Machine learning-based dynamic allocation management system for educational resources

Abstract

The application relates to the technical field of intelligent distribution of educational resources, and discloses a dynamic distribution management system of educational teaching resources based on machine learning. The system comprises the following steps: collecting, through a data acquisition module, resource request time stamps, transmission rates and computing load data of learning terminals in real time; processing the data by a multi-resolution analysis module to generate a resource use mode atlas containing time, space and content dimensions; analyzing the atlas by a resource demand mapping module using a pre-trained deep belief network to output accurate resource quota suggestions; optimizing the suggestions based on network topology by a delay compensation module to form a delay-resistant distribution scheme; executing the distribution by a resource scheduling module and starting audit tracking; and updating network parameters according to the audit data by an adaptive optimization module. The system realizes accurate prediction and dynamic adaptive optimization of resource demand, and improves resource utilization efficiency and teaching service quality.

Description

Technical Field

[0001] This invention relates to the field of intelligent allocation technology of educational resources, specifically to a dynamic allocation and management system for educational resources based on machine learning. Background Technology

[0002] The digital allocation and management of educational resources is a core challenge in the field of educational informatization. Most current mainstream resource allocation systems rely on static configuration or dynamic adjustments based on simple linear rules. These systems typically allocate computing, storage, and network resources based on historical average usage data or preset fixed thresholds. Monitoring methods are often limited to periodic sampling and alarms for individual devices or isolated resource indicators, lacking in-depth insights into the inherent correlations between different types of data. This discrete, single-point threshold-based monitoring model struggles to effectively capture the complex and collaborative changes in resource demands arising from different course content, teaching segments, and teacher-student interactions during teaching activities.

[0003] Existing technical solutions have shortcomings, the core issues being low prediction accuracy and poor adaptability in resource allocation. Static allocation schemes cannot adapt to the inherent dynamism and unpredictability of the teaching process. Dynamic adjustment mechanisms based on fixed rules, due to their relatively simple and rigid rule logic, struggle to model and respond to the nonlinear and multi-scale characteristics present in resource consumption patterns. This often leads to over- or under-allocation of resources, reducing resource utilization and affecting the smoothness of the teaching experience.

[0004] Existing systems generally lack a closed-loop mechanism capable of autonomously learning and optimizing through continuous operation. Once deployed, their allocation strategies tend to have fixed parameters, failing to adapt to feedback and adjustments based on the actual effects of resource allocation. As teaching models evolve, new teaching tools are applied, and user behaviors change, existing strategies gradually become ineffective, leading to performance degradation over time. A dynamic allocation management technology is needed that can accurately characterize the inherent patterns of resource usage and possess continuous self-optimization capabilities to address the fundamental shortcomings of existing methods in terms of accurate prediction and long-term adaptability. Summary of the Invention

[0005] The purpose of this invention is to provide a machine learning-based dynamic allocation and management system for educational resources to solve the problems mentioned in the background.

[0006] To achieve the above objectives, this invention provides a machine learning-based dynamic allocation and management system for educational resources, the system comprising:

[0007] The data acquisition module is used to collect real-time resource consumption data streams through a monitoring agent embedded in the learning terminal. The real-time resource consumption data streams include resource request timestamp sequences, content transmission rate fluctuation curves, and device computing load trajectories.

[0008] The multi-resolution analysis module is used to perform multi-resolution analysis and processing on the real-time resource consumption data stream to generate a resource usage pattern map, which includes time-dimensional pattern clusters, spatial-dimensional pattern clusters and content-dimensional pattern clusters.

[0009] The resource demand mapping module is used to input the resource usage pattern map into the pre-trained deep belief network, perform resource demand mapping calculation, and output a resource quota suggestion value for each learning terminal.

[0010] The delay compensation module is used to perform transmission delay compensation processing on the resource quota recommendation value based on the network topology, and generate a delay-adapted resource allocation scheme.

[0011] The resource scheduling module is used to execute resource scheduling instructions according to the delayed resource allocation scheme and activate resource usage audit trail;

[0012] An adaptive optimization module is used to update the weight parameters of the deep belief network using audit trail data, thereby achieving adaptive optimization of the model.

[0013] Preferably, the data acquisition module collects real-time resource consumption data streams through a monitoring agent embedded in the learning terminal, including:

[0014] Configure the monitoring agent to capture resource request timestamp sequences at a fixed sampling frequency, wherein the resource request timestamp sequences record the initiation time and duration of each resource access event;

[0015] The content transmission rate fluctuation curve is extracted by a traffic sniffer, which represents the change in data transmission rate per unit time.

[0016] The device computational load trajectory is read from the device performance counter, and the device computational load trajectory reflects the timing data of CPU utilization, memory usage ratio and disk I / O pressure.

[0017] The acquired data stream is time-aligned and missing value interpolated to ensure the continuity and integrity of the data stream.

[0018] Preferably, the multi-resolution analysis module performs multi-resolution analysis processing on the real-time resource consumption data stream, including:

[0019] The wavelet transform algorithm is used to decompose the data stream into sub-signals at different time scales, and to extract short-term fluctuation components and long-term trend components.

[0020] Feature extraction is performed on the sub-signals at each time scale, and statistical features including mean, variance, skewness, and kurtosis are calculated.

[0021] A time-dimensional pattern cluster is constructed based on the feature extraction results, and the time-dimensional pattern cluster represents the periodic pattern of resource use;

[0022] The learning terminals are grouped by geographical location using a spatial clustering algorithm to generate spatial dimension pattern clusters that reflect regional differences in resource demand.

[0023] By using topic modeling techniques to analyze the distribution of content types, content dimension pattern clusters are generated to identify popular content resources.

[0024] Preferably, the resource demand mapping module inputs the resource usage pattern graph into the pre-trained deep belief network, including:

[0025] Deep belief networks are composed of multiple restricted Boltzmann machine layers stacked together, with each layer learning resources using different abstract features of the pattern graph;

[0026] During the training phase, historical resources are used to pretrain network parameters using data, and weights are optimized using a contrastive divergence algorithm.

[0027] In the mapping calculation stage, the time dimension pattern cluster, the spatial dimension pattern cluster, and the content dimension pattern cluster are used as input feature vectors;

[0028] The network output layer uses the Softmax activation function to generate a resource quota suggestion value for each learning terminal. The resource quota suggestion value is a continuous numerical vector.

[0029] Preferably, the delay compensation module performs transmission delay compensation processing based on the network topology, including:

[0030] Obtain the internal network topology of the educational institution and extract the link delay parameters and bandwidth capacity between router nodes;

[0031] Establish a delay propagation model to calculate the expected transmission delay from the server to the learning terminal for the resource quota recommendation value;

[0032] Adjust the recommended resource quota value based on the expected transmission delay, and increase the amount of redundant resource allocation for high-latency links;

[0033] A jitter buffering mechanism is introduced to compensate for the impact of latency fluctuations on the stability of resource allocation and to generate a latency-adapted resource allocation scheme.

[0034] Preferably, establishing the delay propagation model includes:

[0035] The network topology is abstracted as a directed graph, where nodes represent network devices and edges represent physical links;

[0036] A delay weight is assigned to each edge, and the delay weight is dynamically updated based on historical delay measurement data;

[0037] Dijkstra's algorithm is used to calculate the shortest latency path from the resource server to each learning terminal;

[0038] The expected transmission delay value is obtained by accumulating the delay weight along the shortest delay path.

[0039] Preferably, the resource scheduling module executes resource scheduling instructions including:

[0040] The resource scheduler parses the latency-adapted resource allocation scheme and generates specific resource allocation commands.

[0041] Resource allocation commands include bandwidth reservation commands, cache allocation commands, and computing resource reservation commands;

[0042] The software-defined network controller issues resource allocation commands to network devices.

[0043] Monitor the resource allocation process and record the deviation between the actual allocated resources and the planned value.

[0044] Preferably, the resource scheduling module activates resource usage audit tracking, including:

[0045] The audit trail module records the timestamp, execution terminal, and resource type for each resource allocation operation;

[0046] Collect resource utilization efficiency metrics, including resource utilization rate, response time, and error rate;

[0047] Audit trail data is stored in a time-series database for subsequent analysis, and audit reports are generated periodically to summarize the effectiveness of resource allocation.

[0048] Preferably, the adaptive optimization module updates the deep belief network using audit trail data, including:

[0049] Extract the discrepancy sequence between actual resource usage and recommended quota values ​​from audit trail data;

[0050] Calculate the loss function value of the difference sequence and backpropagate the error using the gradient descent algorithm;

[0051] Adjust the weight parameters of the deep belief network, minimize the loss function, and use the updated deep belief network to perform the next round of resource demand prediction, thus enabling the model to learn online.

[0052] Preferably, the system further includes a priority management module for performing the following processes: obtaining the user role identifier and resource access history of the learning terminal; determining the basic priority level based on the user role identifier; analyzing the access frequency and duration in the resource access history and calculating the dynamic priority adjustment factor; multiplying the basic priority level by the dynamic priority adjustment factor to generate a resource allocation priority weight; and weighting and fusing the resource allocation priority weight with the resource quota recommendation value to generate a priority-adjusted resource quota recommendation value.

[0053] Compared with the prior art, the beneficial effects of the present invention are:

[0054] Multi-resolution analysis of real-time resource consumption data streams enables the analysis of the periodicity of resource request timestamp sequences, the suddenness of content transmission rate fluctuations, and the trends in device computational load trajectories across different time scales. This technology integrates discrete events and continuous fluctuations into a unified resource usage pattern map encompassing a cluster of temporal, spatial, and content-based patterns. It reveals deep correlations that traditional single-point threshold monitoring cannot uncover. This multi-dimensional and multi-scale detailed characterization of resource usage behavior provides unprecedented insight into the true, dynamically changing nature of resource demand, advancing the basis of resource allocation from surface phenomena to the level of intrinsic patterns and laws.

[0055] By utilizing deep belief networks to process resource usage pattern graphs, their deep nonlinear structure can learn the complex mapping relationship between high-dimensional pattern features and optimal resource quotas. This mapping capability overcomes the limitations of shallow models such as linear regression or simple decision trees, accurately fitting the nonlinear and abrupt resource demands arising from the intertwined influence of multiple factors in teaching activities. Online updates to network weight parameters based on resource usage audit tracking data make the entire mapping model a dynamically evolving organism. The model can perceive resource usage pattern drifts caused by course content updates, adjustments to teaching pace, or changes in user habits, and tracks these changes by continuously fine-tuning its internal parameters. This mechanism ensures that resource quota recommendations are not only based on historical patterns but also respond in real-time to the current and future operational status of the system, achieving co-evolution between allocation strategies and the dynamic environment, thus maintaining high accuracy and adaptability of allocation decisions over a longer time horizon. Attached Figure Description

[0056] Figure 1 This is a schematic diagram illustrating the working principle of the machine learning-based dynamic allocation and management system for educational resources as described in this invention.

[0057] Figure 2 A flowchart for the multi-resolution analysis module to process real-time resource consumption data streams;

[0058] Figure 3 A flowchart illustrating the transmission delay compensation process performed by the delay compensation module;

[0059] Figure 4 This is a graph illustrating the effect of resource quota compensation based on network latency.

[0060] Figure 5 This is a chart for resource scheduling performance analysis and allocation deviation monitoring. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Please see Figure 1 This invention provides a machine learning-based dynamic allocation and management system for educational resources. The system includes: a data acquisition module, a multi-resolution analysis module, a resource demand mapping module, a delay compensation module, a resource scheduling module, and an adaptive optimization module. Specific implementation methods are as follows:

[0063] The data acquisition module continuously collects real-time resource consumption data streams through a monitoring agent embedded in each learning terminal. This data stream includes resource request timestamp sequences, content transmission rate fluctuation curves, and device computational load trajectories, providing raw input for subsequent analysis. The multi-resolution analysis module receives the data stream and processes it using multi-resolution analysis methods to generate a structured resource usage pattern map. This map includes time-dimensional, spatial-dimensional, and content-dimensional pattern clusters, thereby capturing multi-dimensional characteristics of resource usage. The resource demand mapping module takes the resource usage pattern map as input and performs resource demand mapping calculations through a pre-trained deep belief network, outputting a suggested resource quota value for each learning terminal. This value represents the expected resource allocation in the form of a continuous numerical vector. The latency compensation module, based on the network topology of the educational institution, performs transmission latency compensation processing on the suggested resource quota value. By modeling the impact of latency, it generates a latency-adapted resource allocation scheme, ensuring the feasibility of the allocation scheme in the actual network environment. The resource scheduling module executes specific resource scheduling instructions according to the latency-adapted resource allocation scheme, while simultaneously activating the resource usage audit trail function to record the allocation and execution process. The adaptive optimization module uses data collected from the audit trail to update the weight parameters of the deep belief network, achieving adaptive optimization of the model and enabling the system to dynamically adjust prediction accuracy based on actual usage. All modules are connected via data flow pipelines, forming a closed-loop control system. The output of the data acquisition module is fed to the multi-resolution analysis module, which in turn passes the output to the resource demand mapping module, and so on, until the adaptive optimization module completes its parameter update, thereby achieving continuous optimization of resource allocation.

[0064] Example 1: See Figure 2The data acquisition module collects real-time resource consumption data streams through a monitoring agent embedded in the learning terminal. In practice, the monitoring agent is configured with a fixed sampling frequency, the value of which is set according to the typical load characteristics of the teaching network environment, ranging from 10 milliseconds to 100 milliseconds. The monitoring agent is pre-installed or silently deployed as an independent software component on each learning terminal. It runs continuously in the background as a system service or daemon process, and its startup is synchronized with the learning terminal's operating system startup. The monitoring agent is configured to capture resource request timestamp sequences at a fixed sampling frequency. Recording these sequences is achieved by hooking system calls or intercepting network API functions. Whenever an application on the learning terminal initiates a resource access request, the monitoring agent intercepts the event. The resource request timestamp sequence records the initiation time and duration of each resource access event. The initiation time is recorded using a high-precision system clock, with timestamp accuracy down to the microsecond level. The duration is obtained by calculating the difference between the resource request completion time and the resource request initiation time. The resource request timestamp sequence is stored in a circular buffer in chronological order. The size of the circular buffer is configured according to the memory capacity. When the circular buffer reaches its capacity limit, the oldest timestamp record will be overwritten by the newest record.

[0065] Content transmission rate fluctuation curves are extracted using a traffic sniffer. In practice, the traffic sniffer is integrated into the monitoring agent. It captures all data packets flowing through the learning terminal by setting the network interface card to promiscuous mode. The content transmission rate fluctuation curve represents the change in data transmission rate per unit time. Each unit time is divided into fixed time windows, the length of which is coordinated with a fixed sampling frequency. The traffic sniffer counts the total number of bytes successfully received or sent within each time window, then divides the total number of bytes by the time window length to obtain the average transmission rate within that time window. The average transmission rates of consecutive time windows are connected chronologically to form the content transmission rate fluctuation curve. The traffic sniffer filters captured data packets, processing only network traffic related to access to teaching resources. Filtering rules are set based on target IP address ranges, port numbers, or protocol types. The content transmission rate fluctuation curve data is also temporarily stored in a circular buffer and synchronized with the resource request timestamp sequence.

[0066] The device computational load trajectory, which reads device performance counters, is a set of registers or software interfaces provided by the operating system kernel to record hardware resource usage. The monitoring agent reads these counters by calling the operating system's performance data query interface; this query is performed periodically at a fixed sampling frequency. The device computational load trajectory reflects time-series data on CPU utilization, memory usage, and disk I / O pressure. CPU utilization is obtained by reading the processor time percentage provided by the operating system's task manager; memory usage is calculated as the ratio of used physical memory to total physical memory; and disk I / O pressure is quantified by monitoring disk queue length and read / write operations per second. Each data point in the device computational load trajectory includes a timestamp and a set of performance counter readings, arranged in chronological order to form a time series. The monitoring agent optimizes the reading operation of the device performance counters, using batch reading to reduce system call overhead and avoid significant performance impact on the normal use of the learning terminal.

[0067] The collected data streams undergo time alignment and missing value imputation. In specific implementations, time alignment is performed on three types of data streams: resource request timestamp sequences, content transmission rate fluctuation curves, and device computational load trajectories. Since these three types of data streams may be independently collected by different sub-modules of the monitoring agent, their timestamp references may have slight deviations. Time alignment uses a unified reference time source, such as a system clock synchronized with the network time protocol, to convert the timestamps of all data streams onto the same time axis. Time alignment uses a fixed time interval as a reference, resampling each data stream onto a common time grid, with the common time grid interval consistent with the fixed sampling frequency. Missing value imputation handles data point loss caused by network jitter, excessive system load, or brief monitoring agent failures. The missing value imputation algorithm uses linear interpolation; for a missing value between two known data points, an estimated value is calculated based on the values ​​of the preceding and following data points in a time ratio. In some embodiments, for cases with multiple consecutive missing data points, a predictive imputation method based on historical data patterns is used. This predictive imputation method uses an autoregressive model to predict the missing value based on several data points preceding the missing point. Time alignment and missing value imputation are both completed before the data stream leaves the monitoring agent, ensuring the continuity and integrity of the data stream. The processed data stream is encapsulated into standard-format data packets, each containing a header and payload. The header records the learning terminal identifier, data stream type, time range, and data quality identifier, while the payload stores the processed sequence data. These data packets are transmitted to the central data processing server via an encrypted channel, providing high-quality input for subsequent multi-resolution analysis modules. The monitoring agent is designed with resource efficiency in mind, implementing data compression and selective upload strategies during data acquisition. For example, when resources are idle, the data upload frequency is reduced to minimize network bandwidth consumption.

[0068] Example 2: See Figure 3The multi-resolution analysis module performs multi-resolution analysis on the real-time resource consumption data stream. In its implementation, the multi-resolution analysis module receives a preprocessed data stream from the data acquisition module. This data stream includes a time-aligned resource request timestamp sequence, content transmission rate fluctuation curves, and device computational load trajectories. The multi-resolution analysis module uses a wavelet transform algorithm to decompose the data stream into sub-signals at different time scales. The wavelet transform algorithm uses the Daubechies wavelet basis function, with the order chosen as 4 based on the data characteristics to achieve multi-scale signal analysis. The wavelet transform decomposition process converts the original data stream into high-frequency short-term fluctuation components and low-frequency long-term trend components. The high-frequency short-term fluctuation components correspond to detailed information at smaller time scales, while the low-frequency long-term trend components correspond to approximate information at larger time scales. The decomposition level is set to 5 levels to cover fluctuation patterns from the second to the hour level. The mathematical expression for the wavelet transform can be expressed as:

[0069]

[0070] in: Indicates the first Wavelet coefficients at the layer decomposition scale, This represents the sequence of real-time resource consumption data streams input. This represents a discrete wavelet function based on the Daubechies wavelet basis function, with the scaling parameter being... And the translation parameters are , Indices representing data points. Wavelet coefficients. The absolute value of the value reflects the energy distribution of the signal at the corresponding scale and is used for subsequent feature extraction. The decomposed sub-signals are stored as a multi-dimensional array structure, with each dimension corresponding to a decomposition scale, which facilitates subsequent processing.

[0071] Feature extraction is performed on sub-signals at each time scale. The feature extraction process calculates statistical features for each sub-signal, including mean, variance, skewness, and kurtosis. The mean represents the average level of the sub-signal, the variance represents the degree of fluctuation, the skewness describes the asymmetry of the distribution, and the kurtosis describes the sharpness of the distribution. A sliding window method is used for feature extraction, with the window length matched to the time scale of the sub-signal; for example, a shorter window is used for high-frequency components, and a longer window is used for low-frequency components. The calculated feature values ​​are organized into feature vectors, the dimension of which is equal to the number of statistical features multiplied by the number of decomposition levels. These feature vectors serve as input data for constructing pattern clusters. Time-dimensional pattern clusters are constructed based on the feature extraction results using the K-means clustering algorithm. The number of clusters in the K-means clustering algorithm is determined by the elbow rule, which selects the inflection point as the optimal number of clusters based on the sum of squared errors curves under different cluster numbers. The time-dimensional pattern clusters represent the periodic patterns of resource usage, such as daily peak usage patterns or weekly off-peak patterns. The cluster center vector represents the typical usage pattern, and the cluster label is used to identify different pattern types. In some embodiments, the time-dimensional pattern clusters are updated using an incremental clustering algorithm, which allows new patterns to be added online without retraining the entire model, adapting to dynamic changes in resource usage patterns.

[0072] The generation of spatial dimension pattern clusters is achieved through a spatial clustering algorithm. This algorithm groups learning terminals according to their geographical location. Geographical location data is obtained from the IP address resolution of the learning terminals or from their built-in GPS modules, and the coordinate information is converted to a unified coordinate system such as WGS84. The spatial clustering algorithm uses the DBSCAN algorithm, which is based on density clustering and can handle clusters of arbitrary shapes and identify noise points. The parameters of the DBSCAN algorithm include the neighborhood radius and the minimum number of points. The neighborhood radius is set according to the geographical distribution of educational institutions, and the minimum number of points is adjusted according to the terminal density. Spatial dimension pattern clusters reflect the differences in regional resource demand, such as the different resource usage intensities of a teaching building area and a library area. Each cluster is associated with a set of learning terminal identifiers and geographical boundary information. Content dimension pattern clusters are analyzed using topic modeling techniques. Topic modeling uses an LDA model to process the distribution of resource content types. Content types are extracted from resource URLs, file extensions, or metadata to generate topic-word distributions and document-topic distributions. The hyperparameters of the LDA model are estimated using Gibbs sampling, and the number of topics is optimized using a perplexity index. Content dimension pattern clusters identify popular content resources such as video streams or document downloads, and the cluster probability represents the popularity of content types. It is understandable that the output of the multi-resolution analysis module is a resource usage pattern graph. The resource usage pattern graph integrates time-dimensional pattern clusters, spatial-dimensional pattern clusters, and content-dimensional pattern clusters into a unified data structure. The graph is stored in the form of a graph database, where nodes represent pattern clusters and edges represent relationships between clusters.

[0073] The resource demand mapping module inputs the resource usage pattern map into the pre-trained deep belief network. In practice, the deep belief network consists of multiple stacked Restricted Boltzmann Machine (RBM) layers. The number of RBM layers is set to three based on the problem complexity, with the number of nodes in each layer decreasing progressively to learn abstract features. Each RBM layer learns different abstract features from the pattern map. The bottom-layer RBMs process raw features such as statistical feature values, the middle-layer RBMs learn combined features, and the top-layer RBMs capture semantic features. The pre-training phase of the deep belief network uses historical resource usage data, which includes resource consumption records from the past few months. The data has been cleaned and standardized. Pre-training optimizes the weights using a contrastive divergence algorithm, which performs a one-step Gibbs sampling to approximate the gradient. The learning rate is set to 0.01, and a momentum term is used to accelerate convergence. The pre-training process is iterative until the reconstruction error stabilizes. During the mapping computation phase, resource use pattern graphs are used as input feature vectors. These input feature vectors are constructed by concatenating the center vectors of pattern clusters in the time dimension, the identifier encodings of pattern clusters in the spatial dimension, and the probability distributions of pattern clusters in the content dimension. The vector dimension is fixed at 100 dimensions to fit the network input layer. The forward propagation of the deep belief network computes the activation values ​​of each restricted Boltzmann machine layer using the Sigmoid function. The network output layer uses the Softmax activation function, which transforms the original output into a probability distribution. The network outputs a resource quota suggestion value for each learning terminal. The resource quota suggestion value is a continuous numerical vector, with vector elements corresponding to different resource types such as bandwidth, storage, and computing resources. The vector values ​​are normalized to the range of 0-1 to represent the allocation ratio.

[0074] Example 3: The delay compensation module performs transmission delay compensation on resource quota suggestions based on the network topology. In specific implementation, the delay compensation module receives resource quota suggestions from the resource demand mapping module. The resource quota suggestions are a resource allocation scheme for each learning terminal represented in vector form. The delay compensation module obtains the internal network topology map of the educational institution. The network topology map is periodically pulled from the network management system via SNMP protocol or RESTAPI interface. The network topology map includes the interconnection relationships of network devices such as routers, switches, access points, and servers. The network topology map is stored in memory as a directed graph data structure. The vertices of the directed graph represent network devices, and the edges represent physical links or logical connections. The link delay parameters and bandwidth capacity between router nodes are extracted. The link delay parameters are obtained by measuring the round-trip time by sending ICMP probe packets to each link endpoint. The bandwidth capacity is read from the network device configuration file or queried via the NETCONF protocol. The link delay parameters and bandwidth capacity are attached as attributes to the corresponding edges of the directed graph, forming a weighted directed graph data structure. The delay compensation module establishes a delay propagation model, which transforms a directed graph into a computational model to quantify the delay impact of resource quota recommendations being transmitted from the central server to each learning terminal.

[0075] The specific process of establishing a delay propagation model includes abstracting the network topology into a directed graph. The vertex set of the directed graph contains all network device nodes, and the edge set contains all valid connections. A delay weight is assigned to each edge of the directed graph. The delay weight is dynamically updated based on historical delay measurement data, which comes from continuously running network monitoring probes. These probes measure link delays at fixed intervals and record them in a time-series database. The delay weight is calculated using a sliding window averaging method, with the window size set to the 100 most recent measurements to reduce the impact of instantaneous fluctuations. Dijkstra's algorithm is used to calculate the shortest delay path from the resource server to each learning terminal. Dijkstra's algorithm uses the vertex where the resource server is located as the source point, traverses all vertices in the directed graph, and calculates the minimum delay path from the source point to each vertex. Dijkstra's algorithm maintains a priority queue to store vertices to be processed. The priority queue is sorted according to the current estimated delay. In each iteration, the vertex with the minimum delay is processed, and the delay estimates of its neighboring vertices are updated. The predicted transmission delay value is obtained by accumulating delay weights along the shortest delay path. The calculation process involves summing the delay weights of all edges along the path from the resource server vertex to the learning terminal vertex. The mathematical expression for the predicted transmission delay value can be represented as:

[0076]

[0077] in: Indicates from the resource server apex To the corresponding vertex of the learning terminal The expected transmission delay value, This represents the shortest delay path found using Dijkstra's algorithm. Representing a path One of the edges in, Representing an edge The delay weight assigned above. Expected transmission delay value. The unit is milliseconds, reflecting the time overhead of data packets being transmitted from the server to the learning terminal. Resource quota recommendations are adjusted based on the expected transmission delay value, using a linear compensation method to increase redundant resource allocation for high-latency links. The redundant resource allocation is calculated by multiplying the original quota recommendation value by a delay compensation coefficient, which is a monotonically increasing function of the expected transmission delay value.

[0078] In some embodiments, for links where the expected transmission delay exceeds a threshold, a nonlinear compensation strategy is employed. This strategy adds more redundant resources to cope with packet loss and retransmission when the delay is high. A jitter buffering mechanism is introduced to compensate for the impact of delay fluctuations on resource allocation stability. This mechanism sets up a dynamic buffer at the receiving end, with the buffer size adaptively adjusted based on the delay jitter variance. The delay jitter variance is calculated by statistically analyzing the standard deviation of delay values ​​over a recent period, and the buffer size is proportional to the delay jitter variance. The jitter buffering mechanism smooths out sudden delay changes caused by network congestion, ensuring the stability of the resource allocation scheme. A delay-adapted resource allocation scheme is generated, containing adjusted resource quota values ​​and buffer parameters, and encapsulated in a structured data format. This scheme is sent to the resource scheduling module via a message queue, with the message format including the learning terminal identifier, resource type, adjusted quota value, and effective timestamp. It is understood that the delay compensation module periodically recalculates the expected transmission delay value. Recalculation trigger conditions include network topology changes, significant changes in delay measurements, or the expiration of fixed time intervals.

[0079] See Figure 4 This diagram illustrates the impact of network transmission latency on resource quota allocation and the corresponding compensation effect. The figure clearly presents the difference between the original suggested resource quota value and the final quota after latency adaptation through a comparative bar chart, while a line graph reflects the trend of network transmission latency from the central server to each learning terminal. It can be observed that as transmission latency increases, the system automatically applies compensation strategies to increase resource allocation, especially on high-latency links, where the compensation mechanism allocates more redundant resources to ensure service quality. This dynamic compensation mechanism effectively addresses the impact of network condition fluctuations on resource allocation stability, ensuring that learning terminals receive reliable resource support in different network environments.

[0080] Example 4: The resource scheduling module executes resource scheduling instructions according to the latency-adapted resource allocation scheme. In specific implementation, the resource scheduling module receives the latency-adapted resource allocation scheme from the latency compensation module. The latency-adapted resource allocation scheme is encapsulated in JSON or XML format, including the learning terminal identifier, resource type, adjusted quota value, and effective timestamp. The resource scheduler parses the latency-adapted resource allocation scheme. The parsing process includes syntax verification and semantic checks to ensure that the scheme data is complete and conforms to the predefined pattern. The parsed data is converted into an internal data structure to facilitate the generation of specific resource allocation commands. The resource allocation commands include bandwidth reservation instructions, cache allocation instructions, and computing resource reservation instructions. The bandwidth reservation instructions specify the minimum and maximum network bandwidth reserved for each learning terminal. The cache allocation instructions allocate storage space on the edge server for caching popular resources. The computing resource reservation instructions allocate virtual machine or container resources to ensure the priority of computing tasks. Referring to Table 1, when generating resource allocation commands, the resource scheduler refers to the resource type mapping table, which converts abstract resource types into parameters that can be recognized by specific network devices.

[0081] Table 1: Resource Allocation Command Parameter Mapping Table

[0082]

[0083] The software-defined network controller (SDN) issues resource allocation commands to network devices. The SDN communicates with network devices using the OpenFlow protocol or other southbound interfaces, issuing flow table rules to reserve bandwidth. These rules match the IP addresses and port numbers of the learning terminals, and action fields set bandwidth limits or priority queues. For cache allocation commands, the SDN interacts with the storage management system, allocating a specified size of cache space and setting caching policies such as LRU or LFU via API calls. Compute resource reservation commands are implemented through the cloud management platform interface, creating or adjusting the resource configuration of virtual machine instances. The resource allocation execution process is monitored, with monitoring data collected from network devices, servers, and the cloud platform. The deviation between the actual allocated resource amount and the planned value is recorded, calculated using the following formula:

[0084]

[0085] in: Indicates resource type The percentage of allocation deviation, Indicates the actual amount of resources allocated. This indicates the planned allocation of resources and the actual allocation of resources. Read the planned resource allocation amount in real time from the device counter. The resource allocation scheme is adapted for latency. An alarm is triggered when the deviation exceeds a threshold, and the alarm information is logged to the log system for subsequent analysis. The resource scheduling module activates resource usage audit trails. The audit trail module runs as an independent subsystem, monitoring resource allocation operation events, including command issuance, execution start, and completion times. The audit trail module records the timestamp, execution terminal, and resource type for each resource allocation operation. The timestamp is synchronized to Coordinated Universal Time using a high-precision clock. The execution terminal identifier is extracted from the resource allocation command. Resource types are categorized as bandwidth, cache, or compute resources. Recorded data is stored in key-value pairs, with each operation corresponding to one audit record. Each audit record includes the operation ID, timestamp, terminal ID, resource type, and operation status.

[0086] Resource utilization efficiency metrics are collected, including resource utilization rate, response time, and error rate. Resource utilization rate is calculated as the ratio of actual resource consumption to allocated resources. Response time measures the time interval from resource request to response reception. Error rate is the proportion of failed operations to total operations. Metric data is collected from multiple sources: resource utilization rate is obtained from network device interface counters and system performance monitors; response time is measured by injecting probe packets or analyzing application logs; and error rate is statistically analyzed from error logs and transaction rollback records. The collection process uses an asynchronous approach to avoid impacting the main workflow. The aggregated metric data generates a time-series dataset. Audit trail data is stored in a time-series database. The data is partitioned and indexed by time to support efficient querying. Each data point includes a timestamp, metric name, and numerical label. The storage process includes data cleaning and compression. Cleaning rules remove outliers, and compression algorithms reduce storage space usage. Audit reports are generated periodically. The audit report generator runs at fixed intervals, such as daily or weekly, summarizing resource allocation effectiveness. Report content includes allocation accuracy statistics, efficiency trend analysis, and a list of anomalies. Report formats support PDF and HTML, and are distributed via email or a web portal. Audit reports are used for system optimization and compliance checks.

[0087] See Figure 5 This presentation showcases the performance and allocation accuracy monitoring results of the resource scheduling module. Using a combined bar chart, it simultaneously presents the average allocation amount and allocation deviation rate for different types of resources, comprehensively reflecting the execution effect of the resource scheduling process. The charts clearly show the performance differences of different resource categories, such as bandwidth, cache, and computing resources, during the scheduling process. The allocation amount indicator reflects the scale of resource scheduling, while the deviation rate measures the degree of agreement between actual and planned allocations. This visualization analysis helps identify performance bottlenecks and optimization directions in resource scheduling, providing important reference for continuous system improvement and ensuring that educational resources are efficiently and accurately allocated to various learning terminals.

[0088] Example 5: The adaptive optimization module updates the deep belief network using audit trail data. In specific implementation, the adaptive optimization module extracts historical operation records from the audit trail database of the resource scheduling module. The audit trail database stores resource allocation operations and actual usage data in time-series format. It extracts the difference sequence between actual resource usage and the recommended quota value from the audit trail data. The extraction process is implemented through a sliding query within a time window. Each time window corresponds to a complete resource allocation cycle. Each data point in the difference sequence represents the numerical difference between the actually reported resource consumption at a specific time point and the recommended quota value predicted by the deep belief network. The difference sequences are grouped according to the learning terminal identifier and resource type, forming multiple independent time-series datasets. Each dataset corresponds to the historical prediction and performance deviation of a resource type for a learning terminal.

[0089] The loss function value of the discrepancy sequence is calculated using the Huber loss function. The Huber loss function combines the advantages of mean squared error and absolute error, providing a linear penalty to reduce the impact of outliers when the error is large, and a squared penalty to maintain smoothness when the error is small. The mathematical expression of the Huber loss function involves a threshold parameter δ. When the absolute error is less than or equal to δ, a squared term is used for calculation; when the absolute error is greater than δ, a linear term is used. The loss function value is calculated based on the current batch of discrepancy sequence data. The batch size is dynamically adjusted according to the system load, typically set to the discrepancy data from the most recent 100 time points. The gradient descent algorithm is used for backpropagation of the error. The gradient descent algorithm uses a random mini-batch update strategy, randomly sampling a mini-batch of samples from the discrepancy sequence in each iteration to calculate the gradient of the loss function with respect to the weights of the deep belief network. Gradient calculation is implemented through backpropagation, starting from the output layer of the deep belief network and calculating the partial derivative of each weight parameter with respect to the loss function layer by layer. The partial derivative represents the direction and magnitude of the weight adjustment. The learning rate of the gradient descent algorithm adopts an adaptive adjustment strategy, with an initial learning rate set to 0.001, which is dynamically decayed based on convergence.

[0090] The weight parameters of the deep belief network are adjusted, following the gradient descent principle. The new weight value equals the original weight value minus the learning rate multiplied by the gradient value. The weight adjustment process is performed layer-by-layer, starting from the highest hidden layer of the deep belief network and propagating downwards to the input layer. The weights and biases of each restricted Boltzmann machine layer are updated separately. An L2 regularization term is added during the update process to prevent overfitting; the L2 regularization coefficient is set to 0.01. Minimizing the loss function is the ultimate goal of the weight adjustment. The optimization process is iterative until the loss function converges to a stable range or the maximum number of iterations is reached. The updated deep belief network is used for the next round of resource demand prediction, enabling online model learning. During online learning, newly generated audit trail data is continuously added to the training set, while older data has its weights decayed over time, ensuring the model adapts to the latest resource usage patterns.

[0091] The system also includes a priority management module, which works in conjunction with the adaptive optimization module. The priority management module receives input from the identity management system and resource access logs. It acquires user role identifiers and resource access history records for each learning terminal. User role identifiers are synchronized from the unified identity authentication system and include categories such as student, teacher, and administrator. Resource access history records are extracted from application layer logs, recording the resource request time, resource type, and access duration for each learning terminal. A basic priority level is determined based on the user role identifier. This basic priority level is obtained by querying a preset priority mapping table, where different user roles correspond to different basic priority values; for example, the basic priority for a teacher is 0.8, and for a student it is 0.5. The system analyzes the access frequency and duration in the resource access history. Access frequency is calculated by counting the number of resource requests within a specified time window, and duration is calculated by determining the average active duration of each session. The analysis uses a sliding time window algorithm, with the window size typically set to 24 hours. A dynamic priority adjustment factor is calculated based on the access frequency and duration. This dynamic priority adjustment factor is normalized to convert the original indicators into values ​​between 0 and 1. The normalization formula uses a min-max scaling method to ensure the comparability of indicators with different dimensions. The base priority level is multiplied by the dynamic priority adjustment factor to generate a resource allocation priority weight. The multiplication operation is performed element-wise, resulting in a comprehensive priority score for each learning terminal. The resource allocation priority weight is then weighted and fused with the resource quota recommendation value to generate a priority-adjusted resource quota recommendation value. The weighted fusion formula is: the priority-adjusted resource quota recommendation value equals the original resource quota recommendation value multiplied by the resource allocation priority weight, where the multiplication operation is performed element-wise along the resource dimension.

[0092] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0093] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machine learning-based dynamic allocation and management system for educational resources, characterized in that, The system includes: The data acquisition module is used to collect real-time resource consumption data streams through a monitoring agent embedded in the learning terminal. The real-time resource consumption data streams include resource request timestamp sequences, content transmission rate fluctuation curves, and device computing load trajectories. The multi-resolution analysis module is used to perform multi-resolution analysis processing on the real-time resource consumption data stream, including: The wavelet transform algorithm is used to decompose the real-time resource consumption data stream into sub-signals of different time scales, and extract short-term fluctuation components and long-term trend components. Feature extraction is performed on the sub-signals at each time scale, and statistical features including mean, variance, skewness, and kurtosis are calculated. A time-dimensional pattern cluster is constructed based on the feature extraction results, and the time-dimensional pattern cluster represents the periodic pattern of resource use; The learning terminals are grouped by geographical location using a spatial clustering algorithm to generate spatial dimension pattern clusters that reflect regional differences in resource demand. Utilize topic modeling techniques to analyze content type distribution, generate content dimension pattern clusters, and identify popular content resources; Generate a resource usage pattern map, which includes time-dimensional pattern clusters, spatial-dimensional pattern clusters, and content-dimensional pattern clusters; The resource demand mapping module is used to input the resource usage pattern map into a pre-trained deep belief network, perform resource demand mapping calculations, and output a resource quota suggestion value for each learning terminal, including: Deep belief networks are composed of multiple restricted Boltzmann machine layers stacked together, with each layer learning resources using different abstract features of the pattern graph; During the training phase, historical resources are used to pretrain network parameters using data, and weights are optimized using a contrastive divergence algorithm. In the mapping calculation stage, the time dimension pattern cluster, the spatial dimension pattern cluster, and the content dimension pattern cluster are used as input feature vectors; The network output layer uses the Softmax activation function to generate a resource quota suggestion value for each learning terminal. The resource quota suggestion value is a continuous numerical vector. The delay compensation module is used to perform transmission delay compensation processing on the resource quota recommendation value based on the network topology, and generate a delay-adapted resource allocation scheme. The resource scheduling module is used to execute resource scheduling instructions according to the delayed resource allocation scheme and activate resource usage audit trail; An adaptive optimization module is used to update the weight parameters of the deep belief network using audit trail data to achieve adaptive model optimization, including: Extract the sequence of differences between actual resource usage and recommended resource quota values ​​from audit trail data; Calculate the loss function value of the difference sequence and backpropagate the error using the gradient descent algorithm; Adjust the weight parameters of the deep belief network, minimize the loss function, and use the updated deep belief network. The network will conduct the next round of resource demand forecasting to enable online model learning.

2. The machine learning-based dynamic allocation and management system for educational resources according to claim 1, characterized in that, The data acquisition module collects real-time resource consumption data streams through a monitoring agent embedded in the learning terminal, including: Configure the monitoring agent to capture resource request timestamp sequences at a fixed sampling frequency, wherein the resource request timestamp sequences record the initiation time and duration of each resource access event; The content transmission rate fluctuation curve is extracted by a traffic sniffer, which represents the change in data transmission rate per unit time. The device computational load trajectory is read from the device performance counter, and the device computational load trajectory reflects the timing data of CPU utilization, memory usage ratio and disk I / O pressure. Time alignment and missing value interpolation are performed on the real-time resource consumption data stream to ensure its continuity and integrity.

3. The machine learning-based dynamic allocation and management system for educational resources according to claim 1, characterized in that, The delay compensation module performs transmission delay compensation processing based on the network topology, including: Obtain the internal network topology of the educational institution and extract the link delay parameters and bandwidth capacity between router nodes; Establish a delay propagation model to calculate the expected transmission delay from the server to the learning terminal for the resource quota recommendation value; Adjust the recommended resource quota value based on the expected transmission delay, and increase the amount of redundant resource allocation for high-latency links; A jitter buffering mechanism is introduced to compensate for the impact of latency fluctuations on the stability of resource allocation and to generate a latency-adapted resource allocation scheme.

4. The machine learning-based dynamic allocation and management system for educational resources according to claim 3, characterized in that, The establishment of the delayed propagation model includes: The network topology is abstracted as a directed graph, where nodes represent network devices and edges represent physical links; A delay weight is assigned to each edge, and the delay weight is dynamically updated based on historical delay measurement data; Dijkstra's algorithm is used to calculate the shortest latency path from the resource server to each learning terminal; The expected transmission delay value is obtained by accumulating the delay weight along the shortest delay path.

5. The machine learning-based dynamic allocation and management system for educational resources according to claim 1, characterized in that, The resource scheduling module executes resource scheduling instructions including: The resource scheduler parses the latency-adapted resource allocation scheme and generates specific resource allocation commands. Resource allocation commands include bandwidth reservation commands, cache allocation commands, and computing resource reservation commands; The software-defined network controller issues resource allocation commands to network devices. Monitor the resource allocation process and record the deviation between the actual allocated resources and the planned value.

6. The machine learning-based dynamic allocation and management system for educational resources according to claim 1, characterized in that, The resource scheduling module activates resource usage audit tracking, including: The audit trail module records the timestamp, execution terminal, and resource type for each resource allocation operation; Collect resource utilization efficiency metrics, including resource utilization rate, response time, and error rate; Audit trail data is stored in a time-series database for subsequent analysis, and audit reports are generated periodically to summarize the effectiveness of resource allocation.

7. The machine learning-based dynamic allocation and management system for educational resources according to claim 1, characterized in that, The system also includes a priority management module, which performs the following processes: obtaining the user role identifier and resource access history of the learning terminal; determining the basic priority level based on the user role identifier; analyzing the access frequency and duration in the resource access history, and calculating the dynamic priority adjustment factor. Multiply the base priority level by the dynamic priority adjustment factor to generate the resource allocation priority weight; then weight and fuse the resource allocation priority weight with the resource quota recommendation value to generate the priority-adjusted resource quota recommendation value.

Images