substituting the known values:

[0119]

[0120] ω k (t) = 0.85;

[0121] The overall constant function can be simplified to

[0122] The integration interval is 30 seconds:

[0123] Θ(t)=0.0867×30=2.601

[0124] According to the system's threshold settings, the semantic activity integral value Θ(t) is evaluated as follows: low activity < 1.0, medium activity 1.0-2.0, and high activity > 2.0. In this task phase, the value obtained was 2.601, indicating a high activity state. The system infers that the current behavior exhibits high cognitive volatility and clear structure, indicating that the user is not distracted or perfunctory, but rather undergoing a process of deep understanding or conceptual reconstruction.

[0125] In contrast, the system analyzed the behavior pattern of another student on the same question and recorded no obvious changes in facial expression during this stage. The mouse behavior is stable, without replay or jump. Path matching is a linear question-answering type, and the system sets a path weight ω for it k (t) = 0.5, and the integral is calculated as follows:

[0126]

[0127] If the value is lower than 1.0, the system determines that the behavior is a low-activity stable operation segment and can be not used as deep reasoning input. It is only used for answering success rate statistics and behavior profile recording.

[0128] The system enters the final stage of semantic layer data fusion and interpretation mechanism, which is to fuse all the collected, processed, calculated and encoded multimodal learning data into structured input, send it to the cognitive state mapping engine, generate the current student's cognitive state label, and provide a decision basis for teaching feedback and personalized intervention. In the time period from 760 seconds to 790 seconds, the system has completed the micro-expression, behavior trajectory, and learning path decision-making during the process of answering the light reflection and refraction picture question, and has carried out ternary feature modeling and formed a behavior tensor. It includes the i-th type of micro-expression (eyebrow shaking intensity is 0.83), the j-th type of interactive behavior (multiple hovering and replaying, click action acceleration is 0.042), and the k-th type of path pattern (the confidence weight of the replay→pause→answering path is 0.85). At the same time, the semantic vector L vThe system has been coded and generated in the previous stage, with a semantic intensity value of Ψ = 3.46, which is characterized by a cognitive conflict behavior combination of high-frequency visual response + long-term silence; the teaching context variable set C n These factors include: a speaking rate of 210 words per minute (range 180–240 words per minute), a knowledge point complexity level of 4 (range 1–5), the third transition point in the chapter, and a switching frequency of 0.14 times per minute (range 0.1–0.3). These factors collectively indicate that the current teaching phase places moderate to high cognitive load on students. The system integrates the data from these three modules to form a structured input sequence, which is then fed into the state inference engine. The confidence value for the current cognitive state is calculated using the following nonlinear mapping function:

[0129]

[0130] Inside the system, the state space S m It includes five cognitive state labels: S1 deep understanding, S2 surface memory, S3 passive acceptance, S4 false confidence, and S5 cognitive dissonance. Corresponding to N = 5, the system obtains five feature mapping functions φ1, φ2, ..., φ5 through training. The original mapping values generated for the current input are as follows:

[0131] φ1=0.41, φ2=0.38, φ3=0.23, φ4=0.19, φ5=0.76

[0132] Sensitivity weight λ of the state label n The settings are as follows (value range 0.5–1.5): λ1 = 1.1, λ2 = 0.9, λ3 = 0.8, λ4 = 0.7, λ5 = 1.4, and the system normalization constant κ = 5.0 (set according to the number of states N to ensure that the confidence value is within a reasonable range). The confidence value calculation process is:

[0133]

[0134] Calculating item by item:

[0135] log(1+0.451)=log(1.451)≈0.372;

[0136] log(1+0.342)=log(1.342)≈0.294;

[0137] log(1+0.184)=log(1.184)≈0.169;

[0138] log(1+0.133)=log(1.133)≈0.125;

[0139] log(1+1.064)=log(2.064)≈0.726;

[0140] The cumulative total is 0.372+0.294+0.169+0.125+0.726=1.686

[0141] Then finally:

[0142]

[0143] This confidence value is marked as significantly higher risk in the system's cognitive label threshold table. In particular, because item 5, S5 (cognitive discontinuity), has the highest individual mapping value and is assigned the highest sensitivity weight, λ5 = 1.4, the system ultimately labels the cognitive state at this stage as cognitive discontinuity, meaning the student has exhibited a significant disruption in understanding and requires instructional intervention. This label is immediately fed back to the teacher's terminal, triggering a chain of system intervention suggestions (e.g., gentle reminders, video replays, and subsequent personalized practice). It is also recorded in the learning state trajectory map, forming the foundational data for subsequent group cognitive analysis.

Claims

1. The distance education data processing system is characterized by The following steps are involved: S1. Data collection collaboration mechanism using cognitive scenario modeling: S1.

1. Predefine task intent and data expectations for each distance learning task, including lectures, answering questions, and discussions. S1.

2. All terminal acquisition devices, including microphones, cameras, and mice / touch screens, are coordinated to align with the current task objectives. A simultaneous acquisition and screening model is adopted to retain data segments that are strongly related to cognitive tasks. Semantic timestamps and task-binding tags are added to each data segment to create a traceable data layer structure. S2. Using semantic layer data fusion and interpretation mechanism: S2.

1. Map the collected data, including silence + eye movement drift → prediction confusion, into a unified learning semantic vector; introduce a ternary modeling approach of micro-expression + interactive behavior + temporal decision path to identify the potential cognitive intention corresponding to the feature combination; S2.

2. Integrate teaching context information, including the difficulty of teaching content, explanation speed, and knowledge point switching rhythm; build a cognitive state mapping model: infer the current cognitive state label from multimodal perception data; S3. Adopting the teaching intervention decision chain generation mechanism: S3.1 Based on the student's current cognitive state label, search the historical database for intervention trajectory VS effect feedback data; use the sequence modeling algorithm to generate a predicted intervention behavior sequence; S3.

2. Adopt a dynamic combined suggestion chain including gentle prompts → content reconstruction → personalized practice → tutoring invitation; choose whether to intervene in any of these nodes, so that the system and teachers can form a co-governance intervention mechanism; S4. Multi-user cognitive trajectory collaborative analysis mechanism: S4.

1. Superimpose all students’ cognitive state sequences into a group cognitive trajectory map; identify collective cognitive breakpoints; S4.

2. Create a map of hot misunderstanding areas to prompt teachers to repeat, change the lecture, or insert problem sections.

2. The remote education data processing system according to claim 1, characterized in that The method for constructing a data acquisition collaborative mechanism for cognitive scenario modeling includes: The teaching platform performs semantic recognition of current teaching activities, generates teaching intention identification, marks teaching objectives, and carries a single set of data expectation points, predicts the data type and manifestations including facial micro-movements, gaze drift, and interaction delay in the current teaching situation; each type of data expectation point forms a dynamic mapping model based on historical data and teacher settings to guide the scheduling of subsequent collection behaviors.

3. The remote education data processing system according to claim 2, characterized in that The method for constructing a data acquisition collaborative mechanism for cognitive scenario modeling includes: Real-time evaluation of whether each frame or set of data is consistent with the expected behavior of the task; using a cognitive relevance evaluation function, running locally on the terminal, and providing instant feedback on the value index of the data; the relevance function comprehensively considers the degree of match between behavioral characteristics and predefined cognitive templates, while also weighing signal interference and response sensitivity: in: R(t) is the cognitive relevance score of the data at time t; α is the interference tolerance constant preset by the system, which is used to limit the impact of low-quality data on the score; A(x) is the cognitive behavior matching function, and its derivative is represents the rate at which the degree of match between the data behavior and the target cognitive template changes over time; Γ(x) represents the signal confidence weight function currently sampled by the data acquisition device, which is used to dynamically adjust the impact of data quality; the integration interval [t0, t] represents the continuous process from the start of the current task phase to the current one; the cognitive strength of the data is evaluated in real time, and the threshold θ is set to determine whether the current segment should be retained or discarded.

4. The remote education data processing system according to claim 3 is characterized in that The method for constructing a data acquisition collaborative mechanism for cognitive scenario modeling includes: Based on optimizing system resources and collection density, a future data value trend prediction mechanism is introduced to assist in determining whether high-frequency collection is needed. in: V(Δt) is the system's prediction of the overall trend of data value over the next Δt period starting from the current time t; Λ(t′) is the student's sensitivity coefficient to the task response at time t′, with a high value indicating a volatile state and high acquisition value. It represents the derivative of the task synchronization function with respect to time, reflecting whether the learning behavior is still closely following the task process; if V(Δt)<∈, that is, the future trend is insufficient, the system reduces the collection rate and retains the summary information.

5. The remote education data processing system according to claim 1 is characterized in that The method for constructing a semantic layer data fusion and interpretation mechanism is as follows: Based on the real-time collection of multimodal behavioral data including eye movement trajectories, mouse movements, and silent periods, the original perceptual features are encoded and compressed and mapped into a unified learning semantic vector; the semantic vector is an expression of the cognitive meaning of the behavioral combination; This includes continuous eye drift + long periods of silence being interpreted as students being confused but not actively providing feedback; frequent revisions of answers + no facial emotion being understood as knowledge retrieval driven by recall; and establishing cognitive input units.

6. The remote education data processing system according to claim 5, characterized in that The method for constructing a semantic layer data fusion and interpretation mechanism is as follows: A ternary feature combination modeling method based on micro-expression + interactive behavior + decision path; extracting micro-expression frequency bands including eyebrow trembling and eyelid shrinking from the visual expression recognition model, and recognition operation features including looking back behavior and mouse hovering from the behavioral trajectory, and simultaneously analyzing the path evolution logic of students in teaching tasks, and constructing a feature tensor with the three as a unified structure.

7. The remote education data processing system according to claim 6, characterized in that The method for constructing a semantic layer data fusion and interpretation mechanism is as follows: The generated semantic vector L v , ternary model output tensor and teaching context variables C including explanation speed, content difficulty, and knowledge point conversion frequency n Fusion,forms a structured input sequence, which is then fed into the cognitive state mapping engine,,generating the current student's cognitive state label through nonlinear,reasoning function; Label generation is based on the following state inference function: in: Υ(S m ) is the student’s state label set S at the current moment m The prediction confidence value in S m It includes deep understanding, cognitive rupture, surface memory, passive acceptance of false confidence state labels; N is the dimension of cognitive state space; φ n (·) is the feature mapping function of the nth state, and its input is the learning semantic vector L v , behavioral feature tensor and context factor C n ;λ n is the sensitivity weight of the state label, which controls the response to the input feature; κ is the normalization coefficient, which is used to adjust the range of the predicted value.

Images