A method for real-time analysis and scheduling of multi-source data in online education

Abstract

This invention discloses a real-time analysis and scheduling method for multi-source data in online education, belonging to the field of online education technology. It includes: S1, dividing the accessed multi-source heterogeneous educational data into data blocks, extracting primary features, and labeling them to form a uniformly encoded data block sequence; S2, using hierarchical feature extraction and multimodal fusion based on confidence assessment, and using a confidence threshold as a gate to split the data blocks into high-confidence real-time streams and low-confidence delayed streams; S3, constructing dedicated analysis agents, and adaptively scheduling analysis tasks to available computing nodes for parallel execution; S4, using a dual deep Q-network to quantify the multi-dimensional learning status of the analysis results of each agent, and using the quantification results to adjust the block parameters of S1 and the scheduling strategy of S3 through a closed-loop feedback link. Using this method, multimodal data standardization processing can be achieved, confidence-based splitting reduces ineffective computing power, agents dynamically adapt to the classroom, and manual operation and maintenance costs can be reduced.

Description

Technical Field

[0001] This invention relates to the field of online education technology, and in particular to a method for real-time analysis and scheduling of multi-source data in online education. Background Technology

[0002] With the large-scale development of online education platforms, the number of learners online simultaneously in a single classroom can reach thousands or even tens of thousands. This generates massive amounts of multimodal and heterogeneous educational data, including but not limited to learning behavior logs (clicks, pauses, answers, progress), live audio and video streams (teacher's presentation and audio), learner eye movements and physiological signals (attention-related neurophysiological data), text feedback data (course comments, bullet comments, homework feedback), and instant messaging interaction data (group chats, Q&A forum discussions). These data differ significantly in sampling frequency, data format, temporal granularity, and semantic level, constituting a typical multi-source heterogeneous data stream.

[0003] Existing online education data analysis systems mainly suffer from the following technical deficiencies: Firstly, the data processing layer generally adopts the "full data collection first, then offline batch processing" model, which stores data of different modalities in a data warehouse and then performs learning analysis through offline computing tasks. This results in a delay of minutes or even hours from data generation to the output of analysis results, which cannot meet the timeliness requirements of real-time intervention in live classrooms (such as reminders for decreased attention and knowledge point reinforcement push). Some solutions that attempt streaming processing execute the complete analysis pipeline indiscriminately on all data blocks, causing congestion of computing resources during high-concurrency periods, and the processing delay of key learning tasks is actually lengthened.

[0004] Secondly, the feature extraction layer often uses single-modal or simply spliced ​​multimodal features, lacking the ability to perceive the differences in data quality between different modalities. When a certain modality has missing data or noise due to network jitter, equipment failure, or abnormal learner behavior, the system still merges it with high-quality modalities with equal weight, resulting in the fused features being contaminated by low-quality modalities, and the accuracy of downstream learning classification drops significantly. At the same time, the existing solutions lack explicit modeling of the confidence level of the analysis results, and cannot distinguish between "high-confidence reliable conclusions" and "low-confidence conclusions to be confirmed", allowing misjudgments driven by low-quality data to directly enter the teaching decision-making process.

[0005] Third, task scheduling generally adopts static priority or fixed resource allocation strategies, which pre-allocate computing resources to different analysis tasks such as learning analysis, behavior recognition, sentiment analysis, and attention assessment. This makes it impossible to adaptively adjust according to the dynamic changes in real-time learning status. For example, when there is a large-scale decline in attention in the classroom, the computing demand for attention assessment tasks surges, but static scheduling cannot promptly migrate resources from low-priority tasks (such as routine progress statistics), causing critical tasks to queue up and wait.

[0006] Furthermore, different analysis tasks have different dependencies on input modalities. Existing schedulers lack explicit descriptions of task-modal conditional dependencies, leading to tasks being scheduled to computation nodes that do not have the required modal data, resulting in redundant data transmission and wasted computation. Fourth, the output of analysis results is mostly based on single-dimensional learning indicators (such as online time and answer accuracy), lacking joint quantification of multi-dimensional learning information such as learners' audiovisual attention resource allocation, changes in emotional tendencies, and interaction quality. Moreover, there is a lack of closed-loop feedback mechanism between the analysis results and the teaching scheduling process. The analysis conclusions are only used for post-event report statistics and cannot drive the dynamic optimization of the data processing pipeline. Summary of the Invention

[0007] The purpose of this invention is to provide a method for real-time analysis and scheduling of multi-source data in online education, thereby solving the aforementioned technical problems.

[0008] To achieve the above objectives, this invention provides a method for real-time analysis and scheduling of multi-source data in online education, comprising the following steps: S1. Spatiotemporal segmentation and unified coding of multi-source educational data: The accessed multi-source heterogeneous educational data is divided into data blocks according to a fixed time window, and primary features are extracted from each modality of data and labeled with task identifiers, condition masks and location types to form a unified coding sequence of data blocks; S2. Hierarchical feature extraction and multimodal fusion based on confidence assessment: shallow basic feature extraction and deep cross-modal attention weighted fusion are performed sequentially on the unified coded data blocks. The analysis confidence of each data block is calculated based on data quality, modal integrity and temporal stability. The data blocks are divided into high-confidence real-time stream and low-confidence delayed stream by using a confidence threshold as a gate. After the low-confidence block is corrected by neighbor voting, it is judged whether the corrected confidence meets the standard. If it meets the standard, it is transferred to the real-time stream. S3. Real-time dynamic task scheduling based on multi-agent and conditional mask: Construct four dedicated analytical agents for learning analysis, behavior recognition, sentiment analysis, and attention evaluation. Based on the matching degree between the conditional dependency vector of the task and the conditional mask of the data block, the dynamic priority driven by the learning state transition matrix, and the multi-objective optimization strategy, the analytical tasks are adaptively scheduled to available computing nodes for parallel execution. S4. Quantitative Output and Adaptive Feedback Control of Learning Status: A dual deep Q-network (DDQN) is used to quantify the learning status of each agent in multiple dimensions based on the analysis results. The output includes a joint learning status index that includes focus, emotional tendency, interaction quality and audiovisual attention resource allocation. The quantification results are then used to adjust the block parameters of S1 and the scheduling strategy of S3 through a closed-loop feedback link.

[0009] Preferably, the multi-source educational data in S1 includes five categories: learning behavior log modality, audio and video stream modality, eye-tracking and physiological signal modality, text feedback modality, and interactive communication modality; the spatiotemporal segmentation adopts a fixed time window. Overlap ratio with adjacent blocks The sliding window mechanism; uniform coding adds a task identifier to each modality feature. Conditional mask and location type ,in, Number the data blocks, The modal index is, and The condition mask is automatically generated through data integrity verification, indicating whether the modal data is complete and usable.

[0010] Preferably, the deep cross-modal attention weighted fusion in S2 specifically includes: First, calculate the quality score for each modality. : ; in, For signal-to-noise ratio, The missing rate, For temporal continuity, , , As the weight for quality scoring, and + + =1; Then, the quality score is normalized using softmax to obtain the fusion weight. ; Final fusion features for: ; in, For the first The data block number The shallow features extracted from the modalities are transformed into deep features of the single modality by the fundamental transformation.

[0011] Preferably, the formula for calculating the confidence level in S2 is: ; in, To analyze the confidence scores, As a comprehensive indicator of data quality, Modal completeness, which is the proportion of effective modes to the total number of modes. For timing stability, , , To analyze confidence weights, and + + =1; Confidence gating setting threshold , High-confidence blocks are directly fed into the real-time pipeline. Low-confidence blocks are transferred to the delay queue; The corrected formula for neighbor voting is: ; in, This is the corrected analysis confidence score. For the front and back A set of neighbors consisting of adjacent data blocks To preset the hyperparameters of the neighbor window, This represents the total number of neighboring blocks. For voting weight; If the confidence level is corrected The data block is then fed into the real-time stream.

[0012] Preferably, all four dedicated analytical agents in S3 perform tasks using a reasoning-action-observation paradigm; each analytical task uses a conditional dependency vector. Describes the conditional dependency on the input modality, and calculates the conditional matching degree during scheduling. : ; in, For the first A complete multimodal data block after S1 encoding. ∈{0,1}, =1 means the task requires the first... Modal class, =0 means that this mode is not needed; When the matching degree is not lower than the threshold When this happens, data blocks are allocated to the corresponding tasks to avoid scheduling tasks to computing nodes that do not have the required modalities.

[0013] Preferably, the S3 learning state space includes four states: high concentration (normal learning), medium concentration (passive learning), low concentration (requiring intervention), and abnormal behavior (requiring alert). The state transition probability is determined by the state transition matrix. Description: During operation, the matrix dynamically adjusts its corresponding elements and normalizes them row by row based on real-time detected features such as sudden drops in attention, surges in interaction, and abnormal behavior. The formula for calculating the dynamic scheduling priority of tasks is: ; in, Based on task priority, Let be the transition probability from the current state to the target state. Given the current learning situation, The learning status of the target group to be intervened in. For data confidence, , , As the scheduling weight, and + + =1, schedule tasks from highest to lowest priority.

[0014] Preferably, the objective function of multi-objective optimization scheduling in S3 is... for: ; in, To average processing latency, To average resource consumption, The average analysis accuracy. , , Optimize scheduling weights for multi-objective purposes. + + =1; The constraint is the real-time task processing latency. And calculate resource utilization rate When the delay was close to the upper limit, it was temporarily increased. Weighting is increased when resource utilization approaches its limit. Weights.

[0015] Preferably, in S4, DDQN models the learning quantification as a Markov decision process, using the deep fusion features and the intermediate results of each agent as the state. Actions are determined by the combined discrete values ​​of three dimensions: focus, emotional inclination, and interaction quality. The reward function is: ; in, For indicator functions, To predict learning labels for the model, Labeling as true learning progress To accurately predict the reward coefficient, The penalty coefficient for incorrect prediction. Action confidence; Training is performed using an experience replay mechanism, and the loss function includes temporal difference error and round-level constraint terms.

[0016] Preferably, in S4, the decomposition of audiovisual attention resource allocation adopts a forward coding model and variance partitioning technique, separating the total attention into three orthogonal components: visual attention contribution (VAC), auditory attention contribution (AAC), and shared audiovisual attention contribution (SAC), which satisfy the following: ; Low-confidence results are marked as pending confirmation and accompanied by a confidence value. Real-time alarms are triggered when the learning status is low focus or abnormal.

[0017] Preferably, the closed-loop feedback control in S4 includes two links: feedback link one dynamically adjusts the time window of S1 according to the proportion of low-confidence blocks. Overlap ratio and confidence threshold Feedback link two updates the learning status to the state transition matrix of S3 in real time, and dynamically adjusts the basic priority of tasks based on the accuracy statistics of each agent. and scheduling weight , , This forms a closed-loop adaptive optimization.

[0018] Therefore, the above-mentioned method for real-time analysis and scheduling of multi-source data in online education has the following beneficial effects: 1. This invention introduces a spatiotemporal segmentation and unified encoding mechanism at the data access layer. Heterogeneous multi-source data is divided into data blocks according to a fixed time window and encoded into a unified feature token sequence carrying task identifiers, condition masks, and location types. At the same time, an analysis confidence assessment and threshold determination mechanism is introduced at the feature layer. High-confidence data blocks directly enter the real-time analysis pipeline, while low-confidence data blocks are transferred to a delay queue and corrected by neighbor voting before processing. This ensures the real-time performance of key data while avoiding low-quality data from occupying real-time computing resources, thereby reducing the processing latency and computing congestion during high-concurrency periods.

[0019] 2. This invention designs a two-layer structure in the feature extraction layer, consisting of shallow basic feature extraction and deep cross-modal attention weighted fusion. The fusion weights of each modality are dynamically determined by the modality quality score after softmax normalization, and low-quality modalities automatically receive lower weights. At the same time, the analysis confidence of each data block is calculated based on three dimensions: data quality, modality integrity, and temporal stability. The confidence is used as a unified quality gating signal for feature fusion, task scheduling, and result output, thereby preventing low-confidence conclusions from directly entering teaching decisions.

[0020] 3. This invention constructs a multi-agent analysis architecture (four dedicated agents: learning analysis, behavior recognition, sentiment analysis, and attention assessment) at the scheduling layer, and introduces condition masks and task identifiers to explicitly describe the conditional dependencies of each task on the input modality, so that the scheduler only assigns tasks to computing nodes with the required modality; at the same time, it designs a dynamic priority scheduling strategy based on the learning state transition matrix. When real-time learning detects a decline in attention or abnormal behavior, it automatically increases the scheduling priority of the corresponding analysis task and migrates computing resources, realizing the adaptive response of the scheduling strategy to changes in classroom state.

[0021] 4. This invention adopts multi-dimensional learning status quantification based on DDQN at the output layer, outputting a joint learning status index including focus level, audiovisual attention resource allocation, emotional tendency and interaction quality; and feeds back the learning status quantification results to the state transition matrix of the scheduling layer and the block parameters of the data layer, forming a closed-loop optimization link of "data-feature-scheduling-output-data", enabling the system to continuously self-adjust during operation.

[0022] 5. This invention integrates data quality perception, confidence gating, conditional scheduling, and closed-loop feedback into a unified architecture. The four stages are progressively layered, with bidirectional flow of data and control signals. Compared with existing schemes where each stage is designed independently, this invention can simultaneously achieve lower latency for critical tasks, higher accuracy in learning analysis, and better resource utilization under the same computing resources, achieving a synergistic gain effect from the integration of multiple technologies.

[0023] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0024] Fig. 1 This is an overall flowchart of a real-time analysis and scheduling method for multi-source data in online education according to the present invention; Fig. 2 This is a flowchart of the hierarchical feature extraction and confidence level splitting process of the present invention; Fig. 3 This is a diagram of the multi-agent conditional mask scheduling architecture of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0026] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0027] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0028] like Figs. 1-3 As shown, the present invention provides a method for real-time analysis and scheduling of multi-source data in online education, comprising the following steps: S1. Spatiotemporal segmentation and unified coding of multi-source educational data, specifically including: (1) Time-series alignment of five types of multimodal data: access five types of standard online education data sources, modality number The corresponding modalities are, in order: learning behavior log modal, audio and video stream modal, eye-tracking + EEG / fNIRS physiological signal modal, text feedback modal, and interactive communication data modal.

[0029] Discrete event logs and text data are aligned to a global baseline timestamp using linear interpolation. High-frequency audio and video signals and physiological signals are downsampled through feature aggregation to unify the time reference, eliminate time offset of multi-source data, and provide a unified time reference for sliding block partitioning.

[0030] (2) Sliding window spatiotemporal segmentation mechanism: using a fixed time window (Typical industrial value: 1 second), Overlap ratio of adjacent blocks The sliding block rule (typically 0.2) generates a number. data blocks The time interval of a single block is The timing intervals of adjacent blocks overlap by 20% to avoid the loss of timing features at the window boundaries.

[0031] (3) Modal primary feature extraction and information annotation: for data blocks Inner Modality extraction industry-standard primary features : =1 Behavior Log: In-window quiz accuracy rate, video dwell time, and page click statistics; =2 Audio and Video: Audio MFCC features, video keyframe frame difference features; =3 Physiological signals: eye movement fixation dispersion, EEG power at each frequency band, and mean HbO / HbR concentration; =4 Text: Sentence vectors output by the pre-trained language model; =5 Interactions: Raising hands, connecting via voice chat, and message quantity statistics.

[0032] Three types of meta-identifiers are added to the primary features: task identifiers. : Corresponding analysis task number (1 Learning situation, 2 Behavior, 3 Emotion, 4 Attention); Conditional mask Data integrity verification is automatically generated; 1 indicates complete and usable modal data, and 0 indicates missing data or high noise; Location type : Identify the time-series role of data blocks and uniformly mark real-time processing blocks. After encoding, output standardized data blocks. .

[0033] S2. Hierarchical feature extraction and multimodal fusion based on confidence assessment, specifically including: (1) Shallow feature base transformation: for =1 effective modal primary features Shallow transformations are performed using batch normalized (BN) layers with fully connected layers to generate single-modal deep features. ; Missing modes with a value of 0 are directly set to zero vectors.

[0034] (2) Modal quality weighted cross-modal fusion: Modal quality scoring The calculation formula is: ; in, This is the modal signal-to-noise ratio of this block. This represents the missing data rate for this block. The cosine similarity of the features between the current block and the previous block is equal to the sequential continuity. , , As a weight for quality scoring, + + =1, Project Recommendation Weight =0.4、 =0.3、 =0.3; Softmax normalization is performed on all modal quality scores to obtain the fusion weights for each modality. Weighted splicing generates global deep fusion features. for: ; (3) Calculation of confidence level for three-dimensional analysis, wherein the formula for calculating confidence level is: ; in, To analyze the confidence scores, As a comprehensive indicator of data quality, Modal completeness, which is the proportion of effective modes to the total number of modes. For timing stability, , , To analyze confidence weights, and + + =1, Recommendation weight =0.5、 =0.3、 =0.2; Set a fixed confidence threshold (Typical value 0.6) Implement data splitting: like It was determined to be a high-confidence block, and its features were fused. Directly feed into the S3 real-time scheduling pipeline; if If it is determined to be a low-confidence block, it is stored in the delayed buffer and neighbor voting is performed to correct it.

[0035] (4) Neighbor voting confidence correction: setting For the preceding and following blocks A set of neighbors consisting of adjacent data blocks The typical industrial value is 5; the correction formula is: ; in, This is the corrected analysis confidence score. To preset the hyperparameters of the neighbor window, This represents the total number of neighboring blocks. For voting weight, Recommended value: 0.4; after correction, if The data block is upgraded to real-time data and sent to S3; if it is still below the threshold... The cache continues to wait for more context, exceeding the maximum wait time. Then, mark it as "Pending Confirmation" and proceed to the production line.

[0036] S3. Real-time dynamic task scheduling based on multi-agent and condition masking, specifically including: (1) Four types of ReAct-specific analytical agents: Four independent specialized agents are constructed, all adopting the reasoning-action-observation (ReAct) general paradigm, including: Learning analysis agent: Based on behavioral log modalities, it assesses the mastery of knowledge points and deviations in learning progress; Behavioral recognition intelligent agent: Based on audio and video and interactive modalities, it identifies classroom anomalies such as screen absence and frequent switching; Sentiment Analysis Agent: Based on text and speech modalities, it completes educational-level sentiment ABSA analysis; Attention assessment agent: Based on eye movements and physiological modalities, it quantifies the learner's audiovisual attention level.

[0037] (2) Modal condition matching filtering: Define task condition dependency vector , ∈{0,1}, =1 means that this task must use the first... Modal class, =0 means that this mode is not needed; Formula for calculating condition matching degree: ; Matching threshold Recommendation 0.7, only for matches with a score ≥ Data blocks can be allocated to corresponding agents, filtering out invalid computation tasks that do not match the modality.

[0038] (3) Learning status transition matrix and dynamic priority: Define four types of learning status High concentration on normal learning Focus on passive learning Low focus requires intervention. Alerts for abnormal behavior; State transition matrix The initial values ​​are set based on prior knowledge from online teaching; during operation, three types of features are detected: sudden changes in attention, surges in interaction, and abnormal behavior. The probabilities of the corresponding rows of the matrix are adjusted in real time and normalized row by row. Task dynamic priority calculation formula: ; in, The basic priority of the task is (attention and abnormal tasks 0.8, normal learning situation 0.4). Let be the transition probability from the current state to the target state. Given the current learning situation, The learning status of the target group to be intervened in. For data confidence, , , As the scheduling weight, and + + =1, Recommendation weight =0.3、 =0.4、 =0.3, the scheduler follows Tasks are distributed from high to low to idle computing nodes.

[0039] (4) Multi-objective optimization scheduling constraints: The optimization objective function is: ; in, To average processing latency, To average resource consumption, The average analysis accuracy. , , Optimize scheduling weights for multi-objective purposes. + + =1, Recommendation weight =0.4、 =0.4、 =0.2; Hard constraint: Real-time task processing latency =2s, node resource utilization rate =0.85; When the real-time latency approaches the 2-second limit, it will automatically increase. Weighted compression processing time is increasing; node load is approaching 85% limit, increasing... Weight control controls computing power consumption.

[0040] (5) Aggregation of parallel results of intelligent agents: Four types of intelligent agents perform inference in parallel, output intermediate learning results of each dimension, and uniformly aggregate and send them into S4.

[0041] S4. Quantitative output and adaptive feedback control of learning status, specifically including: (1) DDQN Multidimensional Learning Quantification: Modeling learning assessment as a Markov decision process: state S2 deep fusion features Piecing together intermediate analysis results of four types of S3 intelligent agents; action : A combination of discrete tags based on three dimensions: focus, emotional inclination, and interaction quality; Asymmetric reward function: ; in, For indicator functions, To predict learning labels for the model, Labeling as true learning progress To accurately predict the reward coefficient, this embodiment uses a value of 10. The error prediction penalty coefficient is set to 1 in this embodiment. The action confidence level output by S2 is used; DDQN is trained using a general experience replay mechanism, and the loss function is superimposed with the temporal difference error and the classroom round constraint term to output a standardized multidimensional learning level.

[0042] (2) Orthogonal decomposition of audiovisual attention: Using forward encoding + variance partitioning, the total attention is decomposed into three orthogonal indices: VAC (visual attention contribution), AAC (auditory attention contribution), and SAC (auditory-auditory shared attention). VAC, AAC, and SAC are non-overlapping orthogonal attention components. After summing the three, normalization is performed, and the value range is uniformly set to [0,1]. That is, the three satisfy the following constraints: ; For low-confidence quantification results, mark them as "Pending Confirmation" and attach the corresponding information. Numerical values; when a student's learning situation is judged to be low concentration or abnormal, a real-time intervention alert is pushed to the teacher's end.

[0043] (3) Dual closed-loop feedback control link: Feedback link one can reverse control the S1 block parameters, specifically: periodically count the proportion of low-confidence data blocks in the delay queue, and if the proportion is >30%, the window is expanded and the confidence threshold is increased. If the percentage is less than 10% and there is sufficient real-time latency, then reduce the window size to improve performance. Timing resolution, synchronously adjust overlap ratio The feedback link 2 can reversely control the S3 scheduling system, specifically by: real-time learning data... Continuously write the state transition matrix Update transition probabilities; calculate the long-term analysis accuracy of each agent and dynamically adjust the basic task priority. With scheduling weight The two feedback loops continuously iterate, and the entire method can adaptively adapt to classroom data and learner status without human intervention.

[0044] Online implementation examples: Cluster deployment: 12 distributed computing nodes with 8 cores and 16G bandwidth, supporting 4,000 simultaneous online learners in a single classroom. All parameters adopt the above-mentioned engineering recommended values.

[0045] Actual test results: The average processing latency for intervention-type learning tasks is 1.2s (compared to 40~120s for traditional offline solutions); the accuracy of comprehensive learning identification is 11.3% higher than that of traditional fixed-weight fusion solutions; the average resource utilization of the cluster is stable at 72%, with no invalid calculations caused by modality mismatch; the system can automatically adjust the block window and scheduling priority according to the students' attention level on the day, without requiring maintenance personnel to manually configure parameters every day.

[0046] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for real-time analysis and scheduling of multi-source data in online education, characterized in that, Includes the following steps: S1. Spatiotemporal segmentation and unified coding of multi-source educational data: The accessed multi-source heterogeneous educational data is divided into data blocks according to a fixed time window, and primary features are extracted from each modality of data and labeled with task identifiers, condition masks and location types to form a unified coding sequence of data blocks; S2. Hierarchical feature extraction and multimodal fusion based on confidence assessment: shallow basic feature extraction and deep cross-modal attention weighted fusion are performed sequentially on the unified coded data blocks. The analysis confidence of each data block is calculated based on data quality, modal integrity and temporal stability. The data blocks are divided into high-confidence real-time stream and low-confidence delayed stream by using a confidence threshold as a gate. After the low-confidence block is corrected by neighbor voting, it is judged whether the corrected confidence meets the standard. If it meets the standard, it is transferred to the real-time stream. S3. Real-time dynamic task scheduling based on multi-agent and conditional mask: Construct four dedicated analytical agents for learning analysis, behavior recognition, sentiment analysis, and attention evaluation. Based on the matching degree between the conditional dependency vector of the task and the conditional mask of the data block, the dynamic priority driven by the learning state transition matrix, and the multi-objective optimization strategy, the analytical tasks are adaptively scheduled to available computing nodes for parallel execution. S4. Quantitative Output and Adaptive Feedback Control of Learning Status: A dual deep Q-network is used to quantify the learning status of each agent in multiple dimensions based on the analysis results. The output includes a joint learning status index that includes focus, emotional tendency, interaction quality and audiovisual attention resource allocation. The quantification results are then used to adjust the block parameters of S1 and the scheduling strategy of S3 through a closed-loop feedback link.

2. The online education multi-source data real-time analysis and scheduling method according to claim 1, characterized in that: The multi-source educational data in S1 includes five categories: learning behavior log modality, audio and video stream modality, eye-tracking and physiological signal modality, text feedback modality, and interactive communication modality; the spatiotemporal segmentation adopts a fixed time window. Overlap ratio with adjacent blocks The sliding window mechanism; uniform coding adds a task identifier to each modality feature. Conditional mask and location type ,in, Number the data blocks, The modal index is, and The condition mask is automatically generated through data integrity verification, indicating whether the modal data is complete and usable.

3. The online education multi-source data real-time analysis and scheduling method according to claim 2, characterized in that, The deep cross-modal attention-weighted fusion in S2 specifically includes: First, calculate the quality score for each modality. : ； in, For signal-to-noise ratio, The missing rate, For temporal continuity, , , As the weight for quality scoring, and + + =1; Then, the quality score is normalized using softmax to obtain the fusion weight. ; Final fusion features for: ； in, For the first The data block number The shallow features extracted from the modalities are transformed into deep features of the single modality by the fundamental transformation.

4. The online education multi-source data real-time analysis and scheduling method according to claim 3, characterized in that: The formula for calculating the confidence level in S2 is: ； in, To analyze the confidence scores, As a comprehensive indicator of data quality, Modal completeness, which is the proportion of effective modes to the total number of modes. For timing stability, , , To analyze confidence weights, and + + =1; Confidence gating setting threshold , High-confidence blocks are directly fed into the real-time pipeline. Low-confidence blocks are transferred to the delay queue; The corrected formula for neighbor voting is: ； in, This is the corrected analysis confidence score. For the front and back A set of neighbors consisting of adjacent data blocks To preset the hyperparameters of the neighbor window, This represents the total number of neighboring blocks. For voting weight; If the confidence level is corrected The data block is then fed into the real-time stream.

5. The online education multi-source data real-time analysis and scheduling method according to claim 4, characterized in that: In S3, all four dedicated analytical agents perform tasks using a reasoning-action-observation paradigm; each analytical task uses a conditional dependency vector. Describes the conditional dependency on the input modality, and calculates the conditional matching degree during scheduling. : ； in, For the first A complete multimodal data block after S1 encoding. ∈{0,1}, =1 means the task requires the first... Modal class, =0 means that this mode is not needed; When the matching degree is not lower than the threshold When this happens, data blocks are allocated to the corresponding tasks to avoid scheduling tasks to computing nodes that do not have the required modalities.

6. The online education multi-source data real-time analysis and scheduling method according to claim 1, characterized in that: The S3 learning state space includes four states: high concentration (normal learning), medium concentration (passive learning), low concentration (intervention required), and abnormal behavior (alarm required). The state transition probabilities are determined by the state transition matrix. Description: During operation, the matrix dynamically adjusts its corresponding elements and normalizes them row by row based on real-time detected features such as sudden drops in attention, surges in interaction, and abnormal behavior. The formula for calculating the dynamic scheduling priority of a task is: ； in, Based on task priority, Let be the transition probability from the current state to the target state. Given the current learning situation, The learning status of the target group to be intervened in. For data confidence, , , As the scheduling weight, and + + =1, schedule tasks from highest to lowest priority.

7. The online education multi-source data real-time analysis and scheduling method according to claim 1, characterized in that: The objective function of multi-objective optimization scheduling in S3 for: ； in, To average processing latency, To average resource consumption, The average analysis accuracy. , , Optimize scheduling weights for multi-objective purposes. + + =1; The constraint is the real-time task processing latency. And calculate resource utilization rate When the delay was close to the upper limit, it was temporarily increased. Weighting is increased when resource utilization approaches its limit. Weights.

8. The online education multi-source data real-time analysis and scheduling method according to claim 1, characterized in that: In S4, the dual deep Q-network models the learning quantification as a Markov decision process, using the deep fusion features and the concatenation of intermediate results from each agent as the state. Actions are determined by the combined discrete values ​​of three dimensions: focus, emotional inclination, and interaction quality. The reward function is: ； in, For indicator functions, To predict learning labels for the model, Labeling as true learning progress To accurately predict the reward coefficient, The penalty coefficient for incorrect prediction. Action confidence; Training is performed using an experience replay mechanism, and the loss function includes temporal difference error and round-level constraint terms.

9. The online education multi-source data real-time analysis and scheduling method according to claim 1, characterized in that: In S4, the decomposition of audiovisual attention resource allocation employs a forward coding model and variance partitioning technique, separating total attention into three orthogonal components: visual attention contribution (VAC), auditory attention contribution (AAC), and shared audiovisual attention contribution (SAC). These three components satisfy the following: ； Low-confidence results are marked as pending confirmation and accompanied by a confidence value. Real-time alarms are triggered when the learning status is low focus or abnormal.

10. The online education multi-source data real-time analysis and scheduling method according to claim 2, characterized in that, S4's closed-loop feedback control includes two links: Feedback link one dynamically adjusts the time window of S1 based on the proportion of low-confidence blocks. Overlap ratio and confidence threshold Feedback link two updates the learning status to the state transition matrix of S3 in real time, and dynamically adjusts the basic priority of tasks based on the accuracy statistics of each agent. and scheduling weight , , This forms a closed-loop adaptive optimization.

Images