Remote education data intelligent scheduling and storage optimization system

Abstract

The invention relates to the field of data storage scheduling, and discloses a remote education data intelligent scheduling and storage optimization system, which comprises a data preprocessing and topology construction module; the heterogeneous storage medium layer comprises physical storage equipment with multiple performance levels; the initial storage strategy module is connected with the data preprocessing and topology construction module and the heterogeneous storage medium layer; a cognitive state monitoring module; the storage recombination controller is connected with the cognitive state monitoring module and the heterogeneous storage medium layer; and the distribution scheduling engine is used for executing data prefetching and executing cache replacement based on the review urgency of the cache data block. And variable-length semantic granulation segmentation is performed on the media stream through the data preprocessing and topology construction module, and a differentiated initial storage strategy is implemented according to the topology centrality of the knowledge points, so that the effect of dynamically allocating storage resources according to the knowledge structure value is achieved.

Description

Technical Field

[0001] This invention relates to the field of data storage and scheduling technology, specifically to an intelligent scheduling and storage optimization system for distance education data. Background Technology

[0002] With the improvement of network bandwidth and the popularization of online education platforms, massive amounts of teaching media data are being continuously generated and stored. Unlike the linear playback mode of traditional film and entertainment streaming media, user behavior in distance education scenarios has significant non-linear characteristics and cognitive-driven attributes. When watching teaching videos, learners often need to frequently perform actions such as rewinding for review, skipping searches, and repeatedly watching specific difficult points based on their own level of understanding and the logical dependencies between knowledge points. This access mode based on cognitive state poses a severe challenge to the read and write performance and data scheduling strategies of the underlying storage system.

[0003] Existing streaming media storage systems typically employ general-purpose distributed file systems or object storage architectures. Their data management granularity is primarily based on files or fixed-size physical fragments, lacking the ability to perceive the internal semantic structure and knowledge point relationships of teaching content. This separation between physical storage and logical semantics makes it difficult for the system to distinguish between core foundational knowledge and peripheral extension content when facing high-concurrency learning requests, often resulting in the use of a uniform erasure coding strategy and media allocation scheme. This not only wastes high-performance storage resources but also makes high-frequency hot data on low-speed media a bottleneck for the overall I / O performance of the system.

[0004] Furthermore, traditional data prefetching and cache replacement mechanisms primarily rely on linear extrapolation along the timeline and statistical algorithms based on historical access frequencies. However, learners' review behavior is governed by memory decay, exhibiting strong temporal jumps and periodicity. Existing technologies cannot predict users' need to review content due to forgetting, resulting in low data prefetching hit rates. When users make non-linear jumps based on the knowledge graph, the system often needs to initiate high-latency random read requests from the underlying disk. Moreover, logically related teaching content is often scattered in physical storage. This random I / O operation, combined with the decoding overhead of high-compression video encoding formats, severely impacts the smoothness and real-time performance of the learning experience. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent scheduling and storage optimization system for distance education data. This system solves the problems of resource allocation imbalance, low random read / write efficiency, and high data loading latency in existing streaming media storage systems when facing non-linear learning and review behaviors in distance education, due to the lack of perception of the topology of teaching knowledge and the forgetting patterns of users.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent scheduling and storage optimization system for distance education data, comprising:

[0007] The data preprocessing and topology construction module is used to receive the original teaching media stream, identify semantic boundaries through multimodal feature fusion, segment the original teaching media stream into several variable-length semantic particles, parse the knowledge points between each semantic particle, and construct a teaching topology graph based on logic, while calculating the topological centrality of each semantic particle in the teaching topology graph.

[0008] The heterogeneous storage medium layer includes physical storage devices with multiple performance levels, used to store the data blocks and corresponding check blocks of the semantic granules;

[0009] An initial storage strategy module, connected to the data preprocessing and topology construction module and the heterogeneous storage medium layer, is used to allocate erasure coding parameters and initial physical storage levels to each semantic particle according to the topology centrality.

[0010] The cognitive state monitoring module is used to record the user's historical interaction behavior data with the semantic particles, calculate the memory retention rate based on the historical interaction behavior data, calculate the predicted review time window based on the decay trend of the memory retention rate, and generate a set of unprocessed semantic particles containing the identifiers of the semantic particles that are about to enter the predicted review time window.

[0011] A storage reorganization controller, connected to the cognitive state monitoring module and the heterogeneous storage medium layer, is used to read the set to be processed when the system input / output load is lower than a preset idle threshold, and to control the heterogeneous storage medium layer to perform physical address remapping operations on the semantic particles in the set to be processed.

[0012] The distribution scheduling engine is used to respond to access requests for target semantic particles, calculate the topological potential of adjacent nodes based on the teaching topology graph to perform data prefetching, and perform cache replacement based on the review urgency of cached data blocks.

[0013] Preferably, the data preprocessing and topology construction module divides the original teaching media stream into several variable-length semantic particles, specifically including:

[0014] The original teaching media stream is sampled according to a preset time granularity, and the spectrogram energy features of the audio track, the keyframe pixel change rate features of the video track, and the keyword density features of the text data are extracted at the sampling points.

[0015] The above features are normalized and weighted and fused to generate a time-varying information density feature vector;

[0016] The local minimum point of the information density feature vector magnitude is detected. When the decrease of the information density feature vector magnitude exceeds a preset magnitude threshold and the duration of the low value exceeds a preset time threshold, the time corresponding to the local minimum point is determined as the semantic boundary segmentation point, and the semantic particles are generated based on the semantic boundary segmentation point.

[0017] Preferably, the initial storage strategy module specifically includes:

[0018] Set a high threshold and a low threshold, determine the semantic particles whose topological centrality is higher than the high threshold as core nodes, assign low ratio erasure coding parameters or multi-replica strategies to the core nodes, and write them into the high-performance solid-state drive storage pool in the heterogeneous storage medium layer.

[0019] Semantic particles whose topological centrality is between the high threshold and the low threshold are identified as ordinary nodes, standard erasure coding parameters are assigned to the ordinary nodes, and written into the standard performance storage pool in the heterogeneous storage medium layer.

[0020] Semantic particles with topological centrality below the low threshold are identified as edge nodes, high-ratio erasure coding parameters are assigned to the edge nodes, and written into the large-capacity hard disk storage pool in the heterogeneous storage medium layer.

[0021] Perform metadata aggregation operation to extract the unique identifier and topological centrality of the semantic particles, associate the allocated erasure coding parameters and bind the physical logical block address returned after data writing, and generate global metadata entries.

[0022] The global metadata entry records the unique identifier, topological centrality, erasure coding parameters, and physical logical block address of the semantic particle.

[0023] Preferably, the cognitive state monitoring module calculates the memory retention rate by specifically including:

[0024] The interaction depth is calculated based on the user's viewing completion rate, rewatch frequency, and quiz score, and the inherent difficulty coefficient of the semantic particles is calculated based on the average number of replays by users across the entire platform.

[0025] A nonlinear function is constructed using the interaction depth and the inherent difficulty coefficient to calculate the user's personalized memory strength for semantic particles;

[0026] The personalized memory strength is used as the decay time constant, and the memory retention rate over time is calculated using the exponential decay function.

[0027] When the memory retention rate decays to a preset critical threshold, the corresponding time point is marked as the theoretical review time, and the predicted review time window is set based on the theoretical review time.

[0028] Preferably, the storage reorganization controller performing the physical address remapping operation specifically includes:

[0029] Perform topological sorting on a set of semantic particles in the set to be processed that are related on the teaching topology graph to generate a linear access sequence that satisfies the cognitive pre-dependency logic;

[0030] In the heterogeneous storage medium layer, allocate a contiguous physical address space not less than the total data volume of the linear access sequence;

[0031] The semantic particles in the linear access sequence are read from their original scattered physical addresses and written sequentially into the continuous physical address space in the order of the linear access sequence, so that logically adjacent semantic particles satisfy the physical proximity constraint in the continuous physical address space.

[0032] Update the global metadata table, update the index pointer of the semantic particles to point to the new address in the contiguous physical address space, and mark the old storage space as invalid.

[0033] Preferably, when the storage reorganization controller performs the physical address remapping operation, it simultaneously performs the following operations:

[0034] Read the predicted review probability corresponding to each semantic particle in the set to be processed;

[0035] When the predicted review probability is higher than the preset dimensionality threshold, the continuous physical address space is allocated to a storage medium with a performance level higher than the original storage level of the semantic granules.

[0036] The transcoding engine is invoked to convert the original high-compression-ratio encoding format of the semantic particles into an I-frame dense format that does not require reference to the preceding and following frames for decoding, and the converted data is written into the continuous physical address space.

[0037] Set a lifecycle timer for the transformed data block. When the predicted review time window is exceeded, trigger a data degradation or deletion operation.

[0038] Preferably, the data prefetching performed by the distribution scheduling engine specifically includes:

[0039] Using the currently visited semantic granular node as the center, determine the candidate nodes in its k-hop neighborhood;

[0040] The knowledge quality of each candidate node is calculated, which is directly proportional to the topological centrality of the candidate node and inversely proportional to the user's memory retention rate of the candidate node.

[0041] Calculate the topological potential energy of each candidate node relative to the currently visited node. The topological potential energy is proportional to the knowledge quality and inversely proportional to the power of the shortest path length from the candidate node to the currently visited node on the teaching topology graph.

[0042] The candidate nodes are sorted in descending order based on the topological potential energy, and a preset number of candidate nodes with the highest ranking are selected. The semantic granular data corresponding to the selected candidate nodes are then asynchronously loaded into the edge cache nodes.

[0043] Preferably, the cache replacement performed by the distribution scheduling engine specifically includes:

[0044] When the space occupancy rate of an edge cache node exceeds the warning threshold, all cached data blocks currently residing in the edge cache node are traversed.

[0045] A retention score is calculated for each cached data block. The retention score is proportional to the topological centrality of the semantic granules corresponding to the cached data block and inversely proportional to the absolute value of the time difference between the predicted review time of the cached data block and the current system time.

[0046] Based on the ascending order of the retention scores, the cached data blocks with the lowest retention scores are prioritized for elimination until the freed space meets the writing requirements of new data.

[0047] Preferably, the computational interaction depth specifically includes:

[0048] If it is detected that a user repeatedly reviews a content at a specific timestamp, the value of the interaction depth is increased;

[0049] If an increase in a user's quiz score is detected, the value of the interaction depth is increased.

[0050] Preferably, when the data preprocessing and topology construction module calculates the topological centrality, it comprehensively calculates the in-degree centrality and the betweenness centrality.

[0051] The in-degree centrality representation depends on the number of subsequent knowledge points of the semantic particle;

[0052] The betweenness centrality represents the frequency of the semantic particles as bridges that must be passed through the shortest path between any two nodes in the teaching topology graph.

[0053] This invention provides an intelligent scheduling and storage optimization system for distance education data. It has the following beneficial effects:

[0054] 1. This invention performs variable-length semantic granular segmentation of media streams through data preprocessing and topology construction modules, and implements differentiated initial storage strategies based on the topological centrality of knowledge points. This allocates low-ratio erasure coding to core nodes with high topological centrality and stores them on high-performance solid-state drives, while allocating high-ratio erasure coding to edge nodes and storing them on large-capacity hard disks. This achieves the effect of dynamically allocating storage resources based on the value of knowledge structure, and solves the mismatch between the high storage costs caused by the unified redundancy strategy and single medium used in existing educational data storage systems and the performance bottleneck of high-concurrency access to core data.

[0055] 2. This invention uses a cognitive state monitoring module combined with a forgetting function to predict the user's review time window, and utilizes a storage reorganization controller to perform physical address remapping and encoding conversion of data during system idle periods. Furthermore, it predicts a set of logically related semantic particles that will be accessed and pre-migrates them to a continuous physical address space, converting the encoding format to a low-decoding-latency I-frame intensive format. This achieves the effect of transforming the random I / O load during user review into a sequential I / O load and eliminating decoding latency, solving the problem of data loading lag and excessive response latency caused by high-frequency review and skipping access in remote education scenarios.

[0056] 3. This invention executes a data prefetching strategy based on topological potential and a cache replacement strategy based on forgetting urgency through a distribution scheduling engine. It then comprehensively considers node knowledge quality, topological distance, and user memory retention rate to calculate data preloading priority, and determines the eviction order of cached data based on the time difference between the distance and the predicted review time. This achieves the effect that the edge caching system can actively adapt to the user's nonlinear learning path and cognitive pattern, and solves the problems that existing educational data storage and scheduling cannot cope with leapfrog learning behavior and that general cache replacement algorithms cannot perceive future review needs, resulting in low cache hit rate. Attached Figure Description

[0057] Figure 1 This is a system module architecture diagram of the present invention;

[0058] Figure 2 This is a flowchart of the operation of the data preprocessing and topology construction module and the initial storage strategy module of the present invention;

[0059] Figure 3 This is a flowchart of the operation of the storage reorganization controller and distribution scheduling engine of the present invention. Detailed Implementation

[0060] Please see the appendix Figure 1 -Appendix Figure 3This invention provides a remote education data intelligent scheduling and storage optimization system, including: a data preprocessing and topology construction module 101, an initial storage strategy module 102, a cognitive state monitoring module 103, a storage reorganization controller 104, a distribution scheduling engine 105, and a heterogeneous storage medium layer 106.

[0061] The data preprocessing and topology construction module 101 is configured to receive the input raw teaching media stream. By analyzing the multimodal features of the media stream, the data preprocessing and topology construction module 101 identifies the semantic boundaries of knowledge points and divides the continuous time stream into several variable-length semantic particles. Then, it further analyzes the teaching syllabus and the dependency logic between knowledge points, constructs a teaching topology graph containing semantic particle nodes and dependency edges, and calculates the topological centrality of each node in the graph structure.

[0062] The initial storage strategy module 102 is connected to the data preprocessing and topology construction module 101. The initial storage strategy module 102 is configured to receive semantic particles and their corresponding topological centrality, and assign erasure coding parameters to each semantic particle according to a preset mapping rule. For semantic particles with high topological centrality, the initial storage strategy module 102 assigns a low-ratio erasure coding or a multi-replica strategy; for semantic particles with low topological centrality, it assigns a high-ratio erasure coding strategy. The initial storage strategy module 102 also writes the processed data blocks to the initial physical address of the heterogeneous storage medium layer 106.

[0063] The cognitive state monitoring module 103 is configured to record historical interaction data of user groups with semantic particles. It integrates a memory strength calculation unit to calculate the user's memory retention rate for specific knowledge points based on a modified Ebbinghaus forgetting curve. Based on the decay trend of memory retention rate, the cognitive state monitoring module 103 calculates the predicted review time window when the user will generate review needs and generates a set of pending processing semantic particles containing identifiers of upcoming review cycles.

[0064] The storage reorganization controller 104 is connected to the cognitive state monitoring module 103 and the heterogeneous storage media layer 106. The storage reorganization controller 104 is configured to operate during idle periods when the system's input / output load is below a preset threshold. Based on the set of data to be processed, the storage reorganization controller 104 controls the heterogeneous storage media layer 106 to perform physical address remapping operations. Specifically, the storage reorganization controller 104 requests contiguous physical storage space, migrates multiple semantic particles that logically belong to the same review topic but are physically dispersed, and sequentially writes them into this contiguous space. Simultaneously, based on the predicted review probability, it performs encoding format conversion and storage media layer upgrades on the data blocks.

[0065] The distribution scheduling engine 105 is configured to respond to real-time access requests from users. Based on the teaching topology graph, the distribution scheduling engine 105 calculates the topological potential difference between the currently accessing node and its neighboring nodes, and preloads data from the heterogeneous storage medium layer 106 to the edge cache nodes according to the potential difference. The distribution scheduling engine 105 is also configured to perform cache replacement operations based on the review urgency of the cached data blocks when cache space is insufficient.

[0066] The heterogeneous storage media layer 106 includes physical storage devices at various performance levels, such as non-volatile memory, solid-state drives, and hard disk drives. This layer supports sequential writes and random reads based on logical block addresses and responds to data migration commands from the storage reorganization controller 104.

[0067] The data preprocessing and topology construction module 101 performs structured processing and feature quantification of the original teaching media stream. The specific implementation process of the data preprocessing and topology construction module 101 includes semantic granular segmentation of unstructured data, construction of teaching-dependent topology graphs, and calculation of node topological centrality.

[0068] The data preprocessing and topology construction module 101 first receives the input teaching media data stream, which includes video tracks, audio tracks, and text data. The data preprocessing and topology construction module 101 then executes a variable-length segmentation strategy based on multimodal feature fusion, specifically including the following sub-steps:

[0069] Step S210: Feature vector extraction.

[0070] The data preprocessing and topology construction module 101 samples the teaching media data stream at a preset time granularity (e.g., 0.5 seconds) and extracts multimodal features at each sampling point. For audio tracks, short-time Fourier transform is used to extract spectroscopic energy features, and the pauses and intonation changes of speech are characterized by calculating short-time energy and zero-crossing rate. For video tracks, keyframes are extracted and the pixel change rate or histogram difference between adjacent frames is calculated to characterize the switching of visual scenes. For text data, a keyword extraction algorithm (such as TF-IDF) is used to calculate the term density within the current time window. The specific algorithm implementation for the above feature extraction belongs to the existing technology in the fields of signal processing and natural language processing, and will not be elaborated here.

[0071] Step S220: Information density curve generation and boundary determination.

[0072] The data preprocessing and topology construction module 101 normalizes the extracted audio energy features, video rate of change features, and text density features, and then weights and fuses them to generate a time-varying information density feature vector. The data preprocessing and topology construction module 101 sets a sliding window to scan along the time axis and identifies the semantic boundaries of the teaching content by detecting local minima of the feature vector magnitude.

[0073] The specific determination logic is as follows: The system presets an amplitude threshold (e.g., the decrease rate of the feature vector magnitude relative to the average of the previous time window exceeds 30%) and a time threshold (e.g., the duration of the low-value state exceeds 2 seconds). The data preprocessing and topology construction module 101 calculates the rate of change of the feature vector magnitude in real time. When it detects that the magnitude decrease exceeds the amplitude threshold at a certain moment, and the duration of the low-value state exceeds the time threshold, that moment is determined as a split point.

[0074] Step S230: Semantic granule generation.

[0075] Based on the aforementioned set of segmentation points, the data preprocessing and topology construction module 101 physically segments the teaching media data stream into several independent units of varying time lengths, i.e., a set of semantic particles. Each semantic particle Logically, each unit corresponds to a relatively complete teaching knowledge point, covering the explanation, demonstration, or derivation process of that knowledge point, ensuring the atomicity of subsequent storage and scheduling.

[0076] After semantic granular segmentation, the data preprocessing and topology construction module 101 further constructs a mathematical model based on the teaching syllabus and the logical relationships between knowledge points. The data preprocessing and topology construction module 101 parses the accompanying course metadata or extracts the pre- and post-relationships between knowledge points through semantic analysis, constructing a directed weighted graph. Among them, the node set With semantic granule set One-to-one correspondence, edge set This represents the cognitive dependency path between knowledge points. If learning semantic particle v2 requires understanding semantic particle v1 as a prerequisite, then a directed edge is constructed in the graph from v1 to v2.

[0077] To provide a quantitative basis for subsequent storage redundancy strategies, the data preprocessing and topology construction module 101 calculates the topological centrality of each node in the topology graph. Unlike general network metrics, it employs a weighted design specifically for educational scenarios to characterize the structural importance of particular knowledge points within the entire curriculum. The topological centrality of a node is defined as follows:

[0078] ;

[0079] In the formula, Represents the topological centrality of a node. Representative node The in-degree represents the number of subsequent semantic particles that directly depend on this semantic particle; Represents any two nodes in the graph and The total number of shortest paths between them. This indicates the nodes passed through in these shortest paths. The number of paths, in a physical sense, represents the degree to which a node serves as an essential bridge in the course learning path, reflecting the pivotal nature of the node. and These are preset weighting coefficients used to adjust the proportion of fundamental and pivotal aspects in the total centrality.

[0080] Through the above processing, the data preprocessing and topology construction module 101 transforms the linear media stream into graph-structured data with weighted information. Each semantic particle v3 contained in a node of the graph-structured data not only contains specific audio and video data but is also bound to its corresponding topological centrality. and in the diagram The adjacency table in the database contains metadata that is output to subsequent modules to guide storage distribution and scheduling decisions.

[0081] The initial storage strategy module 102 performs differentiated data persistence operations based on the topological attributes calculated by the preceding modules. Its main function is to establish a mapping relationship between the importance of semantic granularities and the reliability and performance parameters of the underlying storage system, thereby achieving knowledge-value-based resource allocation during the data writing phase. This process specifically includes steps such as dynamic erasure coding parameter mapping, heterogeneous media tiered storage, and metadata binding.

[0082] The initial storage strategy module 102 receives semantic granular data and its corresponding topological centrality. .

[0083] Step S310, semantic granularity classification.

[0084] The system sets high and low thresholds to divide semantic particles into core nodes, ordinary nodes, and edge nodes. Core nodes typically correspond to basic knowledge points located at the center of the topology graph or serving as common prerequisites for multiple learning paths; edge nodes correspond to extended knowledge points located at the end of the topology graph or with fewer dependencies; ordinary nodes correspond to semantic particles with topological centrality between the high and low thresholds, and typically serve as transitional knowledge nodes connecting core basic knowledge and peripheral extended content.

[0085] Step S320: Perform dynamic erasure coding parameter mapping based on centrality.

[0086] The initial storage strategy module 102 does not employ a uniform redundancy strategy for all data. Instead, it calls a pre-defined mapping function and adjusts the redundancy based on centrality. Determine erasure coding configuration parameters .in, This represents the total number of data blocks after encoding. Represents the number of original data blocks; redundancy is determined by the ratio. / Decide.

[0087] For semantic particles (i.e., topological centrality) that are determined to be core nodes (≥The high threshold set by the system), the initial storage strategy module 102 allocates a highly redundant erasure coding strategy or a multi-replica strategy. For example, using... This configuration allows the system to tolerate up to 50% of data blocks being corrupted, or it can directly use three-replica mirrored storage. This strategy gives the core nodes high data redundancy, and the core nodes are semantic granularities with topological centrality exceeding a high threshold.

[0088] For semantic particles (i.e., topological centrality) that are determined to be edge nodes (The initial storage policy module 102 allocates a high-storage-efficiency erasure coding policy based on the system's low threshold). For example, using... This configuration results in low redundancy. This strategy leverages the relatively low access frequency and non-critical path characteristics of edge nodes to preserve data with minimal storage overhead, optimizing the overall cost-effectiveness of the storage system. For ordinary nodes falling between these two extremes, a standard erasure coding configuration (such as...) is used. The specific encoding algorithm for the erasure coding mentioned above can be either Reed-Solomon codes or local repair codes, which are existing technologies in the field of channel coding and will not be discussed further here.

[0089] Step S330: Physical address allocation based on media performance.

[0090] The system's heterogeneous storage media layer 106 includes various storage devices with different performance and costs, such as solid-state drives (SSDs) based on non-volatile memory high-speed interface (NVMe) and shingled magnetic recording (SMR) hard disk drives (HDDs). The initial storage strategy module 102 writes them to different physical media layers according to the category of semantic granules.

[0091] Specifically, for semantic granules with high topological centrality, the initial storage strategy module 102 directly routes their data blocks and parity blocks to a high-performance solid-state drive (SSD) storage pool with a high-speed memory interface. Since core nodes are often essential points on the user's learning path and are accessed significantly more frequently than other nodes, placing them in low-latency media effectively reduces I / O latency under high concurrency. Conversely, for semantic granules with low topological centrality, the initial storage strategy module 102 routes them to a large-capacity SMR HDD storage pool. This type of data is typically optional or supplementary material with low access frequency; storing it in low-cost media optimizes storage resource utilization.

[0092] Step S340: Global metadata generation.

[0093] The initial storage strategy module 102 performs metadata aggregation and generation operations. First, it extracts the unique identifier of each semantic particle and its calculated topological centrality value; then, it associates it with the erasure coding parameters assigned to that particle in step S320; finally, it binds the physical logical block address returned after data writing is completed in step S330. Finally, it combines these four key attributes to construct a complete global metadata entry.

[0094] The global metadata entries generated by the initial storage strategy module 102 record the unique identifier of each semantic particle, the corresponding topological centrality value, and the erasure coding parameters used. This includes the physical logical block address (LBA) of each data shard in the heterogeneous storage media layer. This metadata is synchronized to the system's metadata server, serving as an index for subsequent data retrieval, reassembly, and scheduling. Through these steps, the system completes the initial alignment of the physical storage structure and logical topology during the data ingestion phase.

[0095] The cognitive state monitoring module 103 is configured to record user interaction data with teaching content and calculate the user's memory state for specific semantic particles based on the interaction data, thereby providing a time-dimensional predictive basis for subsequent memory reorganization. This process mainly involves quantitative modeling of memory strength, real-time calculation of memory decay curves, and window prediction of review thresholds.

[0096] Step S410: Extraction of multi-dimensional interactive behavior features.

[0097] The cognitive state monitoring module 103 retrieves historical operation records of specific users for specific semantic granules from the system log. To accurately quantify the user's initial grasp of the knowledge points, the cognitive state monitoring module 103 calculates an interaction depth index. This index is not based solely on the number of clicks, but rather integrates viewing completion rate, replay frequency of key segments, and scores on in-class quizzes.

[0098] Specifically, a higher completion rate, i.e. a higher test score, indicates a deeper level of interaction. For repeated viewing operations at specific timestamps, the system identifies them as repeated access behaviors targeting specific semantic particles, which also contribute to the increase in interaction depth.

[0099] Meanwhile, the cognitive state monitoring module 103 calculates the inherent difficulty coefficient of each semantic particle by aggregating the historical learning data of all users on the platform. This coefficient is based on the global error rate or average number of retakes of the test questions corresponding to that knowledge point, and the numerical range is normalized to the interval [0,1).

[0100] Step S420, memory strength calculation.

[0101] The cognitive state monitoring module 103 uses a preset nonlinear function to calculate the user's memory strength for specific semantic particles. The calculation formula is as follows:

[0102] ;

[0103] In the formula, It is the basic memory factor, used to adjust the overall time scale; This is a difficulty-sensitive factor used to control the degree to which the difficulty of knowledge weakens memory strength. According to this formula, memory strength is related to interaction depth. Positive correlation with the difficulty of the knowledge point Negative correlation.

[0104] Step S430, memory retention rate deduction.

[0105] The cognitive state monitoring module 103 uses a modified Ebbinghaus forgetting function to extrapolate the decay trajectory of memory retention rate in real time. Unlike the general forgetting curve, this scheme introduces the aforementioned calculated personalized memory strength. As a decay time constant, the time interval between the current moment and the user's last completion of learning this semantic particle is defined as: The current memory retention rate The calculation is as follows:

[0106] ;

[0107] In the formula, Indicates user semantic granules After time The memory retention rate after that is usually in the range of (0,1]. This represents the forgetting time interval, i.e., the time elapsed since the current moment when the user forgot. Last time this semantic particle was completed The length of time spent learning or interacting; Indicates user semantic granules The memory strength, which is a parameter of the decay function, is determined by the depth of user interaction and the difficulty of the knowledge points. The base of the natural logarithm (a mathematical constant).

[0108] The modified Ebbinghaus forgetting function can dynamically reflect the differences in the rate at which different users forget knowledge points of varying difficulty. For simple knowledge points or content involving deep interaction, A larger value results in a smaller absolute value of the exponent, leading to a lower memory retention rate. Over time The decay rate is relatively slow; conversely, for content that is difficult to access and only superficially browsed, the memory loss rate is relatively fast.

[0109] Step S440: Review critical point time window prediction.

[0110] The cognitive state monitoring module 103 performs time window prediction for the review critical point. The system presets a critical threshold for memory retention rate. (For example, 0.6), this threshold represents the optimal intervention point when the user is about to forget key information and urgently needs to review it. The cognitive state monitoring module 103 solves the equation In the formula, Indicates user semantic granules In the future Predicted memory retention rate at that time; This represents the theoretical review moment to be solved, that is, the absolute time point when the memory retention rate decays to a critical value; This represents the system's preset critical threshold for memory retention (e.g., 0.6), indicating the optimal memory state threshold for suggesting review intervention to the user. This equation is then used to calculate the theoretical review time for each semantic particle for the user.

[0111] Based on this theory, the cognitive state monitoring module 103 sets a preprocessing buffer time for the review period. Define interval To predict the review time window. When the system's current time... When the time window has not yet arrived but is approaching, the cognitive state monitoring module 103 marks the relevant semantic particles as soon to be active. The cognitive state monitoring module 103 further aggregates a group of semantic particles with similar predicted review time windows and correlations on the teaching topology map, generating a set to be processed. This set contains a group of logically strongly correlated data block identifiers that will be frequently accessed by the user, and is sent to the storage reorganization controller 104 to trigger subsequent proactive optimization operations.

[0112] The storage reorganization controller 104 performs a physical address remapping operation. This operation aims to utilize the storage system's I / O idle periods to rearrange logically related but physically scattered data blocks into a contiguous physical address space according to a predicted access order. This process transforms potential random read loads from users into sequential read loads, thereby reducing seek latency and protocol overhead of the underlying storage devices. The specific implementation of this mechanism includes steps such as load status determination, topology serialization and sorting, contiguous space allocation, and data migration and rewriting.

[0113] Step S510, I / O load status monitoring.

[0114] The storage reorganization controller 104 first monitors the current input / output operations per second (IOPS) and throughput metrics of the heterogeneous storage media layer. When it detects that the current I / O load of the system is lower than a preset idle threshold, the storage reorganization controller 104 activates the background reorganization process and reads the pending set generated by the cognitive state monitoring module 103.

[0115] Step S520: Semantic granule topological serialization.

[0116] The storage reorganization controller 104 performs topological serialization processing on the semantic particles in the set to be processed. Since the semantic particles in the set to be processed may have complex network dependencies (such as one-to-many or many-to-one) in the teaching topology graph, the storage reorganization controller 104 calls the topological sorting algorithm to flatten the subgraph structure into a linear access sequence. In the formula, This represents a linear access sequence generated after topological sorting, where the order of elements in the sequence represents the suggested review or access order. This represents each semantic particle (i.e., a teaching knowledge unit) in the sequence, where the subscript indicates its position in the sequence; This represents the total number of semantic particles in the set to be processed.

[0117] This sequence strictly follows the cognitive pre-dependency logic of knowledge points, that is, for any two elements in the sequence and ,like Therefore, there is no logical basis for this from the perspective of cognition. point to This strong dependency path ensures that when a user reviews according to the teaching logic, the data they request is monotonically increasing in logical address.

[0118] Step S530: Pre-allocate contiguous physical address space.

[0119] The storage reorganization controller 104 performs pre-allocation of a contiguous physical address space. The storage reorganization controller 104 calculates the sum of all semantic granular data in the sequence and retrieves unoccupied contiguous logical block address ranges in the heterogeneous storage media layer (e.g., a high-performance storage pool). This range is not less than the calculated total data volume. If a sufficiently large single contiguous range does not exist in the storage system, the storage reorganization controller 104 performs fragmentation merging or requests multiple large-capacity block groups, maximizing address contiguousness within each block group. During this process, the system locks the contiguous LBA range to prevent preemption by other concurrent write operations.

[0120] Step S540, Physical migration and remapping.

[0121] Storage reorganization controller 104 performs physical migration and remapping of data. For each semantic particle in the sequence, storage reorganization controller 104 reads data blocks from its current storage location (which may be scattered across different disk sectors, different stripes, or different server nodes) and writes them sequentially into the consecutive LBA intervals requested in step S530.

[0122] Specifically, assuming semantic granules Originally stored in physical address After the migration, it was written to the new physical address. At this point, two adjacent semantic particles in the sequence and Satisfying the physical proximity constraint on the new physical storage medium, i.e. Adjacent The tail address. This physical-level reordering differs from logical links or shortcuts at the file system level; instead, it performs a migration and writing of data from the original physical address to the new physical address.

[0123] For hard disk drives (HDDs), this rearrangement eliminates the mechanical seek time when the read / write head is reading different data points; for solid disk drives (SSDs), it increases the probability of read operations hitting the same flash page or block, thus improving internal parallelism.

[0124] Step S550: Update the global metadata table.

[0125] After the data writing is complete, the storage reassembly controller 104 updates the system's global metadata table. The system will then... The index pointers of each semantic particle are updated to point to the new physical address. and mark the old physical address If an item is in an invalid state, the storage space it occupies is released for subsequent garbage collection. Through the above steps, the system has completed data layout optimization at the physical layer before the user initiates a review request, so that the future review process only requires a single continuous streaming read operation on the disk to obtain complete knowledge topic data.

[0126] During the physical rearrangement process described above, the storage reorganization controller 104 simultaneously performs media dimensionality upgrades and encoding conversions based on cognitive prediction. This mechanism dynamically adjusts the storage media level and encoding format of the data at the physical layer based on the review probability generated by the cognitive state monitoring module, eliminating decoding latency by trading storage space for access speed. This process is specifically integrated into the data migration pipeline and includes steps such as review probability determination, cross-level media upgrades, low-latency format transcoding, and lifecycle management.

[0127] Step S560: Determine the urgency of review.

[0128] For each semantic particle in the set to be processed, the storage reorganization controller 104 reads its corresponding predicted review probability or review urgency score. The system presets an upgrade threshold to determine whether performance enhancement is needed for the data block corresponding to the semantic particle. This threshold is usually set to a high value (e.g., 0.8), and upgrade operations are only triggered for difficult knowledge or core review content that are highly likely to be accessed by users in the near future, in order to balance the overall resource consumption of the system.

[0129] Step S570, cross-level media enhancement.

[0130] The storage reorganization controller 104 performs cross-level media promotion. When the semantic granularity is determined to satisfy a predicted review probability greater than the dimensionality increase threshold, the storage reorganization controller 104 modifies the target write address in step S540. Specifically, instead of rearranging within the same level of storage media, the storage reorganization controller 104 allocates the target physical address to a higher-performance storage level. For example, the data source is read from a large-capacity, low-speed mechanical hard drive and written to non-volatile memory (NVDIMM) or a high-performance NVMe solid-state drive. Through this proactive data migration, the system ensures that when a user initiates a review request, the data resides on the physical medium with the highest IOPS and lowest read / write latency, avoiding I / O bottlenecks caused by retrieving data from slow disks.

[0131] Step S580, low-latency encoding format transcoding.

[0132] The storage reorganization controller 104 performs encoding format transcoding with a latency reduction focus. In distance education scenarios, to save bandwidth and storage, the original teaching videos usually use high compression ratio encoding formats (such as H.265 / HEVC). This format requires a lot of terminal CPU resources for decoding during playback, resulting in a delay in the loading of the first frame.

[0133] In this step, the storage reconstruction controller 104 invokes the transcoding engine to execute a format conversion function during the data transfer process from the source medium to the target medium. The conversion logic is as follows:

[0134] ;

[0135] In the formula, This indicates the target encoding format after transcoding, which is the data format that will be written to the high-performance storage layer. This indicates an I-frame dense format or a full keyframe format, which requires no reference frame for decoding and has extremely low playback latency. This indicates the source encoding format, i.e., the original high compression ratio format of the storage (such as H.265 / HEVC). This indicates the predicted review probability (or review urgency score) for that semantic particle. This represents the format conversion threshold. When the predicted review probability is higher than this threshold, the system triggers a transcoding operation to a low-latency format.

[0136] For semantic particles with high review probability, the storage reassembly controller 104 decompresses and re-encodes them into an "I-frame intensive" format or a low-compression format containing only keyframes, or even, in extreme cases, into uncompressed RAW or YUV format. Although this conversion leads to a significant increase in data volume, the decoding process does not require reference to preceding and following frames, allowing the terminal device to render directly, greatly reducing playback latency.

[0137] The above strategy is based on a core technical judgment: in the upcoming review window, users' need for timely access is higher than consideration of storage costs.

[0138] Step S590, Lifecycle Management.

[0139] The storage reorganization controller 104 sets a time-to-live (TTL) timer for the data blocks that have undergone dimensionality upgrades and transcoding. Since the upgraded data occupies significant storage resources in the high-performance storage layer and is quite large, a retention period is set. This timer is typically set to the length of a predicted review window (e.g., 24 hours). When the time window ends and the user has not accessed the data, or after the user has completed the review, the system triggers a degradation mechanism, deleting the data block from the high-performance layer or re-transcoding it back to a high-compression format and migrating it back to the low-cost storage layer, thus completing the dynamic reclamation and closed-loop management of storage resources. Through this mechanism, the system achieves flexible adaptation of storage configuration to the user's cognitive state.

[0140] The distribution scheduling engine 105 performs a topology-guided data prefetching operation. This operation dynamically calculates the access probability distribution of adjacent nodes in the teaching knowledge graph during real-time learning and preloads high-probability data blocks from the heterogeneous storage medium layer into the memory or cache of the edge computing nodes accordingly. This mechanism overcomes the limitation of traditional streaming media, which can only perform prefetching based on a linear temporal order, and can adapt to non-linear, jump-like learning behaviors. The process specifically includes steps such as current state localization, node quality calculation, potential field distribution deduction, and asynchronous preloading.

[0141] Step S610: Access request response and location.

[0142] The distribution and scheduling engine 105 responds to access requests initiated by user terminals for specific semantic granularities. The distribution and scheduling engine 105 locates the node in the teaching dependency topology graph and, using this node as the center, determines its location using a breadth-first search (BFS) algorithm. All nodes within the skip neighborhood are used to form a candidate prefetch set. Here... The value is usually set to a small integer (such as 2 or 3) to limit the computational range and reduce system overhead. The specific implementation logic of breadth-first search belongs to the existing technology in the field of graph theory algorithms, and will not be elaborated here.

[0143] Step S620: Calculate the knowledge quality of candidate nodes.

[0144] The distribution scheduling engine 105 calculates the knowledge quality of each node in the candidate set. In the topological potential model, the quality of a node determines its predicted access probability. This scheme combines the structural importance of knowledge points with the user's cognitive forgetting state to define this quality parameter. The specific calculation logic is as follows: the topological centrality of the node is compared with the user's forgetting level. Multiply.

[0145] in, This refers to the memory retention rate calculated in real-time by the cognitive state monitoring module. The calculation logic for knowledge quality above indicates that the more core a knowledge point is within the curriculum, the lower the user's memory retention rate (i.e.,...). The closer a value is to 0, the greater the knowledge quality of that node, and the higher the likelihood that a user will stay at or jump to that node.

[0146] Step S630: Calculate the topological potential of candidate nodes.

[0147] The distribution and scheduling engine 105 calculates the topological potential energy of candidate nodes based on the potential energy field generated by the current node. This potential energy value represents the strength of a user's tendency to jump from the current knowledge point to the target knowledge point. The calculation formula is as follows:

[0148] ;

[0149] In the formula, This represents the shortest path length (number of hops) from the current node to the target node in the topology graph. It is a very small positive constant used to prevent the denominator from being zero when calculating its own potential energy; This is the distance damping coefficient, used to simulate the characteristic that user attention rapidly decreases as the jump distance increases; This indicates the current node. In this state, candidate nodes The topological potential energy; Indicates candidate nodes; Indicates candidate nodes The quality of knowledge is determined by the product of the node topological centrality defined in step S620 and the degree of user forgetting.

[0150] This formula states that the prefetch priority is directly proportional to the knowledge quality of the target node and inversely proportional to the power of the topological distance. Even if some nodes are far apart in the physical file list, if they are logically adjacent to the current node and happen to be in the user's forget zone, their potential value will still be significantly higher than that of linearly adjacent nodes.

[0151] Step S640: Prefetch list generation and asynchronous I / O.

[0152] The distribution and scheduling engine 105 sorts the candidate set in descending order based on the calculated potential energy value, and extracts the top N nodes to generate a prefetch list. The distribution and scheduling engine 105 then sends asynchronous I / O read instructions to the underlying storage system.

[0153] If the target data block has already undergone physical rearrangement by the storage reorganization controller 104 (i.e., is in a physically contiguous state), the distribution scheduling engine 105 will initiate a batch read based on the logical block address range, reading data segments containing multiple high-potential granules into memory at once. If the data block has not been rearranged, the distribution scheduling engine 105 will initiate parallel random read requests. The read data is filled into the circular buffer of the edge node. When the user actually triggers the jump operation, the system directly pushes the data stream from memory to the client, thereby eliminating network round-trip time (RTT) and disk I / O latency, achieving low-latency data loading.

[0154] When the local cache space of an edge node or terminal device reaches saturation, the distribution scheduling engine 105 executes a cache replacement strategy based on forgetting urgency. This strategy aims to overcome the shortcomings of traditional Least Recently Used (LRU) or Least Frequently Used (LFU) algorithms, which only focus on historical access frequency and cannot perceive future cognitive needs. The core logic of this strategy is to calculate the retention value of cached data blocks relative to the user's memory curve, and prioritize the elimination of data blocks that are neither core knowledge points nor far from the predicted review time. This process specifically includes cache status monitoring, retention score calculation, and elimination execution.

[0155] Step S650, cache space monitoring.

[0156] The distribution and scheduling engine 105 monitors the storage space occupancy rate of edge cache nodes in real time. When it detects that the current occupancy rate exceeds a preset warning threshold (e.g., 85% of the total capacity) and a new high-potential data block is requested to be written, the system triggers the cache replacement procedure. The distribution and scheduling engine 105 traverses all data block indices currently residing in the cache and constructs a candidate eviction set. To prevent system jitter, the distribution and scheduling engine 105 typically does not evict just one data block, but instead calculates the target space size to be released and marks a batch of data blocks that need to be evicted.

[0157] Step S660: Retain score calculation.

[0158] The dispatch scheduling engine 105 dispatches each cache block in the candidate set. Calculate the retention score This score quantifies the necessity of keeping the data block in the cache at the current moment. During the calculation, the distribution scheduling engine 105 first retrieves the topological centrality of the semantic particle corresponding to the data block from the global metadata table. And the current user's predicted review time for this semantic particle. .

[0159] Based on the above parameters, the distribution scheduling engine 105 calculates the retention score according to the following formula:

[0160] ;

[0161] In the formula, Represents the current absolute timestamp of the system; denominator term This represents the absolute value of the time difference between the current moment and the predicted review time. It is a very small positive constant (e.g., 0.001) to prevent numerical overflow caused by the denominator being zero when the current time is exactly equal to the predicted review time.

[0162] The filtering logic embodied in this formula is as follows:

[0163] Firstly, the time dimension. When Approaching When the denominator approaches zero, the whole score is retained. The value increases sharply. This corresponds to a scenario where review is imminent, and the data is about to be accessed. Therefore, the system gives it a very high retention priority to prevent accidental deletion just before access.

[0164] Secondly, the content dimension. (Sub-item) This ensures that core knowledge points (high centrality) have a higher retention weight than peripheral knowledge points under the same time conditions. Even if two data blocks are at the same time away from review, the system will prioritize protecting the data located at the center of the teaching topology because it is more likely to be indexed again or used as a stepping stone.

[0165] Conversely, for data blocks whose memory retention rate is lower than the preset repair threshold or higher than the preset safety threshold, the retention score is lower due to the larger denominator, making them the preferred targets for elimination.

[0166] Step S670: Elimination execution and space release.

[0167] The distribution scheduling engine 105 sorts all candidate cache blocks in ascending order by their retention scores. The distribution scheduling engine 105 then selects the block with the lowest score. Each cache block is selected as a eviction target until the freed space meets the writing requirements of new data. For a selected eviction block, the distribution and scheduling engine 105 checks its data modification status. If the data block has not been modified during caching (read-only data), the distribution and scheduling engine 105 directly marks it as invalid in the index table, overwriting its physical address; if the data block contains incremental data such as user-generated notes or annotations, the distribution and scheduling engine 105 first writes it back to the underlying heterogeneous storage medium layer, and then performs the space release operation. Through this mechanism, the system manages cache space based on predicted review needs.

Claims

1. A remote education data intelligent scheduling and storage optimization system, characterized in that, include: The data preprocessing and topology construction module is used to receive the original teaching media stream, identify semantic boundaries through multimodal feature fusion, segment the original teaching media stream into several variable-length semantic particles, parse the knowledge points between each semantic particle, and construct a teaching topology graph based on logic, while calculating the topological centrality of each semantic particle in the teaching topology graph. The heterogeneous storage medium layer includes physical storage devices with multiple performance levels, used to store the data blocks and corresponding check blocks of the semantic granules; An initial storage strategy module, connected to the data preprocessing and topology construction module and the heterogeneous storage medium layer, is used to allocate erasure coding parameters and initial physical storage levels to each semantic particle according to the topology centrality. The cognitive state monitoring module is used to record the user's historical interaction behavior data with the semantic particles, calculate the memory retention rate based on the historical interaction behavior data, calculate the predicted review time window based on the decay trend of the memory retention rate, and generate a set of unprocessed semantic particles containing the identifiers of the semantic particles that are about to enter the predicted review time window. A storage reorganization controller, connected to the cognitive state monitoring module and the heterogeneous storage medium layer, is used to read the set to be processed when the system input / output load is lower than a preset idle threshold, and to control the heterogeneous storage medium layer to perform physical address remapping operations on the semantic particles in the set to be processed. The distribution scheduling engine is used to respond to access requests for target semantic particles, calculate the topological potential of adjacent nodes based on the teaching topology graph to perform data prefetching, and perform cache replacement based on the review urgency of cached data blocks.

2. The intelligent scheduling and storage optimization system for distance education data according to claim 1, characterized in that, The data preprocessing and topology construction module divides the original teaching media stream into several variable-length semantic particles, specifically including: The original teaching media stream is sampled according to a preset time granularity, and the spectrogram energy features of the audio track, the keyframe pixel change rate features of the video track, and the keyword density features of the text data are extracted at the sampling points. The above features are normalized and weighted and fused to generate a time-varying information density feature vector; The local minimum point of the information density feature vector magnitude is detected. When the decrease of the information density feature vector magnitude exceeds a preset magnitude threshold and the duration of the low value exceeds a preset time threshold, the time corresponding to the local minimum point is determined as the semantic boundary segmentation point, and the semantic particles are generated based on the semantic boundary segmentation point.

3. The intelligent scheduling and storage optimization system for distance education data according to claim 1, characterized in that, The initial storage strategy module specifically includes: Set a high threshold and a low threshold, determine the semantic particles whose topological centrality is higher than the high threshold as core nodes, assign low ratio erasure coding parameters or multi-replica strategies to the core nodes, and write them into the high-performance solid-state drive storage pool in the heterogeneous storage medium layer. Semantic particles whose topological centrality is between the high threshold and the low threshold are identified as ordinary nodes, standard erasure coding parameters are assigned to the ordinary nodes, and written into the standard performance storage pool in the heterogeneous storage medium layer. Semantic particles with topological centrality below the low threshold are identified as edge nodes, high-ratio erasure coding parameters are assigned to the edge nodes, and written into the large-capacity hard disk storage pool in the heterogeneous storage medium layer. Perform metadata aggregation operation to extract the unique identifier and topological centrality of the semantic particles, associate the allocated erasure coding parameters and bind the physical logical block address returned after data writing, and generate global metadata entries. The global metadata entry records the unique identifier, topological centrality, erasure coding parameters, and physical logical block address of the semantic particle.

4. The intelligent scheduling and storage optimization system for distance education data according to claim 1, characterized in that, The cognitive state monitoring module calculates the memory retention rate specifically by: The interaction depth is calculated based on the user's viewing completion rate, rewatch frequency, and quiz score, and the inherent difficulty coefficient of the semantic particles is calculated based on the average number of replays by users across the entire platform. A nonlinear function is constructed using the interaction depth and the inherent difficulty coefficient to calculate the user's personalized memory strength for semantic particles; The personalized memory strength is used as the decay time constant, and the memory retention rate over time is calculated using the exponential decay function. When the memory retention rate decays to a preset critical threshold, the corresponding time point is marked as the theoretical review time, and the predicted review time window is set based on the theoretical review time.

5. The intelligent scheduling and storage optimization system for distance education data according to claim 1, characterized in that, The storage reorganization controller performs physical address remapping operations, specifically including: Perform topological sorting on a set of semantic particles in the set to be processed that are related on the teaching topology graph to generate a linear access sequence that satisfies the cognitive pre-dependency logic; In the heterogeneous storage medium layer, allocate a contiguous physical address space not less than the total data volume of the linear access sequence; The semantic particles in the linear access sequence are read from the original scattered physical addresses and written into the continuous physical address space in the order of the linear access sequence, so that logically adjacent semantic particles satisfy the physical proximity constraint in the continuous physical address space. Update the global metadata table, update the index pointer of the semantic particles to point to the new address in the contiguous physical address space, and mark the old storage space as invalid.

6. The intelligent scheduling and storage optimization system for distance education data according to claim 1, characterized in that, When the storage reorganization controller performs the physical address remapping operation, it simultaneously performs the following operations: Read the predicted review probability corresponding to each semantic particle in the set to be processed; When the predicted review probability is higher than the preset dimensionality threshold, the continuous physical address space is allocated to a storage medium with a performance level higher than the original storage level of the semantic granules. The transcoding engine is invoked to convert the original high-compression-ratio encoding format of the semantic particles into an I-frame dense format that does not require reference to the preceding and following frames for decoding, and the converted data is written into the continuous physical address space. Set a lifecycle timer for the transformed data block. When the predicted review time window is exceeded, trigger a data degradation or deletion operation.

7. The intelligent scheduling and storage optimization system for distance education data according to claim 6, characterized in that, The data prefetching process performed by the distribution and scheduling engine specifically includes: Using the currently visited semantic granular node as the center, determine the candidate nodes in its k-hop neighborhood; The knowledge quality of each candidate node is calculated, which is directly proportional to the topological centrality of the candidate node and inversely proportional to the user's memory retention rate of the candidate node. Calculate the topological potential energy of each candidate node relative to the current visited node. The topological potential energy is proportional to the knowledge quality and inversely proportional to the power of the shortest path length from the candidate node to the current visited node on the teaching topology graph. The candidate nodes are sorted in descending order based on the topological potential energy, and a preset number of candidate nodes with the highest ranking are selected. The semantic granular data corresponding to the selected candidate nodes are then asynchronously loaded into the edge cache nodes.

8. The intelligent scheduling and storage optimization system for distance education data according to claim 7, characterized in that, The cache replacement performed by the distribution scheduling engine specifically includes: When the space occupancy rate of an edge cache node exceeds the warning threshold, all cached data blocks currently residing in the edge cache node are traversed. A retention score is calculated for each cached data block. The retention score is proportional to the topological centrality of the semantic granules corresponding to the cached data block and inversely proportional to the absolute value of the time difference between the predicted review time of the cached data block and the current system time. Based on the ascending order of the retention scores, the cached data blocks with the lowest retention scores are prioritized for elimination until the freed space meets the writing requirements of new data.

9. The intelligent scheduling and storage optimization system for distance education data according to claim 4, characterized in that, The computational interaction depth specifically includes: If it is detected that a user repeatedly reviews a content at a specific timestamp, the value of the interaction depth is increased; If an increase in a user's quiz score is detected, the value of the interaction depth is increased.

10. The intelligent scheduling and storage optimization system for distance education data according to claim 1, characterized in that, When the data preprocessing and topology construction module calculates the topological centrality, it comprehensively calculates the in-degree centrality and the betweenness centrality. The in-degree centrality representation depends on the number of subsequent knowledge points of the semantic particle; The betweenness centrality represents the frequency of the semantic particles as bridges that must be passed through the shortest path between any two nodes in the teaching topology graph.

Images