Education data collaborative processing method and system for multi-scene interconnection of schools

Abstract

The present application relates to the field of education informatization and wisdom education data processing technology, in particular to a multi-scene interconnected education data collaborative processing method and system for schools, families and other scenes; comprising: constructing a multi-dimensional orthogonal state space and initializing a target entity baseline state vector; converting multi-scene terminal original interaction events into state bias vectors through a domain knowledge graph; fitting an actual state vector in the same time window in combination with a decay weight and a baseline state vector; comparing the actual state vector with an expected vector obtained from the baseline state vector and a target timestamp to generate a cross-domain residual distribution map; generating a collaborative intervention instruction based on the cross-domain residual distribution map and delivering it to a scene execution end for closed-loop control; the present application realizes second-level education collaborative processing, reduces cross-system data coupling, data transmission volume and privacy exposure risk.

Description

Technical Field

[0001] This invention relates to the field of educational informatization and smart education data processing technology, specifically to a collaborative processing method and system for educational data interconnection across multiple scenarios involving schools, bureaus, and families. Background Technology

[0002] In the interconnected educational data processing involving multiple scenarios such as bureaus, schools, and families, continuous status assessment and collaborative intervention around student, class, or school entities are gradually becoming an important technical direction for achieving refined processing of regional educational data.

[0003] Currently, for multi-source heterogeneous data generated by education bureau management platforms, school logistics systems, and home terminals, the common approach is to centrally aggregate the original detailed data and then perform unified analysis. If cross-scenario status assessment and linkage intervention are required, it is usually necessary to perform field alignment, semantic transformation, and result aggregation on the central platform to obtain the comprehensive evaluation results of the corresponding entities in the current time period.

[0004] However, when processing multi-scenario data from schools, bureaus, and families using the above methods, it is necessary to directly aggregate different types of raw events such as cafeteria consumption, daily schedule records, content browsing, and attendance / exercise. This leads to the easy mixing of business meanings for similar indicators in different systems, making it impossible to stably characterize the continuous state of target entities under a unified semantic coordinate system. Furthermore, it is difficult to distinguish and cumulatively fit the temporal impact of events from different scenarios within the same time window, resulting in cross-domain data semantic inconsistency and insufficient real-time state evolution capabilities. At the same time, this processing method requires the transmission and storage of a large amount of raw data, which may lead to bandwidth consumption exceeding the preset communication threshold and security risks related to the exposure of privacy data. Moreover, it relies on manual or semi-manual methods to interpret and distribute the analysis results, making it difficult to quickly form an automated closed-loop intervention based on the source of anomalies, resulting in collaborative control delays exceeding the preset time threshold. Summary of the Invention

[0005] To address the aforementioned technical issues, this invention provides a method and system for collaborative processing of educational data across multiple scenarios involving schools, bureaus, and families. Specifically, the technical solution of this invention includes: constructing a multi-dimensional orthogonal state space and initializing a baseline state vector for the target entity within this space; acquiring original interaction events generated by multiple scenario terminals on the data acquisition side, parsing entity identifiers in the original interaction events to match the corresponding target entities, and converting the original interaction events into state bias vectors using a preset domain knowledge graph to achieve semantic state synchronization across domain scenarios; extracting the state bias vector corresponding to the target entity within the same time window, performing multi-dimensional state vector fitting on the baseline state vector and the state bias vector in each independent orthogonal dimension, and calculating the currently accumulated actual state vector; comparing the actual state vector with the expected vector, performing cross-domain residual distribution calculation, and generating a cross-domain residual distribution map; wherein the expected vector is obtained by inputting the baseline state vector and the target timestamp corresponding to the current time window into a preset comprehensive evaluation model for prediction; generating collaborative intervention instructions based on the cross-domain residual distribution map, and sending the collaborative intervention instructions to the scenario execution terminal on the control side for closed-loop control.

[0006] Preferably, the business dimensions of macro-management scenario, teaching and logistics scenario and family routine scenario are extracted, and the business dimensions are mapped to the coordinate axes of multi-dimensional orthogonal state space; the historical evaluation features of the target entity are obtained, and the initial coordinate points are established in the multi-dimensional orthogonal state space based on the historical evaluation features; the initial coordinate points are transformed into multi-dimensional feature vectors, and the multi-dimensional feature vectors are confirmed as the baseline state vectors.

[0007] Preferably, the original interactive events generated by multiple terminals in the edge processing device are captured in real time; the original interactive events are input into the domain knowledge graph for semantic feature extraction, and the corresponding dimension of state micro-change values ​​are output. The domain knowledge graph at least includes entities and relationship networks of educational routines, nutritional health indicators and psychological cognitive development; the state micro-change values ​​are vectorized and encapsulated, and event timestamps and scene tags are attached to generate state bias vectors; the state bias vectors are asynchronously pushed to the preset collaborative bus.

[0008] Preferably, the state bias vector is extracted from the preset cooperative bus; the decay weight is calculated based on the event timestamp carried by the state bias vector, and a weighted summation operation is performed on the state bias vectors belonging to the same dimension to obtain the cumulative bias vector; the baseline state vector and the cumulative bias vector are fitted by vector addition to output the actual state vector.

[0009] Preferably, the baseline state vector and the target timestamp are input into the comprehensive evaluation model to obtain the expected vector output by the comprehensive evaluation model; the difference between the actual state vector and the expected vector in each independent orthogonal dimension is calculated, and the difference is split according to the scene label carried by the state bias vector to obtain the cross-domain residual matrix; the cross-domain residual matrix is ​​spatially mapped and rendered to generate a cross-domain residual distribution map, wherein the cross-domain residual distribution map is used to locate the abnormal dimension that causes the overall evaluation to shift.

[0010] Preferably, the residual values ​​in the cross-domain residual distribution map are extracted and the absolute values ​​of the residuals are calculated. The absolute values ​​of the residuals are compared with a preset tolerance threshold. If the absolute value of the residuals is higher than the tolerance threshold, it is determined that there is an abnormal deviation in the corresponding dimension. A first type of collaborative intervention instruction is generated according to the preset intervention rule base and sent to the corresponding scene execution terminal to perform the action adjustment. If the absolute value of the residuals is equal to the tolerance threshold, it is determined that the corresponding dimension is in a critical state. A second type of collaborative intervention instruction is generated and sent to the corresponding scene execution terminal to perform high-frequency monitoring. If the absolute value of the residuals is lower than the tolerance threshold, it is determined that the corresponding dimension is operating normally. A third type of collaborative intervention instruction is generated and sent to the corresponding scene execution terminal to maintain the current operating state.

[0011] Preferably, the scenario execution end includes a canteen management execution system and a converged media execution system; when the cross-domain residual distribution map indicates that the absolute value of the residual in the physical dimension is higher than the tolerance threshold, an instruction to adjust the nutritional ratio of the menu is issued to the canteen management execution system; when the cross-domain residual distribution map indicates that the absolute value of the residual in the psychological cognition dimension is higher than the tolerance threshold, an instruction to push precise content is issued to the converged media execution system; the vector change data generated after executing the collaborative intervention instruction is extracted, and the vector change data is recorded in the preset student comprehensive quality evaluation database as a basis for process evaluation.

[0012] This educational data collaborative processing system, designed for multi-scenario interconnection between schools, bureaus, and families, includes a spatial modeling unit, a semantic synchronization unit, a vector fitting unit, a residual calculation unit, and a closed-loop control unit. The spatial modeling unit constructs a multi-dimensional orthogonal state space and initializes baseline state vectors for target entities within this space. The semantic synchronization unit acquires raw interaction events generated by multiple scenarios from the data acquisition side, parses entity identifiers within these events to match corresponding target entities, and transforms the raw interaction events into state bias vectors using a pre-defined domain knowledge graph to achieve semantic state synchronization across domain scenarios. The vector fitting unit divides continuous data according to a pre-defined time series. Within a time window, the state bias vector corresponding to the target entity is extracted. Multidimensional state vector fitting is performed on the baseline state vector and the state bias vector in each independent orthogonal dimension to calculate the current accumulated actual state vector. The residual calculation unit compares the actual state vector with the expected vector, performs cross-domain residual distribution calculation, and generates a cross-domain residual distribution map. The expected vector is predicted by inputting the baseline state vector and the target timestamp corresponding to the current time window into a preset comprehensive evaluation model. The closed-loop control unit generates collaborative intervention commands based on the cross-domain residual distribution map and sends these commands to the scene execution terminal on the control side for closed-loop control.

[0013] Preferably, the semantic synchronization unit is deployed in an edge processing device, specifically used to capture raw interaction events generated by terminals in multiple scenarios in real time; the raw interaction events are input into a domain knowledge graph for semantic feature extraction, and the corresponding dimension of state micro-change values ​​are output. The domain knowledge graph at least includes entities and relationship networks related to educational routines, nutritional health indicators, and psychological cognitive development; the state micro-change values ​​are vectorized and encapsulated, and event timestamps and scene tags are added to generate a state bias vector; the state bias vector is asynchronously pushed to a preset collaborative bus.

[0014] Preferably, the closed-loop control unit is specifically used to extract residual values ​​from the cross-domain residual distribution map and calculate the absolute value of the residuals, and compare the absolute value of the residuals with a preset tolerance threshold; if the absolute value of the residuals is higher than the tolerance threshold, it is determined that there is an abnormal deviation in the corresponding dimension, a first type of collaborative intervention instruction is generated according to the preset intervention rule base, and the first type of collaborative intervention instruction is sent to the corresponding scene execution terminal to perform the action adjustment; if the absolute value of the residuals is equal to the tolerance threshold, it is determined that the corresponding dimension is in a critical state, a second type of collaborative intervention instruction is generated, and the second type of collaborative intervention instruction is sent to the corresponding scene execution terminal to perform high-frequency monitoring; if the absolute value of the residuals is lower than the tolerance threshold, it is determined that the corresponding dimension is operating normally, a third type of collaborative intervention instruction is generated, and the third type of collaborative intervention instruction is sent to the corresponding scene execution terminal to maintain the current operating state.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] 1. This invention captures raw interaction events generated by terminals in multiple scenarios in real time in edge processing devices, and extracts heterogeneous events such as cafeteria consumption, family routines, content browsing, and access control movement into corresponding dimensional state micro-change values ​​by combining domain knowledge graphs. This enables the pre-compression of raw detailed events into a unified format of state bias vectors, thereby reducing the central platform's reliance on heterogeneous field alignment and complex semantic transformations, and improving the efficiency of cross-scenario semantic state synchronization. By asynchronously pushing only state bias vectors with timestamps and scene tags without uploading the complete raw data, the burden of cross-network segment transmission and central storage pressure can be reduced, and the risk of direct exposure of sensitive data such as cafeteria consumption details and family logs can be reduced, thus balancing real-time collaborative processing capabilities with privacy control requirements.

[0017] 2. This invention extracts the state bias vector corresponding to the target entity within the same time window, calculates the decay weight based on the event timestamp, performs weighted summation on each independent orthogonal dimension, and then performs vector addition fitting with the baseline state vector. This can distinguish the different influence intensities of recent and distant events on the current state, thereby solving the problem of difficulty in temporal differentiation and cumulative fitting of events in different scenarios within the same time window, and improving the closeness of the actual state vector to the current state. By performing boundary truncation, abnormal time series removal, and high-concurrency homogeneous aggregation processing on the fitting results, the state solution process can be guaranteed to be stable, transparent, and interpretable, thereby improving the reliability of continuous state evolution processing in regional concurrent scenarios.

[0018] 3. This invention obtains the expected vector by inputting the baseline state vector and the target timestamp into the comprehensive evaluation model, and compares the actual state vector with the expected vector to calculate the cross-domain residual matrix and render the cross-domain residual distribution map. This allows for quantitative identification of whether the current state of the target entity deviates from the expected state in the current time period from the result level. This not only obtains the comprehensive evaluation results, but also locates the abnormal dimensions that cause the deviation and their main source scenarios, improving the interpretability and traceability of the analysis results. By retaining the component contributions of each scenario to the same dimension, it can still identify the source of internal fluctuations when the total residual is close to zero, thereby avoiding misjudgments caused by the mutual cancellation of different scenarios.

[0019] 4. This invention, by setting up fault-tolerant processing mechanisms such as pending confirmation queues, neutral vectors, group mean completion, recent effective expected vector rollback, degradation actions, and manual review in situations such as entity identifier conflicts, insufficient historical data, map mismatch, model unavailability, execution limitations, or push failures, can avoid error bias propagation or closed-loop interruption caused by local data loss, device offline, or module failure. This improves the robustness, maintainability, and engineering implementation capability of the entire educational data collaborative processing method and system in real deployment environments. Attached Figure Description

[0020] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0021] Figure 1 This is a flowchart illustrating the collaborative processing method for educational data interconnection across multiple scenarios involving schools, bureaus, and families, as provided in an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of a module of an educational data collaborative processing system for multi-scenario interconnection between government, schools, and families, provided in an embodiment of this application. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0024] A collaborative processing method for educational data across multiple scenarios involving schools, bureaus, and families includes the following steps: Constructing a multi-dimensional orthogonal state space and initializing baseline state vectors for target entities within this space; acquiring raw interaction events generated by multiple scenario terminals on the data acquisition side, parsing entity identifiers in the raw interaction events to match corresponding target entities, and transforming the raw interaction events into state bias vectors using a pre-defined domain knowledge graph to achieve semantic state synchronization across domain scenarios; extracting the state bias vectors corresponding to the target entities within the same time window, performing multi-dimensional state vector fitting on the baseline state vectors and state bias vectors in each independent orthogonal dimension, and calculating the currently accumulated actual state vector; comparing the actual state vectors with the expected vectors, performing cross-domain residual distribution calculation, and generating a cross-domain residual distribution map; wherein the expected vector is predicted by inputting the baseline state vectors and the target timestamp corresponding to the current time window into a pre-trained comprehensive evaluation model; generating collaborative intervention instructions based on the cross-domain residual distribution map, and issuing the collaborative intervention instructions to the scenario execution end on the control side for closed-loop control.

[0025] This embodiment provides a collaborative processing mechanism for educational data across multiple scenarios involving schools, bureaus, and families, such as... Figure 1 As shown; specifically, taking the regional education collaboration platform deployed by a district education bureau as the main scenario, the platform simultaneously connects the school canteen management system, campus access control and sports data collection equipment, parents' mobile terminals, and the regional comprehensive evaluation platform, and uses student entities as the target entities for continuous status tracking.

[0026] For ease of explanation, the following analysis focuses on Student A's state processing over three consecutive hours. A multidimensional orthogonal state space is constructed on the platform side. This space can be set to five independent dimensions, such as physical dimension, cognitive dimension, psychological dimension, daily routine dimension, and consumption risk dimension. Each independent orthogonal dimension in the multidimensional orthogonal state space maps to non-overlapping business semantics and is processed according to independent coordinate axes during calculation to reduce direct semantic conflicts between data from different sources.

[0027] Based on student A's process evaluation data from the previous semester, the platform determines their initial coordinates, for example, setting their baseline state vector as [0.62, 0.71, 0.68, 0.75, 0.30]. The value corresponding to this dimension is positively correlated with the degree of closeness to the expected health range. The consumption risk dimension can be defined in reverse as a higher value indicating a higher risk. Therefore, positive and negative directions can be uniformly agreed upon in the subsequent graph mapping stage.

[0028] The data collection side continuously receives raw interaction events from terminals in multiple scenarios; the raw event generated by the school cafeteria system could be 12:03, when student A makes a purchase at the high-sugar beverage window; the raw event generated by the home terminal could be that the student slept for less than 6 hours the previous night; the raw event generated by the converged media system could be that a parent spends 180 seconds on a psychological counseling topic; the platform first parses the student ID, device token, or family-bound account from the raw events and maps them uniformly to the target entity, student A.

[0029] The semantic synchronization unit, combined with the domain knowledge graph, transforms each event into a state bias vector. For example, the event of high sugar consumption in the cafeteria is mapped to [physical condition -0.08, cognition 0, psychology 0, rest 0, consumption risk +0.05]; the event of insufficient sleep is mapped to [physical condition -0.03, cognition -0.04, psychology -0.02, rest -0.12, consumption risk 0]; and the event of parents reading psychological counseling content is mapped to [physical condition 0, cognition 0, psychology +0.03, rest 0, consumption risk 0]. In this way, heterogeneous original records from different systems are uniformly transformed into homogeneous state bias vectors, completing the semantic state synchronization across domain scenarios.

[0030] Within the same time window, for example, from 12:00 to 13:00 on a given day, the platform extracts all state bias vectors corresponding to Student A within that window from the collaborative bus and performs cumulative fitting according to each dimension. If there are three bias vectors within the window, and the attenuation weights are 1, 0.8, and 0.9 respectively, then the cumulative bias vectors can be calculated as follows: the first one multiplied by 1 yields [-0.08, 0, 0, 0, +0.05]; the second one multiplied by 0.8 yields [-0.024, -0...]. [0.032, -0.016, -0.096, 0]; Multiply the third by 0.9 to get [0, 0, +0.027, 0, 0]; Summing the three by dimension yields the cumulative bias vector [-0.104, -0.032, +0.011, -0.096, +0.05]; Add this cumulative bias vector to the baseline state vector to get the current cumulative actual state vector: [0.516, 0.678, 0.691, 0.654, 0.35].

[0031] The platform inputs the baseline state vector and the current target timestamp into the comprehensive evaluation model to obtain the expected vector under the current time window; assuming the model output is [0.60, 0.70, 0.72, 0.76, 0.28]; the middle platform calculates the difference between the actual state vector and the expected vector to obtain the cross-domain residual vector: [-0.084, -0.022, -0.029, -0.106, +0.07].

[0032] To facilitate scheduling, the residual vector can be further mapped into a two-dimensional distribution map. For example, the horizontal axis represents the source of the business scenario, the vertical axis represents the state dimension, and the color intensity represents the magnitude of the absolute value of the residual, thus forming a cross-domain residual distribution map. This distribution map shows that the absolute values ​​of the residuals in the rest and physical dimensions exceed the preset deviation evaluation threshold, and also shows that their main source is the superposition of the home sleep scenario and the school cafeteria scenario.

[0033] Based on the residual distribution map, the platform generates collaborative intervention instructions and sends them to the control-side scenario execution end to form a closed-loop control. For example, when the system determines that the work-rest dimension and body shape dimension are both below the corresponding tolerance range, it can simultaneously push regular work-rest reminders to the family end, issue adjustment instructions to the canteen system to limit high sugar recommendations, and convert subsequent new event flows into state bias vectors again to enter the next time window for continued calculation, thus forming a cycle of perception-projection-fitting-residual analysis-control-re-perception.

[0034] As an anomaly handling mechanism, if multiple candidate entities are found to match the same event during entity identifier resolution, a bias vector is not generated temporarily. Instead, the event is placed in the confirmation queue and a secondary determination is made based on the most recent trusted login relationship, terminal binding relationship, and temporal proximity relationship. If a determination still cannot be made, the event is treated as an unmatched entity event, retaining only anonymous statistics without participating in individual state fitting. If no event is collected within a certain time window, the cumulative bias vector is set to zero, and only the baseline state vector is compared with the expected vector to avoid fluctuations in the state calculation results outside the preset range. If the comprehensive evaluation model is unavailable, the platform can fall back to the most recent valid expected vector or use temporal proximity interpolation results for temporary calculation.

[0035] For example, in a certain district's cloud-school-family integrated platform, student A uploaded a sleep deprivation event to their family terminal the night before, the school cafeteria recorded the consumption of sugary drinks at noon, and the parents also read psychological counseling content on the same day. The platform does not upload raw data across network segments, but only uploads state bias vectors to each terminal. The middleware quickly fits the cumulative deviation of student A in the dimensions of daily routine and physical posture, and issues intervention actions to the cafeteria system and the parents' terminal respectively. The purpose of this step is to replace the traditional detailed data aggregation mechanism with a unified state vector mechanism, while retaining cross-domain evaluation capabilities and reducing the risks of cross-system data coupling, data transmission volume, and privacy exposure, thereby achieving second-level educational collaborative processing and closed-loop intervention.

[0036] Furthermore, a multidimensional orthogonal state space is constructed, and a baseline state vector is initialized for the target entity in the multidimensional orthogonal state space. This includes: extracting business dimensions from macro-management scenarios, teaching and logistics scenarios, and family routine scenarios, and mapping the business dimensions to coordinate axes in the multidimensional orthogonal state space; obtaining historical evaluation features of the target entity, and establishing initial coordinate points in the multidimensional orthogonal state space based on the historical evaluation features; converting the initial coordinate points into multidimensional feature vectors, and confirming the multidimensional feature vectors as the baseline state vectors.

[0037] This embodiment provides a mechanism for initializing a multidimensional orthogonal state space. Specifically, in the aforementioned regional education collaboration platform, if the data from schools, families, and education bureaus are simply labeled and spliced ​​together, the same indicator may have inconsistent meanings in different systems. For example, school arrival time reflects attendance in a school setting, but is closer to daily routines in a family setting. Simply merging these data will lead to distortion of the input data for the evaluation model. Therefore, this embodiment first establishes a unified state space by extracting business dimensions and mapping coordinate axes, and then generates baseline state vectors for each target entity in this space.

[0038] Specifically, the platform first extracts relevant dimensions for regional comprehensive evaluation from macro-management scenarios, such as health level, psychological risk trends, and abnormal consumption warnings; from teaching and logistics scenarios, it extracts attendance stability, sports participation, and dietary structure rationality; from family routine scenarios, it extracts features such as sleep regularity and family companionship stability; and maps these business dimensions onto a unified coordinate axis. To avoid dimension redundancy, similar dimensions can be merged through a relevance deduplication mechanism. For example, the frequency of high sugar intake in the canteen and the degree of dietary structure imbalance are uniformly incorporated into the body shape dimension, and the fluctuations in homework completion time and sleep onset time are uniformly incorporated into the routine dimension.

[0039] The platform reads the historical evaluation features of the target entity to establish initial coordinate points. Taking student A as an example, in the historical data of the most recent 90 days, the average sleep compliance rate is 0.75, the physical test compliance rate is 0.64, the psychological questionnaire stability is 0.68, the classroom cognitive engagement is 0.71, and the abnormal consumption index is 0.30. The platform projects these features onto the corresponding coordinate axes according to the preset mapping rules to obtain the initial coordinate points.

[0040] If a dimension is derived from multiple historical features, a weighted average can be used to generate a single coordinate value. For example, the body shape dimension can be formed by the pass rate of physical fitness tests and the balance of diet. Assuming that the historical pass rate of physical fitness tests is 0.64 and the weight is set to 0.6, and the historical balance of diet is 0.59 and the weight is set to 0.4, then the initial value of the body shape dimension is 0.6×0.64+0.4×0.59=0.620, and the calculation result corresponds precisely to the preset baseline setting value of 0.62.

[0041] The platform transforms the initial coordinate point into a multi-dimensional feature vector, which is then written into the entity state database as a baseline state vector. This vector serves as both the starting point for the subsequent superposition of state bias vectors and the input basis for the comprehensive evaluation model to calculate the expected vector. To ensure comparability across different time periods, the platform can also normalize each dimension, for example, by uniformly mapping it to the 0 to 1 range, where 0 indicates that the point is not in the healthy range and 1 indicates that it is in the expected range.

[0042] As an anomaly handling mechanism, if a target entity lacks historical data, such as a newly transferred student having only 7 days of records, the platform can use the average of the same age and class group as the initial completion value, and then combine it with the collected short-term real data to form a mixed initial value; if a dimension is completely missing, such as the family not authorizing the upload of daily routine data, the dimension can be set to a neutral value of 0.5, and marked as low confidence in the metadata, so that the subsequent evaluation model can reduce its influence weight; if the data difference between two candidate dimensions is found to be greater than the preset conflict threshold when deduplicating business dimensions, the dimension with higher preset confidence is retained as the main axis, and the other dimension is converted into a correction item.

[0043] For example, in the same regional deployment, after the education bureau maps the regional health assessment, school logistics and family routines to a five-dimensional state space, a baseline vector [0.62, 0.71, 0.68, 0.75, 0.30] is generated for student A; subsequently, no matter what events are uploaded by the canteen, access control or family terminal, the original historical records will no longer be directly rewritten, but will evolve incrementally on the baseline.

[0044] The purpose of this step is to first establish a unified state coordinate framework that is computable, superimposed, and comparable, so as to achieve semantic consistency when fitting cross-domain data later.

[0045] Furthermore, the system acquires raw interaction events generated by multiple scenarios from the data acquisition side, and transforms these raw interaction events into state bias vectors using a pre-defined domain knowledge graph. This process includes: capturing raw interaction events generated by multiple scenarios in real time on the edge processing device; inputting the raw interaction events into the domain knowledge graph for semantic feature extraction, and outputting state micro-change values ​​for the corresponding dimensions. The domain knowledge graph at least includes entities and relationship networks related to educational routines, nutritional health indicators, and psychological cognitive development; vectorizing and encapsulating the state micro-change values, and attaching event timestamps and scenario tags to generate state bias vectors; and asynchronously pushing the state bias vectors to a pre-defined collaborative bus.

[0046] This embodiment provides a mechanism for transforming raw interactive events into state bias vectors. Specifically, in the aforementioned scheme, directly uploading cafeteria consumption details, family daily routine logs, and multimedia click logs would not only bring network load and privacy pressure, but also cause the middleware to undertake complex heterogeneous parsing work. Therefore, this embodiment forwards semantic extraction to edge processing devices and completes the real-time mapping of raw events to minute state change values ​​through domain knowledge graphs.

[0047] Specifically, edge processing devices can be deployed on the school's local node and the home terminal side respectively; the school's local node monitors events such as cafeteria consumption records, access control card swipes, and sports equipment tracking in real time; the home terminal monitors the daily routine records, study companionship records, and content browsing records confirmed by parents; for each original event, the edge processing device first extracts basic fields, such as entity identifier, event type, timestamp, scene tag, and necessary values.

[0048] The event is fed into a domain knowledge graph; this knowledge graph contains at least three types of entities and their relationship networks: the daily routine entity, the nutrition and health entity, and the psychological and cognitive development entity; for example, there is a negative relationship between high-sugar drinks and body shape fluctuations, a negative relationship between consecutive late nights and decreased cognitive input the next day, and a positive relationship between effective reading of family education guidance content and enhanced psychological support.

[0049] For example, the pre-defined mapping rules within the knowledge graph include: First rule, if the event type is consumption of high-sugar drinks and occurs during lunchtime, the slight change value for the body shape dimension is -0.08, and the slight change value for the consumption risk dimension is +0.05; Second rule, if the sleep deprivation duration exceeds 90 minutes, the slight change value for the sleep schedule dimension is -0.12, and the slight change value for the cognition dimension is -0.04; Third rule, if parents effectively read psychological guidance content for more than 120 seconds, the slight change value for the psychological dimension is +0.03.

[0050] The edge device outputs the corresponding dimension's minute state change value based on this, and then encapsulates it according to a unified vector format, such as entity identifier + timestamp + scene label + five-dimensional change value; if student A consumes a sugary drink at 12:03, a state bias vector <student A, 12:03, cafeteria scene, [-0.08, 0, 0, 0, +0.05]> is generated; this vector is pushed to the collaboration bus in an asynchronous message manner, without carrying specific dishes, amounts, or other detailed fields.

[0051] As an anomaly handling mechanism, if the original event cannot be matched with an existing node in the knowledge graph, such as the emergence of a new food category or a new family behavior tag, the edge device can first perform fuzzy classification, such as mapping to the closest known node based on nutritional or behavioral attributes. If the similarity is lower than a preset threshold, the event is transferred to the knowledge graph queue to be expanded, and only zero vectors or low-confidence vectors are output to prevent error propagation. If the edge device is offline for a short period of time, the state bias vector is cached locally and re-uploaded in chronological order. If the cache exceeds the limit, the vector corresponding to the abnormal event is retained first, and regular stable events with an overall impact less than a preset impact threshold are discarded.

[0052] For example, after student A consumes lunch, the school's edge node does not upload the original transaction record of purchasing one bottle of a sugary drink of a certain brand for 4 yuan. Instead, it directly forms a five-dimensional bias vector with negative physical posture and positive consumption risk based on the knowledge graph and writes it into the collaborative bus. At night, the home terminal records that the student is sleep-deprived, and only forms negative vectors of rest and cognition before uploading them.

[0053] The purpose of this mechanism is to compress the original heterogeneous events into directly computable semantic increments, thereby achieving low-load transmission across network segments and privacy-controlled semantic state synchronization.

[0054] Furthermore, within the same time window, the state bias vector corresponding to the target entity is extracted. Multidimensional state vector fitting is performed on the baseline state vector and the state bias vector in each independent orthogonal dimension to calculate the current cumulative actual state vector. This includes: extracting the state bias vector from the preset cooperative bus; calculating the decay weight based on the event timestamp carried by the state bias vector, and performing a weighted summation operation on the state bias vectors belonging to the same dimension to obtain the cumulative bias vector; and performing vector addition fitting on the baseline state vector and the cumulative bias vector to output the actual state vector.

[0055] This embodiment provides a mechanism for fitting state bias vectors. Specifically, in the aforementioned scheme, if all bias vectors are simply summed, earlier events will have the same impact as the most recently occurred events, which cannot reflect the characteristics of educational scenarios where near-time events correspond to higher weight parameters and far-time events have gradually decreasing weights. Therefore, this embodiment introduces a timestamp-based decay weight and performs weighted summation on each independent dimension to obtain an actual state vector that better reflects the current state.

[0056] Specifically, at the beginning of each time window, the middle platform extracts all state bias vectors within that window from the collaborative bus according to the entity identifier. Assuming that for student A, three vectors are extracted between 12:00 and 13:00: the first one is at 12:03, from the cafeteria scenario, with a value of [-0.08, 0, 0, 0, +0.05]; the second one is at 12:20, from the family supplementary transmission scenario, with a value of [-0.03, -0.04, -0.02, -0.12, 0]; and the third one is at 12:50, from the parent content reading scenario, with a value of [0, 0, +0.03, 0, 0].

[0057] If the current window ends at 13:00, the decay weight can be calculated according to the rule that the closer to the end of the window, the higher the weight. To avoid introducing implicit calculation functions, the platform uses a deterministic linear mapping rule based on the time span to calculate the decay weight; specifically, it is based on the start time of the current time window. Total duration of the time window Timestamp of the event and the preset effective weight range [ , The decay weight of the event. The computational logic can be expressed as: .

[0058] For N state bias vector values ​​belonging to the same dimension The corresponding dimension value of the cumulative bias vector Obtained through the weighted summation formula: ;in, For the summation operator, i is the index number of the state bias vector in the same dimension, with values ​​of 1, 2, ..., N. The above mathematical expression improves the interpretability of the underlying calculation mechanism and avoids the defects caused by implicit operation logic. In the example, for ease of deduction, the first event occurring at 12:03 has a time difference of less than the window start time, and its attenuation weight based on mapping is 0.7; the second event occurring at 12:20 has a weight of 0.8; and the third event occurring at 12:50 is close to the end of the window and has a weight of 0.95.

[0059] Therefore, the weighted vectors become: [-0.056, 0, 0, 0, +0.035], [-0.024, -0.032, -0.016, -0.096, 0], and [0, 0, +0.0285, 0, 0], respectively. After summing by dimension, the cumulative bias vector is [-0.08, -0.032, +0.0125, -0.096, +0.035]. This is then added to the baseline state vector [0.62, 0.71, 0.68, 0.75, 0.30] to obtain the actual state vector [0.54, 0.678, 0.6925, 0.654, 0.335].

[0060] The additive fitting here can be truncated according to business constraints; for example, if the result of a certain dimension is less than 0, it is set to 0; if it is greater than 1, it is set to 1; if the result of a certain dimension is between 0 and 1, the fitting result remains unchanged; this preserves the incremental evolution logic and prevents the state value from going out of bounds; if necessary, nonlinear fitting can also be introduced for some dimensions, such as using a saturation function for the consumption risk dimension, so that multiple abnormal consumptions in a short period of time will not be linearly amplified to an unreasonable level.

[0061] As a backup anomaly handling mechanism, if conflicting bias vectors appear within the same window, such as positive body shape increments from school exercise check-ins and negative body shape increments from high-sugar consumption in the cafeteria, the system will still retain the combined influence of both by weighted summation without manual overwriting. If a vector's timestamp is abnormally earlier than the baseline time or later than the upper bound of the current window, the vector is judged as a timing anomaly, will not participate in this round of fitting, and will be written to the audit log. If the number of bias vectors of an entity within a certain window exceeds a preset threshold, such as exceeding 500, homogeneous aggregation can be performed on the edge side first, and then the aggregated bias vectors can be uploaded to avoid bus congestion.

[0062] For example, after student A continuously generates lunch consumption, overnight sleep supplement, and parent reading behavior, the middle platform does not simply count the events, but performs attenuation weighting based on the relationship between the three vectors and the time proximity, and finally obtains the actual state vector that is closer to 13:00.

[0063] The purpose of this mechanism is to smooth the discrete event stream into a continuously evolving current state expression, thereby achieving unified, stable and real-time state fitting for cross-domain events occurring at different times.

[0064] Furthermore, the actual state vector is compared with the expected vector, and cross-domain residual distribution calculation is performed to generate a cross-domain residual distribution map. This includes: inputting the baseline state vector and the target timestamp into the comprehensive evaluation model to obtain the expected vector output by the comprehensive evaluation model; calculating the difference between the actual state vector and the expected vector in each independent orthogonal dimension, and splitting the difference according to the scene label carried by the state bias vector to obtain the cross-domain residual matrix; performing spatial mapping on the cross-domain residual matrix and rendering to generate a cross-domain residual distribution map, wherein the cross-domain residual distribution map is used to locate the abnormal dimensions that cause the overall evaluation to shift.

[0065] This embodiment provides a cross-domain residual distribution calculation mechanism. Specifically, in the aforementioned state fitting, if only the current actual state vector itself is observed, the platform can only obtain the current state evaluation value, but it is difficult to determine whether the target entity deviates from the expected state of the current time period. Therefore, this embodiment introduces the expected vector and residual matrix to identify the specific abnormal dimensions and source scenarios that cause the overall evaluation to deviate.

[0066] Specifically, the platform first inputs the baseline state vector and the current target timestamp into the comprehensive evaluation model. This model can combine the school stage patterns, school calendar rhythm, historical evolution trajectory, and regional group statistics to obtain the expected state at that moment. To ensure the verifiability of the comprehensive evaluation's deduction logic, the comprehensive evaluation model adopts a deterministic structured computation mechanism with multi-factor incremental compensation instead of an unexplainable prediction mechanism: For the school calendar time period where the target timestamp, such as 13:00, is located, the model extracts the periodic fluctuation parameters of business expectations, such as the expected positive increment of the lunch break schedule; it extracts the statistical mean curve of the regional group of the same age group and calculates the group trend compensation amount at the current moment; finally, the two increments are weighted and summed with the credibility factor of the student's historical evolution trajectory to form a comprehensive expected compensation term, which is directly superimposed on the original baseline state vector.

[0067] In the specific structured decomposition, based on the baseline state vector B and the extracted periodic fluctuation compensation vector... Group trend compensation vector And the credibility factor of the established historical evolution trajectory for cyclical fluctuations. With the credibility factor of group trends ,in and All are dimensionless scalar parameters, and the expected vector E corresponding to the target timestamp can be calculated using a structured derivation formula: This method clearly and transparently calculates the expected result, avoiding the lack of interpretability in complex evaluation models. Taking student A at 13:00 as an example, after the above structured incremental compensation, the model outputs the expected vector [0.60, 0.70, 0.72, 0.76, 0.28]. Subtracting the actual state vector [0.54, 0.678, 0.6925, 0.654, 0.335] obtained in the previous embodiment yields the residual vector [-0.06, -0.022, -0.0275, -0.106, +0.055].

[0068] To further pinpoint the cross-domain sources, each bias vector can be split according to the source scenario and a cross-domain residual matrix can be constructed. For example, let the rows of the matrix represent the scenario source and the columns represent the state dimension. If the canteen scenario contributes -0.056 to body posture and +0.035 to consumption risk, the family routine scenario contributes -0.024 to body posture, -0.032 to cognition, -0.016 to psychology, and -0.096 to routine, and the converged media scenario contributes +0.0285 to psychology, then a 3×5 matrix can be obtained. The middle platform then correlates the contribution of each row with the total residual and renders a cross-domain residual distribution map. In the map, red can be used to represent negative deviation, blue can represent positive deviation, and color depth can represent the absolute value.

[0069] Through this spatial mapping, the platform not only finds that the daily routine dimension is low, but also that this low level is mainly contributed by the family's daily routine scenario; at the same time, the low body size is related to both the canteen consumption scenario and the family sleep scenario; for the education bureau's macro analysis, this graphical result helps to distinguish whether it is a school logistics problem, a family behavior problem, or a combination of the two.

[0070] As a backup anomaly handling mechanism, if the expected vector output by the comprehensive evaluation model has missing terms, such as the psychological dimension having no effective predicted value for the time being, the corresponding dimension can be excluded from the residual matrix of this round and marked with a gray blank in the distribution map; if a column in the residual matrix is ​​close to zero in all scenarios, it indicates that the dimension is running stably and its display priority can be reduced during rendering; if different scenarios produce positive and negative offsets on the same dimension, resulting in the total residual being close to zero, the platform can retain the scene component in the graph to avoid misjudging that the system has no fluctuations simply because the total amount is offset.

[0071] For example, in the analysis results of student A at 13:00, the distribution map generated by the platform shows that the red of the daily routine dimension is the darkest, mainly from the family's daily routine; the physical posture dimension is the second darkest, from the dual contributions of the cafeteria and home; although the psychological dimension has a slight negative bias caused by family sleep, it is offset by the positive incremental part generated by the parents' reading guidance content.

[0072] The purpose of this mechanism is to trace state anomalies from the outcome level to the cross-domain source level through the actual-expected residual relationship, thereby achieving a localizable and interpretable basis for collaborative intervention.

[0073] Furthermore, collaborative intervention instructions are generated based on the cross-domain residual distribution map and sent to the scene execution terminal on the control side for closed-loop control. This includes: extracting residual values ​​from the cross-domain residual distribution map and calculating the absolute value of the residuals; comparing the absolute value of the residuals with a preset tolerance threshold; if the absolute value of the residuals is higher than the tolerance threshold, it is determined that there is an abnormal deviation in the corresponding dimension, generating a first type of collaborative intervention instruction based on a preset intervention rule base, and sending the first type of collaborative intervention instruction to the corresponding scene execution terminal to perform action adjustment; if the absolute value of the residuals is equal to the tolerance threshold, it is determined that the corresponding dimension is in a critical state, generating a second type of collaborative intervention instruction, and sending the second type of collaborative intervention instruction to the corresponding scene execution terminal to perform high-frequency monitoring; if the absolute value of the residuals is lower than the tolerance threshold, it is determined that the corresponding dimension is operating normally, generating a third type of collaborative intervention instruction, and sending the third type of collaborative intervention instruction to the corresponding scene execution terminal to maintain the current operating state.

[0074] This embodiment provides a collaborative intervention control mechanism based on tolerance thresholds. Specifically, the aforementioned residual distribution map can already locate the source of anomalies, but if we still rely on manual analysis of the distribution map and then issue control measures item by item, it will cause processing delays and make it difficult to meet the needs of regional multi-student concurrent scenarios. Therefore, this embodiment achieves hierarchical automatic control through the linkage of tolerance thresholds and rule base.

[0075] Specifically, the platform extracts residual values ​​for each dimension from the cross-domain residual distribution map and calculates the absolute value of the residuals, then compares them with a preset tolerance threshold. The tolerance threshold can be set separately for each dimension. The preset tolerance threshold is calculated based on the variance of the historical residual distribution of the target entity's group in the target dimension. Specifically, it extracts the 95% confidence interval boundary value of the historical residuals of the same age group as the dynamic tolerance threshold for that dimension.

[0076] The pre-defined intervention rule base is constructed from the association features of educational experts' prior guidance strategies and historically high-frequency effective intervention actions. It maps the abnormal scenario sources in the residual matrix to the corresponding execution actions through a knowledge graph. For example, the threshold for the physical dimension is 0.05, the threshold for the daily routine dimension is 0.08, and the threshold for the psychological dimension is 0.06. Taking the aforementioned student A as an example, the absolute values ​​of his residuals are: physical 0.06, cognition 0.022, psychology 0.0275, daily routine 0.106, and consumption risk 0.055. Assuming that the threshold for the consumption risk dimension is 0.05, then physical physique, daily routine, and consumption risk all exceed the threshold.

[0077] When the absolute value of the residual exceeds the tolerance threshold, the platform determines that there is an abnormal deviation in that dimension and generates a first-type collaborative intervention instruction based on the intervention rule base. The rule base can establish a correspondence between dimension, scenario, and action. For example, if the daily routine dimension exceeds the threshold and the main source is the family scenario, a strong reminder for the family and a morning check-up instruction for the next day will be generated. If the body shape dimension exceeds the threshold and the main source is the cafeteria scenario, a cafeteria recommendation strategy adjustment instruction will be generated. If the consumption risk dimension exceeds the threshold, an enhanced abnormal consumption monitoring instruction will be generated.

[0078] When the absolute value of the residual equals the tolerance threshold, the platform determines that it is in a critical state and generates a second type of collaborative intervention instruction. Unlike direct action adjustment, this time it focuses more on high-frequency monitoring. For example, the status assessment originally once per hour is changed to once every 15 minutes, or the edge side is required to temporarily increase the frequency of relevant event uploads to confirm whether it continues to deviate from the expected state.

[0079] When the absolute value of the residual is lower than the tolerance threshold, the platform determines that the dimension is operating normally and generates a third type of collaborative intervention instruction. Its role is not to completely ignore the problem, but to explicitly notify the execution end to maintain the current operating parameters, so as to avoid each subsystem from making local adjustments that conflict with the middle platform strategy.

[0080] As a backup anomaly handling mechanism, if multiple dimensions exceed the threshold simultaneously and the corresponding actions conflict with each other, such as the body shape dimension requiring a reduction in high-calorie supply while another special nutrition dimension requires energy supplementation, the platform can arbitrate based on rule priority, the student's specific health record, or manual approval switch; if the fluctuation frequency of the residual value of a certain dimension is higher than the preset frequency threshold and the amplitude is within the preset tolerance range, a hysteresis interval can be introduced, for example, only if it exceeds the threshold by more than 0.01 will it be upgraded to the first type of instruction, and only if it is lower than the threshold by more than 0.01 will it fall back to the third type of instruction, in order to prevent frequent switching.

[0081] For example, in Student A's current window, the daily routine residual of 0.106 is higher than 0.08, the physical residual of 0.06 is higher than 0.05, and the consumption risk of 0.055 is higher than 0.05. The platform generates the first type of instruction respectively. If the cognitive and psychological dimensions are lower than the threshold, the third type of maintenance instruction is generated. If the physical residual in the next window is exactly equal to 0.05, the system does not immediately force an adjustment, but switches to the second type of high-frequency monitoring.

[0082] The purpose of this mechanism is to transform residual analysis results into hierarchically executable control commands, thereby achieving automatic closed-loop response in educational scenarios.

[0083] Furthermore, the scenario execution end includes the canteen management execution system and the converged media execution system. The first type of collaborative intervention instruction is issued to the corresponding scenario execution end to perform action adjustments, including: when the cross-domain residual distribution map indicates that the absolute value of the residual in the physical dimension is higher than the tolerance threshold, an instruction to adjust the nutritional ratio of the menu is issued to the canteen management execution system; when the cross-domain residual distribution map indicates that the absolute value of the residual in the psychological cognition dimension is higher than the tolerance threshold, an instruction to push precise content is issued to the converged media execution system; the vector change data generated after the execution of the collaborative intervention instruction is extracted, and the vector change data is recorded in the preset student comprehensive quality evaluation database as a basis for process evaluation.

[0084] This embodiment provides an intervention implementation mechanism for specific scenarios. Specifically, in the previous solution, if the collaborative intervention instruction is not associated with specific control parameters, such as only outputting attention to posture or attention to psychology, the execution end will find it difficult to form an operable action, and the closed loop will still be incomplete. Therefore, this embodiment further refines the first type of collaborative intervention instruction to the canteen management execution system and the converged media execution system, and writes the vector change data after execution back to the student comprehensive quality evaluation database.

[0085] Specifically, when the residual distribution map shows that the absolute value of the body dimension residual is higher than the tolerance threshold, the middle platform calls the canteen management execution system interface to issue a menu nutrition ratio adjustment instruction. This instruction is not a fixed replacement of a single dish name, but is described in the form of nutritional structure parameters. For example, in the next lunch cycle, the recommendation weight of high-sugar drinks in the recommended group of student A will be reduced from 0.8 to 0.2, and the weight of low-sugar drinks will be increased from 0.3 to 0.9. At the same time, high-fiber side dishes will be recommended at the meal preparation end. After receiving the instruction, the canteen system can apply it to the window recommendation screen, the campus meal ordering entrance, or the meal preparation engine.

[0086] When the residual distribution map shows that the absolute value of the residual in the psychological cognition dimension is higher than the tolerance threshold, the middle platform issues a precise content push instruction to the converged media execution system. This instruction can specify the content theme, the target audience, and the effective time period. For example, it can push content on regular sleep and concentration recovery to student A's parent account from 18:00 to 20:00, with the push priority set to high. If the content is not read for two consecutive days, a reminder will be triggered again. In this way, the system does not broadcast indiscriminately, but transforms the cross-domain analysis results into targeted content intervention.

[0087] After the execution end completes the action, the platform continues to collect new event streams generated as a result. For example, after the cafeteria system implements adjustments, the lunch consumption event the next day may be mapped as body shape +0.02 and consumption risk -0.01. After parents read and complete the interactive questionnaire, it may generate psychological status +0.04 and daily routine +0.02. The platform records these new vector change data into the student comprehensive quality evaluation database. The database can store fields such as pre-intervention state vector, intervention action type, post-intervention state bias vector, timestamp, and source system for subsequent process evaluation and rule optimization.

[0088] As an exception handling mechanism, if the cafeteria system cannot execute all nutritional ratio adjustments due to inventory limitations, it can send back the restricted information. The middle platform will then generate a degradation action based on the executable items, such as adjusting only the recommended order without adjusting the meal structure. If the integrated media system fails to push notifications, for example, if the parents are offline or refuse to receive the notifications, the platform can automatically switch to SMS, homeroom teacher messages, or scheduled pushes the next day. If no positive vector change is observed after execution, the platform can upgrade to a manual review task after two consecutive windows.

[0089] For example, in the case of student A, the platform generates two actions in conjunction with student A's physical condition and daily routine: first, it sends a low-sugar alternative recommendation instruction to the cafeteria system, and second, it pushes sleep management and family companionship guidance content to the parents' media system. After execution, the cafeteria's consumption structure improves the next day, and the family completes the relevant learning content. These changes are then mapped back to the database as new bias vectors, becoming part of the student's process-based growth trajectory.

[0090] The purpose of this mechanism is to transform abstract state deviations into concrete business actions that can be executed, recorded, and traced, thereby achieving the true implementation of the closed loop of educational intervention.

[0091] A collaborative educational data processing system for multi-scenario interconnection between schools, bureaus, and families includes: a spatial modeling unit, used to construct a multi-dimensional orthogonal state space and initialize baseline state vectors for target entities in the multi-dimensional orthogonal state space; a semantic synchronization unit, used to acquire raw interaction events generated by multi-scenario terminals on the data acquisition side, parse entity identifiers in the raw interaction events to match corresponding target entities, and transform the raw interaction events into state bias vectors through a preset domain knowledge graph to complete semantic state synchronization across domain scenarios; and a vector fitting unit, used to divide continuous time windows according to a preset time series and extract target entities within the same time window. The state bias vector corresponding to the body is fitted with the baseline state vector and the state bias vector in each independent orthogonal dimension to calculate the current accumulated actual state vector; the residual calculation unit is used to compare the actual state vector with the expected vector, perform cross-domain residual distribution calculation, and generate a cross-domain residual distribution map; the expected vector is predicted by inputting the baseline state vector and the target timestamp corresponding to the current time window into the pre-trained comprehensive evaluation model; the closed-loop control unit is used to generate collaborative intervention instructions based on the cross-domain residual distribution map and send the collaborative intervention instructions to the scene execution end on the control side for closed-loop control.

[0092] This embodiment provides an educational data collaborative processing system for multi-scenario interconnection between government, schools, and families, such as... Figure 2 As shown; specifically, the system can be deployed as a collaborative architecture of regional central node + school edge node + home terminal, where each functional unit can be implemented by a logical module, or by a combination of independent program components, container services, or devices with processing capabilities.

[0093] Specifically, the spatial modeling unit is responsible for maintaining a unified state space model and entity baseline library. This unit receives evaluation indicators from the education bureau, school logistics indicators, and family routine indicators, generates a unified coordinate system, and writes corresponding baseline state vectors for each student, class, or school entity.

[0094] The semantic synchronization unit is responsible for connecting the edge input, completing the original event parsing, entity matching and knowledge graph projection, outputting the standardized state bias vector and sending it into the collaborative bus; the vector fitting unit extracts the vector flow from the bus by window, performs time decay, dimension weighting and vector addition to form the actual state vector.

[0095] The residual calculation unit calls the comprehensive evaluation model to obtain the expected vector and generates the residual matrix and residual distribution map.

[0096] The closed-loop control unit reads the residual distribution map, generates hierarchical collaborative intervention instructions based on thresholds and rule bases, and pushes them to the canteen management execution system, the converged media execution system, or other execution terminals;

[0097] To illustrate the collaborative relationship between the units, the following sandbox process can be used: When the canteen edge node sends a body posture bias vector of -0.08 and the home terminal sends a daily routine bias vector of -0.12, the semantic synchronization unit first writes the two vectors into the bus; the vector fitting unit extracts and fits the actual state vector in the 13:00 time window.

[0098] The residual calculation unit calculates the difference between the residual and the expected vector and finds that the rest and body posture exceed the threshold; the closed-loop control unit then issues action instructions to the canteen system and the converged media system respectively; the units are connected through message bus or service interface to form a complete link.

[0099] As an anomaly handling mechanism, if a unit is temporarily unavailable, the system can adopt a tiered degradation strategy. For example, if the residual calculation unit fails, the vector fitting unit can still continue to produce actual state vectors and temporarily store them, and recalculate the residuals after the model recovers. If the closed-loop control unit fails, the system can at least retain the distribution map and alarm results for manual takeover. If the spatial modeling unit updates the coordinate axis definition and becomes incompatible with the old data, the version number mechanism can be used to ensure that the old vectors are calculated according to the old model and the new vectors are calculated according to the new model, avoiding mixed calculations.

[0100] For example, in the deployment of a cloud-based school-home platform in a certain district, the central node integrates spatial modeling units, vector fitting units, residual calculation units, and closed-loop control units, while the school's local server and the home's mobile terminal deploy semantic synchronization capabilities; after the various units process the multiple events of student A on the same day, a complete closed loop is achieved from edge acquisition to mid-platform analysis and then to execution-end actions.

[0101] The purpose of this system is to support the aforementioned methods and processes in a modular manner, so that cross-domain collaborative processing of educational data has a deployable, scalable and maintainable engineering implementation foundation.

[0102] Furthermore, the semantic synchronization unit is deployed in the edge processing device. Specifically, the semantic synchronization unit is used to: capture the original interaction events generated by terminals in multiple scenarios in real time; input the original interaction events into the domain knowledge graph for semantic feature extraction, and output the state micro-change values ​​of the corresponding dimensions. The domain knowledge graph at least includes entities and relationship networks of educational routines, nutritional health indicators and psychological cognitive development; encapsulate the state micro-change values ​​into vectors and attach event timestamps and scene tags to generate state bias vectors; and asynchronously push the state bias vectors to the preset collaborative bus.

[0103] This embodiment provides a semantic synchronization unit implementation method deployed in an edge processing device. Specifically, if the entire semantic synchronization process is completed at the central node, the original data needs to be transmitted centrally across network segments, which is not conducive to privacy control and will also increase the parsing burden on the central node. Therefore, this embodiment pushes the unit down to the school's local server, the canteen's edge gateway, and home terminals.

[0104] Specifically, the semantic synchronization unit deployed in the canteen edge gateway can directly monitor the logs of the consumer sales terminal, smart weighing equipment, and window recommendation screen; the semantic synchronization unit deployed in the home terminal can receive sleep, companionship, and content reading events confirmed by parents; each edge device stores a lightweight copy of the domain knowledge graph or a graph rule cache for local semantic extraction.

[0105] Taking the cafeteria gateway as an example, when a sales terminal event occurs where student A purchases a high-sugar beverage at 12:03, the local unit immediately matches the graph path of high-sugar beverage - negative body posture - positive consumption risk, and outputs a slight change in status value; then, it adds a timestamp and cafeteria scene tag, encapsulates it into a bias vector, and pushes it asynchronously to the bus; because asynchronous push is used, the edge device does not need to wait for the middle platform to confirm before it can continue to process the next event, thereby ensuring the throughput capacity under high concurrency.

[0106] The processing logic of the home terminal is similar; when a parent enters the actual bedtime at 22:30 that is 90 minutes later than the recommended time, the terminal can obtain the negative micro-change value of the sleep schedule and cognition locally, and report it in the form of a bias vector; if the terminal is not connected to the network, the vector can be stored in the local persistent queue first, and then asynchronously transmitted in time order after the network is restored.

[0107] As an anomaly handling mechanism, if the version of the graph copy in the edge device lags behind that of the central node, causing some new events to be unrecognizable, the edge device can send back a graph version insufficiency flag, and the central node will issue an incremental update package. Before the version update is completed, the device can output a neutral vector for such events and retain the original summary. If the push bus is blocked, the device can set priorities according to the importance of the events, such as ensuring that abnormal nutrition events and obvious abnormal work and rest events are pushed first.

[0108] For example, in the scenario of student A, the school cafeteria gateway and the home mobile terminal independently complete the local conversion of the original events to the bias vector, and then send the results in a unified format asynchronously to the collaborative bus; the central node only receives the compressed semantic increment, and does not receive the student's detailed consumption list and complete family log; the purpose of this mechanism is to reduce the central transmission pressure through edge semantic processing, and to realize the collaborative mode of keeping the original details on the edge and projecting the local data to the cloud in the system structure.

[0109] Furthermore, the closed-loop control unit is specifically used for: extracting residual values ​​from the cross-domain residual distribution map and calculating the absolute value of the residuals; comparing the absolute value of the residuals with a preset tolerance threshold; if the absolute value of the residuals is higher than the tolerance threshold, it is determined that there is an abnormal deviation in the corresponding dimension, generating a first type of collaborative intervention instruction based on the preset intervention rule base, and sending the first type of collaborative intervention instruction to the corresponding scene execution end to perform the action adjustment; if the absolute value of the residuals is equal to the tolerance threshold, it is determined that the corresponding dimension is in a critical state, generating a second type of collaborative intervention instruction, and sending the second type of collaborative intervention instruction to the corresponding scene execution end to perform high-frequency monitoring; if the absolute value of the residuals is lower than the tolerance threshold, it is determined that the corresponding dimension is operating normally, generating a third type of collaborative intervention instruction, and sending the third type of collaborative intervention instruction to the corresponding scene execution end to maintain the current operating state.

[0110] This embodiment provides an implementation method for a closed-loop control unit. Specifically, in the implementation of a regional platform, residual calculation results alone are insufficient to support real-time intervention. The results must be automatically converted into structured instruction objects so that they can be accurately parsed and executed by execution terminals in different scenarios. Therefore, this embodiment further explains the internal processing flow of the closed-loop control unit.

[0111] Specifically, the closed-loop control unit may include a residual reading module, a threshold comparison module, a rule matching module, and an instruction distribution module. The residual reading module reads dimension values ​​and source scenarios from the residual distribution map or residual matrix. The threshold comparison module calculates the state level based on the threshold corresponding to the dimension. The rule matching module searches for the corresponding template in the intervention rule base. The instruction distribution module encapsulates the results into standardized instruction messages and sends them to the execution end. The instruction message may include fields such as entity identifier, target dimension, action level, execution scenario, validity period, and receipt address.

[0112] For example, for student A's daily routine dimension, the absolute value of the residual is 0.106, while the threshold is 0.08. The threshold comparison module determines that it belongs to an abnormal deviation, and the rule matching module retrieves the template of abnormal family daily routine - first type of action - parent's strong reminder + homeroom teacher attention, and then generates the first type of collaborative intervention instruction. If the residual of posture in a certain analysis is exactly 0.05, which is equal to the threshold, the closed-loop control unit will generate the second type of instruction, requiring the canteen system to increase the collection frequency of the student's related dining behavior and enter the critical monitoring state. If the residual of the psychological dimension is only 0.02, which is lower than the threshold of 0.06, the third type of maintenance instruction will be output to keep the current push frequency of the converged media system unchanged.

[0113] This unit can also use different distribution protocols for different execution terminals; for example, an internal message interface can be used for the canteen management execution system, with the instruction carrying recommendation weight parameters; a content delivery interface can be used for the converged media execution system, with the instruction carrying topic tags and target parent accounts; and a to-do message method can be used for the homeroom teacher workbench, with the instruction carrying review prompts.

[0114] As an exception handling mechanism, if the action template corresponding to a certain residual pattern does not exist in the rule base, the closed-loop control unit can output a fallback instruction, that is, enter the manual review queue, while retaining the residual value and relevant scenario evidence; if the execution end fails to execute the receipt, the instruction distribution module can resend, switch to a backup channel, or report to manual processing according to a preset strategy; if the same entity receives multiple similar first-type instructions in a short period of time, the system can perform deduplication and merging, retaining only the most stringent one or aggregating the parameters.

[0115] For example, in the case of student A, the closed-loop control unit automatically generates recommended adjustment instructions for the cafeteria system, sleep guidance push instructions for the family media system, and attention to be done for the homeroom teacher based on the threshold results of body shape, daily routine and consumption risk; while for the psychological dimension that does not exceed the threshold, only the third type of instruction to maintain the current state is sent.

[0116] The purpose of this mechanism is to enable the mid-platform analysis results to have standardized control output capabilities, thereby achieving automatic closed-loop connection from residual judgment to execution action.

[0117] I. Experimental conditions and subjects:

[0118] The experimental subjects were: a regional education collaboration platform deployed by a district education bureau and its connected multi-scenario interconnected terminals between the bureau, schools, and families.

[0119] Experimental environment: Simulates macro-management scenarios, teaching and logistics scenarios, and family routine scenarios with high concurrency and continuous access to multi-source heterogeneous data; system integration includes school canteen management system, campus access control and sports data collection equipment, parent mobile terminals, and regional comprehensive evaluation platform.

[0120] Monitoring cycle and frequency: For multiple consecutive natural days and a complete learning cycle, the system synchronously collects and processes the raw interactive events generated by multi-scenario terminals captured in real time by the edge processing device within a continuous time window divided according to a preset time sequence.

[0121] II. Comparison Methods:

[0122] This invention group proposes a method that constructs a multidimensional orthogonal state space, transforms the original events into state bias vectors through a domain knowledge graph, fits the actual state vectors with attenuation weights and compares them with the expected vectors to generate a cross-domain residual distribution map, and finally issues closed-loop collaborative intervention instructions based on the residuals.

[0123] Control Group 1: The traditional method of centrally aggregating raw detailed data is used, and the data from different sources is directly aligned, semantically transformed, and the results are summarized on the central platform for comprehensive evaluation.

[0124] Control group 2: A semi-manual or isolated system was used, where each scenario (such as cafeteria consumption and daily routine records) directly generated early warnings based on the original events of single-dimensional indicators within its own system, and relied on manual interpretation and dissemination of the analysis results.

[0125] Experimental Design: A regional school-family interconnected network of equal size was selected and continuously operated under normal teaching and rest cycles, as well as under artificially introduced specific abnormal combinations (such as a sudden imbalance in the students' collective diet combined with irregular work and rest schedules). Status assessment and intervention were simultaneously conducted using the method of this invention, control group 1, and control group 2. The experiment continued until a sufficient amount of benchmark evaluation data and real cross-domain abnormal events were collected, and a massive amount of valid multi-scenario interaction samples were obtained for data comparison.

[0126] III. Evaluation Indicators:

[0127] Closed-loop collaborative intervention on-time rate: This reflects the proportion of the number of times the system successfully identifies cross-domain abnormal deviations and automatically issues collaborative intervention instructions to the scene execution end to execute closed-loop control within the corresponding time window under massive concurrent data access, out of the total number of actual abnormalities.

[0128] Cross-scenario semantic state synchronization latency: This reflects the average time required for the system to transform heterogeneous original events (such as cafeteria consumption and family routines) into uniform-dimensional state micro-change values ​​and complete synchronization.

[0129] Privacy Exposure and Transmission Bandwidth Utilization: This reflects the proportion of sensitive raw log data (such as cafeteria consumption details and family logs) directly exposed during cross-network segment transmission and the overall bandwidth burden.

[0130] Table 1: Comparison of the effects of different methods in collaborative processing of educational data

[0131] Table 2: Comparison of Identification and Handling Effects of Typical Multi-Scenario Complex Educational Anomalies

[0132] IV. Comparison with the implementation conclusions:

[0133] The method of this invention significantly improves the timeliness of closed-loop collaborative intervention. By capturing the original interactive events generated by multiple terminals in real time in edge processing devices and extracting the state micro-change values ​​of the corresponding dimensions in advance, the heterogeneous alignment dependency of the central platform is reduced, and second-level educational collaborative processing is achieved.

[0134] The method of this invention performs well in terms of privacy protection and network burden reduction. By only asynchronously pushing state bias vectors with timestamps and scene tags, without uploading the complete original pipeline, it significantly reduces the burden of cross-network segment transmission and the risk of direct exposure of sensitive data.

[0135] This invention innovatively constructs a multidimensional orthogonal state space and performs a weighted summation operation on each event within the same time window by combining attenuation weights. This successfully distinguishes the temporal influence intensity of events in different scenarios and improves the stability and interpretability of the actual state solution process.

[0136] This invention introduces an expected vector based on a comprehensive evaluation model. By calculating the difference between the actual state vector and the expected vector, a cross-domain residual distribution map is rendered. This not only quantifies whether the entity state deviates from the expectation, but also accurately locates the abnormal dimension and the source of the cross-domain scenario, providing a solid basis for automatically generating collaborative intervention instructions and issuing closed-loop control.

[0137] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. 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 be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A collaborative processing method for educational data across multiple interconnected scenarios involving schools, bureaus, and families, characterized in that: Includes the following steps: Construct a multidimensional orthogonal state space and initialize a baseline state vector for the target entity in the multidimensional orthogonal state space; The system acquires raw interaction events generated by multiple scenario terminals on the data acquisition side, parses the entity identifiers in the raw interaction events to match the corresponding target entities, and transforms the raw interaction events into state bias vectors through a preset domain knowledge graph to complete the semantic state synchronization of cross-domain scenarios. Divide the continuous time windows according to the preset time series, extract the state bias vector corresponding to the target entity within the same time window, and perform multi-dimensional state vector fitting on the baseline state vector and the state bias vector in each independent orthogonal dimension to calculate the current accumulated actual state vector. The actual state vector is compared with the expected vector, and cross-domain residual distribution calculation is performed to generate a cross-domain residual distribution map; wherein, the expected vector is obtained by inputting the baseline state vector and the target timestamp corresponding to the current time window into a preset comprehensive evaluation model for prediction; Based on the cross-domain residual distribution map, a collaborative intervention command is generated and sent to the scene execution terminal on the control side for closed-loop control. The construction of a multidimensional orthogonal state space and the initialization of a baseline state vector for the target entity in the multidimensional orthogonal state space include: Extract business dimensions from macro-management scenarios, teaching and logistics scenarios, and family routine scenarios, and map these business dimensions to coordinate axes of the multi-dimensional orthogonal state space; Obtain the historical evaluation features of the target entity, and establish initial coordinate points in the multidimensional orthogonal state space based on the historical evaluation features; The initial coordinate points are transformed into multi-dimensional feature vectors, and the multi-dimensional feature vectors are confirmed as the baseline state vectors. The step of comparing the actual state vector with the expected vector, performing cross-domain residual distribution calculation, and generating a cross-domain residual distribution map includes: Input the baseline state vector and the target timestamp into the comprehensive evaluation model to obtain the expected vector output by the comprehensive evaluation model; Calculate the difference between the actual state vector and the expected vector in each independent orthogonal dimension, and split the difference according to the scene label carried by the state bias vector to obtain the cross-domain residual matrix; The cross-domain residual matrix is ​​spatially mapped and rendered to generate the cross-domain residual distribution map, wherein the cross-domain residual distribution map is used to locate the abnormal dimensions that cause the overall evaluation to shift.

2. The method for collaborative processing of educational data in multiple interconnected scenarios between schools, bureaus, and families, as described in claim 1, is characterized in that... The process of acquiring raw interaction events generated by multi-scenario terminals on the data acquisition side, and transforming these raw interaction events into state bias vectors using a preset domain knowledge graph, includes: The original interactive events generated by the multi-scenario terminals are captured in real time in the edge processing device; The original interactive events are input into the domain knowledge graph for semantic feature extraction, and the corresponding dimension of state micro-change values ​​are output. The domain knowledge graph at least includes entities and relationship networks related to educational routines, nutritional health indicators and psychological cognitive development. The state micro-change values ​​are vectorized and encapsulated, and an event timestamp and scene label are attached to generate the state bias vector. The state bias vector is asynchronously pushed to a preset cooperative bus.

3. The educational data collaborative processing method for multi-scenario interconnection between schools, bureaus, and families as described in claim 2, is characterized in that... The process involves extracting the state bias vector corresponding to the target entity within the same time window, performing multi-dimensional state vector fitting on the baseline state vector and the state bias vector in each independent orthogonal dimension, and calculating the currently accumulated actual state vector, including: Extract the state bias vector from the preset cooperative bus; The decay weight is calculated based on the event timestamp carried by the state bias vector, and a weighted summation operation is performed on the state bias vectors belonging to the same dimension to obtain the cumulative bias vector. The baseline state vector and the cumulative bias vector are fitted by vector addition to output the actual state vector.

4. The method for collaborative processing of educational data for multi-scenario interconnection between schools, bureaus, and families according to claim 1, characterized in that, The process of generating collaborative intervention instructions based on the cross-domain residual distribution map and sending these instructions to the scene execution terminal on the control side for closed-loop control includes: Extract the residual values ​​from the cross-domain residual distribution map and calculate the absolute value of the residuals. Then compare the absolute value of the residuals with a preset tolerance threshold. If the absolute value of the residual is higher than the tolerance threshold, it is determined that there is an abnormal deviation in the corresponding dimension. A first type of collaborative intervention instruction is generated according to the preset intervention rule library, and the first type of collaborative intervention instruction is sent to the corresponding scene execution terminal to perform the action adjustment. If the absolute value of the residual is equal to the tolerance threshold, the corresponding dimension is determined to be in a critical state, a second type of collaborative intervention instruction is generated, and the second type of collaborative intervention instruction is sent to the corresponding scenario execution terminal to perform high-frequency monitoring; If the absolute value of the residual is lower than the tolerance threshold, the corresponding dimension is determined to be operating normally, a third type of collaborative intervention instruction is generated, and the third type of collaborative intervention instruction is sent to the corresponding scenario execution terminal to maintain the current operating state.

5. The method for collaborative processing of educational data for multi-scenario interconnection between schools, bureaus, and families according to claim 4, characterized in that, The scenario execution terminal includes a canteen management execution system and a converged media execution system. The step of issuing the first type of collaborative intervention instruction to the corresponding scenario execution terminal to adjust the execution action includes: When the absolute value of the residual in the body dimension indicated by the cross-domain residual distribution map is higher than the tolerance threshold, an instruction to adjust the nutritional ratio of the menu is issued to the canteen management execution system. When the cross-domain residual distribution map indicates that the absolute value of the residual in the psychological cognition dimension is higher than the tolerance threshold, a precise content push instruction is issued to the converged media execution system. Extract the vector change data generated after executing the collaborative intervention instruction, and record the vector change data in a preset student comprehensive quality evaluation database as a basis for process evaluation.

6. An educational data collaborative processing system for multi-scenario interconnection between schools, bureaus, and families, configured to execute the educational data collaborative processing method for multi-scenario interconnection between schools, bureaus, and families as described in any one of claims 1-5, characterized in that, include: A spatial modeling unit is used to construct a multidimensional orthogonal state space and initialize a baseline state vector for the target entity in the multidimensional orthogonal state space. The semantic synchronization unit is used to acquire the original interaction events generated by multi-scenario terminals on the data acquisition side, parse the entity identifiers in the original interaction events to match the corresponding target entities, and convert the original interaction events into state bias vectors through a preset domain knowledge graph to complete the semantic state synchronization of cross-domain scenarios. The vector fitting unit is used to divide a continuous time window according to a preset time series, extract the state bias vector corresponding to the target entity within the same time window, and perform multi-dimensional state vector fitting on the baseline state vector and the state bias vector in each independent orthogonal dimension to calculate the current accumulated actual state vector. The residual calculation unit is used to compare the actual state vector with the expected vector, perform cross-domain residual distribution calculation, and generate a cross-domain residual distribution map; wherein, the expected vector is obtained by inputting the baseline state vector and the target timestamp corresponding to the current time window into a preset comprehensive evaluation model for prediction; A closed-loop control unit is used to generate collaborative intervention instructions based on the cross-domain residual distribution map, and to send the collaborative intervention instructions to the scene execution terminal on the control side for closed-loop control.

7. The educational data collaborative processing system for multi-scenario interconnection between schools, bureaus, and families as described in claim 6, is characterized in that, The semantic synchronization unit is deployed in the edge processing device, and the semantic synchronization unit is specifically used for: Real-time capture of the original interactive events generated by the multi-scenario terminals; The original interactive events are input into the domain knowledge graph for semantic feature extraction, and the corresponding dimension of state micro-change values ​​are output. The domain knowledge graph at least includes entities and relationship networks related to educational routines, nutritional health indicators and psychological cognitive development. The state micro-change values ​​are vectorized and encapsulated, and an event timestamp and scene label are attached to generate the state bias vector. The state bias vector is asynchronously pushed to a preset cooperative bus.

8. The educational data collaborative processing system for multi-scenario interconnection between schools, bureaus, and families as described in claim 6, characterized in that, The closed-loop control unit is specifically used for: Extract the residual values ​​from the cross-domain residual distribution map and calculate the absolute value of the residuals. Then compare the absolute value of the residuals with a preset tolerance threshold. If the absolute value of the residual is higher than the tolerance threshold, it is determined that there is an abnormal deviation in the corresponding dimension. A first type of collaborative intervention instruction is generated according to the preset intervention rule library, and the first type of collaborative intervention instruction is sent to the corresponding scene execution terminal to perform the action adjustment. If the absolute value of the residual is equal to the tolerance threshold, the corresponding dimension is determined to be in a critical state, a second type of collaborative intervention instruction is generated, and the second type of collaborative intervention instruction is sent to the corresponding scenario execution terminal to perform high-frequency monitoring; If the absolute value of the residual is lower than the tolerance threshold, the corresponding dimension is determined to be operating normally, a third type of collaborative intervention instruction is generated, and the third type of collaborative intervention instruction is sent to the corresponding scenario execution terminal to maintain the current operating state.

Images