Dynamic feedback type resource management system for education research platform

Abstract

The invention relates to the technical field of education informatization, and particularly discloses a dynamic feedback type resource management system for an education research platform. The system comprises a real-time data acquisition module, a dynamic feature extraction module, a strategy generation module, a feedback evaluation module and a resource scheduling execution module. By constructing a complete technical chain of real-time data acquisition, dynamic feature extraction, online strategy generation, multi-dimensional feedback evaluation and resource scheduling execution, closed-loop dynamic optimization of the educational research platform resource management process is realized. The system can sensitively capture short-time demand fluctuation caused by environmental changes, interest transfer or emergencies in research activities, and adjusts a resource allocation strategy on a minute-level time scale through an online learning mechanism, thereby effectively solving the problem of strategy lag caused by offline training and updating of a traditional system.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of educational informatization, and particularly relates to a dynamic feedback type resource management system for an educational research and learning platform. BACKGROUND

[0002] The field of educational informatization is committed to improving the efficiency and quality of teaching and learning processes through technical means, and the management and allocation of educational resources is one of the core research directions. As an important carrier connecting theoretical learning and practical experience, the intelligent level of the resource management system of the educational research and learning platform directly affects the organization efficiency and learning effect of the research and learning activities.

[0003] Among them, the dynamic resource management for the educational research and learning platform is the key direction of current technology development, and the goal is to dynamically adjust and allocate various educational resources according to the actual needs of research and learning activities, in order to support personalized and situational learning experience.

[0004] The existing technology usually adopts a machine learning model based on historical data to make resource allocation decisions. These models rely on daily offline training and updating mechanism, and cannot capture the short-term behavior pattern mutations caused by sudden events, changes in student interest or environment in the research and learning scene in real time, resulting in a serious disconnection between resource allocation strategies and actual needs. At the same time, the existing system lacks the ability to process dynamic feedback information in real time, and it is difficult to effectively modify the allocation strategy within the resource scheduling period, causing structural contradictions between resource idling and resource shortage, which seriously restricts the service quality and user experience of the research and learning platform.

[0005] Therefore, there is an urgent need for a resource management system that can realize real-time perception and dynamic scheduling to solve the above technical problems. SUMMARY

[0006] The technical problem to be solved by the present application is to overcome the lag of resource allocation strategy and structural contradiction caused by the inability of the existing resource management system of the educational research and learning platform to respond to dynamic demand changes in real time and lack of immediate feedback correction capability. The purpose of the present application is to provide a dynamic feedback type resource management system for an educational research and learning platform, which realizes online adaptive optimization of resource allocation strategy by constructing a multi-level real-time perception network and a closed-loop feedback control architecture.

[0007] The technical scheme provided by the present application is a dynamic feedback type resource management system for an educational research and learning platform, which comprises a real-time data acquisition module, a dynamic feature extraction module, a strategy generation module, a feedback evaluation module and a resource scheduling execution module. The real-time data acquisition module is responsible for continuously collecting multi-dimensional data streams from user terminals, environmental sensors and platform operation databases, including student behavior logs, device usage status, environmental parameters, resource inventory and reservation queue data.

[0008] The dynamic feature extraction module is connected to the output end of the real-time data acquisition module, performs time window division and feature transformation on the input data stream, and extracts a key feature vector representing the dynamic change of resource demand, wherein the time window length is set as a configurable parameter of 5 minutes to 30 minutes. The strategy generation module receives the feature vector output by the dynamic feature extraction module, and updates the resource allocation strategy model in real time through an online learning algorithm. The model is constructed using a deep reinforcement learning framework, wherein the state space is defined as the current resource demand distribution and available resource capacity, and the action space is defined as the allocation ratio adjustment instruction of each type of resource.

[0009] The feedback evaluation module monitors the system state change after the strategy is executed, calculates the resource utilization rate, response delay and user satisfaction indicators, and generates a strategy effect evaluation report. The resource scheduling execution module generates specific resource allocation instructions based on the latest strategy output by the strategy generation module and the correction signal provided by the feedback evaluation module, and issues the instructions to the platform bottom control system.

[0010] Further, the real-time data acquisition module includes three types of data acquisition subunits. The user behavior acquisition subunit records the interactive operation sequence of students on the research platform through the burying point technology, including page stay time, content click frequency and task completion progress. The device state acquisition subunit establishes a communication connection with the physical device through the Internet of Things protocol, and periodically reports the device working state, geographic position and remaining available time. The environment perception acquisition subunit integrates temperature, humidity and light intensity sensors, collects environment data every 10 seconds and uploads them to the platform data hub.

[0011] Further, the feature extraction process performed by the dynamic feature extraction module includes sliding window statistics, time series pattern recognition and abnormal fluctuation detection. The sliding window statistics calculates the mean, variance and change trend of the resource request amount in each time window.

[0012] The time series pattern recognition uses an adaptive piecewise aggregation approximation algorithm to reduce the dimensionality and extract patterns from long time series data. The abnormal fluctuation detection identifies sudden demand signals that exceed the normal fluctuation range based on the Isolation Forest algorithm, and marks these signals as high-priority features.

[0013] Further, the online learning algorithm of the strategy generation module uses the Proximal Policy Optimization method, and its learning process includes two alternating stages of policy evaluation and policy improvement. The policy evaluation stage estimates the value function of the current policy through Monte Carlo sampling, and the policy improvement stage updates the policy network parameters through the gradient ascent method. The policy network performs model updating every 5 minutes, and the importance sampling technique is used during the updating process to ensure the stability of the learning process.

[0014] Further, the feedback evaluation module constructs a multi-objective evaluation function, which includes three core evaluation dimensions. The resource utilization efficiency dimension calculates the average utilization rate and idle rate of various resources, the response timeliness dimension measures the average time delay from resource request to resource allocation, and the user satisfaction dimension indirectly calculates the satisfaction score by analyzing user feedback data and behavior data. The evaluation results of the three dimensions are fused into a single comprehensive evaluation score through weighted summation, and the weight coefficients are dynamically adjusted according to the platform operation strategy.

[0015] Further, the resource scheduling execution module executes different scheduling logic according to the type of input instruction. For proportional adjustment instructions, the module reallocates resource quotas in the available resource pool according to the specified proportion. For priority adjustment instructions, the module reorders the resource request queue according to the preset priority rules. For emergency scheduling instructions, the module starts the fast channel mechanism to complete the emergency allocation and deployment of resources within 2 minutes.

[0016] Further, the system also includes a strategy cache and rollback mechanism. The strategy cache unit stores all strategy versions used in the last 24 hours and their performance evaluation data. When the comprehensive evaluation score of a new strategy is lower than the average level of historical strategies, the system automatically triggers a strategy rollback operation and switches to the historical strategy version with the best performance to continue running.

[0017] Further, the system adopts a distributed micro-service architecture for deployment, and the modules communicate asynchronously through a message queue. The real-time data acquisition module and the dynamic feature extraction module are deployed on edge computing nodes, the strategy generation module and the feedback evaluation module are deployed in the cloud computing center, and the resource scheduling execution module is directly embedded in the platform core business system, forming a collaborative computing paradigm of edge perception, cloud decision-making, and local execution.

[0018] Compared with the prior art, the beneficial effects of the present application are:

[0019] 1. The present application realizes the closed-loop dynamic optimization of the resource management process of the education research and learning platform by constructing a complete technical chain of real-time data acquisition, dynamic feature extraction, online strategy generation, multi-dimensional feedback evaluation, and resource scheduling execution. The system can accurately capture short-term demand fluctuations caused by environmental changes, interest shifts, or unexpected events in research and learning activities, and adjust resource allocation strategies on a minute time scale through online learning mechanisms, effectively solving the strategy lag problem caused by offline training and updating in traditional systems.

[0020] 2、Feedback evaluation module provides multi-dimensional performance indicators for strategy optimization, and strategy cache and rollback mechanism ensures the decision reliability of the system in complex dynamic environment. This real-time perception, fast response and continuous optimization of resource management mode significantly improves the utilization rate of educational resources and the service quality of research and learning activities, and provides an innovative technical solution for the field of educational information. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 is the overall technical scheme architecture diagram of the dynamic feedback type resource management system for the education research and learning platform provided by the application;

[0022] Figure 2 is the core principle framework diagram of the online learning algorithm of the strategy generation module in the application, which adopts the proximal policy optimization method;

[0023] Figure 3 is the multi-level interaction relationship and data flow diagram of the real-time data acquisition module and the dynamic feature extraction module in the application;

[0024] Figure 4 is the logic flow framework diagram of the multi-objective evaluation function constructed by the feedback evaluation module in the application. DETAILED DESCRIPTION

[0025] Please refer to the attached Figures 1 to 4 The embodiment details a specific implementation of the dynamic feedback type resource management system for the education research and learning platform. The system constructs a complete closed-loop control architecture from data perception to decision execution, and its core goal is to realize the minute-level adaptive optimization of the education resource allocation strategy to cope with the changing demand dynamics in research and learning activities.

[0026] The system adopts a distributed micro-service architecture for deployment, in which the real-time data acquisition module and the dynamic feature extraction module are deployed on the edge computing nodes close to the data source, the strategy generation module and the feedback evaluation module are deployed in the cloud computing center with strong computing power, and the resource scheduling execution module is directly embedded in the core business system of the platform, forming a collaborative computing paradigm of edge side responsible for real-time perception and preliminary processing, cloud side responsible for core decision and evaluation, and platform side responsible for final execution.

[0027] The modules communicate asynchronously through a high-throughput message queue to ensure efficient and reliable transmission of data flow and instruction flow in the system, while ensuring the scalability and fault tolerance of the system as a whole.

[0028] The real-time data acquisition module, acting as the system's sensory nerve endings, is responsible for continuously collecting multi-dimensional, high-frequency data streams from user terminals, environmental sensors throughout the study tour setting, and the platform's operational database. This module is further divided into three functionally independent but collaborative data acquisition sub-units. The user behavior acquisition sub-unit records every interactive sequence of students on the study tour platform in real time by embedding precisely defined tracking code in the front-end page and application interface.

[0029] These operation sequences are transformed into structured log data, whose key fields include, but are not limited to, student identifiers, accessed page Uniform Resource Locators (URLs), page entry and exit timestamp sequences, click event types and frequencies for specific teaching content or interactive elements, and the percentage of task or experiment completion progress. All tracked data is pushed in real-time to the message middleware of the edge computing nodes in the form of event streams. The device status acquisition subunit establishes persistent communication connections with the physical devices managed by the research and learning platform through standard IoT protocols, such as message queue telemetry transmission protocols or restricted application protocols. These physical devices may include experimental instruments, display terminals, mobile learning terminals, etc.

[0030] This subunit proactively sends status query commands to all online devices every 30 seconds and receives working status reports from the devices. The reports include the device's current operating mode, remaining battery power or continuous working time, real-time geographical coordinates, network connection quality indicators, and any alarm information generated by device self-tests.

[0031] The environmental sensing and acquisition subunit integrates a series of high-precision sensors, including temperature, humidity, and light intensity sensors. These sensors are typically deployed as IoT nodes in the actual physical spaces of the study tour, such as classrooms, laboratories, or outdoor campsites. This subunit synchronously collects readings from all sensors at fixed 10-second intervals and packages this environmental parameter data, along with sensor device identifiers and timestamps, uploading it to the platform's data hub for persistent storage and subsequent processing.

[0032] Please refer to the attached document. Figure 3 The multi-dimensional data stream output by the real-time data acquisition module is seamlessly integrated into the dynamic feature extraction module. This module serves as a crucial bridge connecting raw data and intelligent decision-making. Its core task is to extract key feature vectors that can characterize the dynamic changes in resource demand from massive, high-speed raw data.

[0033] The dynamic feature extraction module first divides the input data stream into time windows. The window length is a configurable parameter, allowing system administrators to flexibly set it between 5 and 30 minutes based on actual business scenarios. Within each sliding time window, the module performs three core feature extraction processes.

[0034] The first process is sliding window statistics. For numerical sequence data such as resource request volume, the module calculates in real time the total number of all request events within the current window, the arithmetic mean of the request volume, and the variance of the request volume over time. It then uses linear regression to fit the slope of the change trend of the request volume within the window period. This slope value can intuitively reflect whether the demand is growing rapidly, remaining stable, or declining.

[0035] The second process is temporal pattern recognition. For long-term series data such as user behavior sequences, the module employs an adaptive segmented aggregation approximation algorithm for data dimensionality reduction and pattern extraction. This algorithm first divides the long-term series into multiple intervals of variable length, then calculates the average value of the data within each interval, using these average value sequences to approximate the original sequence, thereby capturing periodic or steady-state patterns in user behavior.

[0036] The third process is abnormal fluctuation detection. This process is based on the Isolation Forest algorithm, which constructs multiple isolation trees by randomly selecting features and split points, thereby quickly isolating a minority of points whose behavior is significantly different from the majority of data points. The module uses this algorithm to identify sudden demand signals that occur within a short period of time, exceeding the historical normal fluctuation range. Once such a signal is detected, the module immediately assigns it a high-priority feature label and correspondingly increases the weight of its feature vector.

[0037] The dynamic feature extraction module ultimately outputs a set of high-dimensional, dense feature vectors. These vectors comprehensively represent the resource demand status, user behavior tendencies, and environmental context of the entire study platform within the most recent time window. These feature vectors are encapsulated into standard data packets and sent to the strategy generation module in real time via a message queue.

[0038] Please refer to the attached document. Figure 2 The strategy generation module is the core of the intelligent decision-making process of the entire system. It receives real-time feature vectors from the dynamic feature extraction module and dynamically updates the resource allocation strategy model based on these vectors using an online learning algorithm. This model is constructed using a deep reinforcement learning framework. Its state space is explicitly defined as a joint representation of the demand distribution of all resource types at the current moment and the available resource capacity of the system, while the action space is defined as a set of instructions for adjusting the allocation ratio of all allocable resource types.

[0039] The online learning algorithm used in the policy generation module is the proximal policy optimization method, an advanced policy gradient algorithm renowned for its excellent learning stability and sample efficiency. Its learning process comprises two alternating phases:

[0040] The strategy evaluation phase and the strategy improvement phase. In the strategy evaluation phase, the system uses the Monte Carlo sampling method to simulate multiple sets of resource allocation actions generated by the current policy network and observes the cumulative rewards brought by these actions to estimate the value function of the current policy, that is, to evaluate the long-term expected return of implementing the policy under the current system state.

[0041] In the policy improvement phase, the system utilizes the value estimate obtained in the policy evaluation phase to update the parameters of the policy network using a gradient ascent method. Its core update objective is to steadily increase the expected return of the policy while ensuring that the difference between the new and old policies is not too large. The loss function of the proximal policy optimization method is... Defined as:

[0042] ;

[0043] in, For time steps Expectation operation, The parameters representing the policy network, This indicates the difference between the new strategy and the old strategy at time step. The probability ratio of taking an action. The advantage value is estimated through the advantage function and measures how good or bad a particular action is compared to the average action in a specific state. It is a hyperparameter. The function will calculate the probability ratio. Limited to the range This achieves stability constraints for policy updates.

[0044] The entire policy network updates its model parameters according to a preset cycle, typically every 5 minutes.

[0045] During each update, the system employs importance sampling to extract a batch of historical interaction data from the experience replay buffer for gradient calculation, further ensuring data efficiency and stability during the learning process. Once the update is complete, the new policy model takes effect immediately, replacing the old model to generate subsequent resource allocation instructions.

[0046] The latest resource allocation strategy output by the strategy generation module is essentially a series of specific adjustment instructions, such as increasing the allocation ratio of experimental device A from the current 20% to 30%, or reordering the priority weights of computing resources. These instructions are simultaneously sent to the resource scheduling execution module and the feedback evaluation module.

[0047] Please refer to the attached document. Figure 4 The feedback evaluation module acts as a "prosecutor" of system performance, continuously monitoring changes in the overall platform system state after strategy execution and quantitatively calculating the actual effect of the strategy. This module constructs a multi-objective evaluation function containing three core evaluation dimensions.

[0048] The first dimension is resource utilization efficiency. This dimension calculates the average utilization rate of various resources within the evaluation period, such as the past 15 minutes, by monitoring the actual usage logs of all resource instances. This is the ratio of actual usage time to total available time, along with the corresponding idle rate. It also analyzes whether resources are overused or experiencing excessively long queues.

[0049] The second dimension is response time. This dimension precisely measures the average time latency from when a user or system issues a resource request until the resource is successfully allocated and ready for use. This latency is broken down into several sub-items such as processing latency, queuing latency, and scheduling latency for detailed statistical analysis.

[0050] The third dimension is user satisfaction. Since it is difficult to directly obtain real-time user satisfaction scores, this module uses an indirect estimation method. By analyzing user feedback data, such as ratings and complaints, as well as behavioral data, such as task completion rate, page exit rate, and willingness to repeat use, a pre-trained regression model is used to calculate a user satisfaction score between 0 and 1.

[0051] Ultimately, the evaluation results from these three dimensions—resource utilization efficiency score, response timeliness score, and user satisfaction score—will be weighted and summed into a single comprehensive evaluation score. The weighting coefficients are not fixed but dynamically adjusted by the platform operations administrator based on the platform's operational strategy priorities at different times. For example, during peak periods of study tours, more emphasis may be placed on response timeliness, while during off-peak periods, more attention may be paid to resource utilization efficiency.

[0052] The feedback evaluation module generates a strategy effectiveness evaluation report periodically, for example, every 10 minutes. This report includes not only the overall evaluation score but also a detailed list of the scores for each sub-item and their changing trends. This report is fed back to the strategy generation module as an important input for its subsequent strategy optimization. More importantly, the feedback evaluation module compares its calculated overall evaluation score with historical data stored in the strategy cache unit.

[0053] The system's built-in policy caching and rollback mechanisms are key safety valves for ensuring the reliability of decision-making. The policy caching unit, as a dedicated storage service, is responsible for persistently storing the version of each policy model used by the system in the last 24 hours, its corresponding parameter snapshot, and the sequence of comprehensive evaluation scores calculated by the feedback evaluation module during the execution of that policy.

[0054] When the feedback evaluation module detects that a new strategy has been launched, and its comprehensive evaluation score is lower than the average performance level of all historical strategy versions in the past 6 hours for two consecutive evaluation periods, the system will automatically trigger a strategy rollback operation.

[0055] The rollback operation is not simply switching to the previous version. Instead, it retrieves the historical policy version that performed best under similar system conditions from the policy cache unit, loads its parameters and model into the policy generation module, and replaces the currently underperforming new policy. This process ensures that the system can quickly recover to a known, stable, and high-performance state when faced with unknown dynamics or when the algorithm exploration produces poor policies.

[0056] The resource scheduling execution module is the final executor of system decisions. It receives the latest policy instructions from the policy generation module and, in conjunction with any corrective signals that the feedback evaluation module may provide, generates specific resource allocation instructions that can be recognized and executed by the platform's underlying control system. Internally, this module has three different scheduling logic branches designed based on the type of input instruction.

[0057] For the most common ratio adjustment commands, the module parses the resource type and target allocation ratio specified in the command, then accesses the platform's resource pool management interface to recalculate and allocate the quotas for various resources according to the new ratio. For example, dynamically adjusting the quota for virtual simulation resources from 15% to 20% of the total resources.

[0058] For priority adjustment instructions, the module will reorder all pending requests in the current resource request queue according to the new priority rules defined in the instruction.

[0059] Sorting algorithms typically calculate a new priority score based on a combination of factors, such as the urgency of the request, user level, and task type, and then sort them in descending order of score.

[0060] For emergency dispatch commands, which are typically triggered by abnormal fluctuation detection features, the module will activate a fast-track mechanism. This mechanism will temporarily bypass the normal queuing queue, directly reserve or allocate the required resources, and ensure that the emergency allocation, activation, and deployment of resources are completed within a 2-minute time limit to meet sudden high-priority needs.

[0061] Through the close collaboration of the five core modules mentioned above, the entire system achieves real-time perception, intelligent decision-making, accurate evaluation, and rapid execution of the entire process of resource management on the educational research platform. The real-time data acquisition module ensures the complete capture of multi-dimensional dynamic data;

[0062] The dynamic feature extraction module transforms raw data into features with decision-making value; the policy generation module utilizes advanced online learning algorithms to enable policies to evolve adaptively within minutes; and the feedback evaluation module provides quantitative, multi-objective guidance for policy optimization.

[0063] The resource scheduling and execution module ensures that optimization strategies are accurately and efficiently translated into actual actions. This closed-loop dynamic optimization mechanism enables the system to effectively cope with short-term fluctuations in resource demand caused by changes in student interests, activity segments, sudden equipment failures, or changes in environmental conditions, significantly improving the efficiency of educational resource utilization and the quality of service experience for study tours.

[0064] This embodiment provides another specific implementation of a dynamic feedback resource management system for educational research platforms. Its core feature is that the abnormal fluctuation detection algorithm in the dynamic feature extraction module and the emergency scheduling logic in the resource scheduling execution module have been enhanced to further improve the system's ability to cope with sudden and high-concurrency resource demand scenarios.

[0065] In the dynamic feature extraction module, in addition to the isolated forest algorithm described in the previous embodiment, this embodiment introduces a parallel mutation point detection mechanism based on an exponentially weighted moving average model. This mechanism calculates the exponentially weighted moving average and standard deviation in real time for the historical request volume time series of each resource type. For each newly arrived request volume data point within a time window, its deviation from the current moving average is calculated. If this deviation exceeds three times the current standard deviation, the time window is immediately marked as a potential mutation point, and the confirmation process is initiated.

[0066] The confirmation process checks whether the data in the next two consecutive time windows continues to deviate. If confirmed, the event is marked as a high-confidence abnormal fluctuation signal, with a priority even higher than anomalies detected by the Isolation Forest algorithm. This dual detection mechanism greatly reduces the probability of missed and false alarms, ensuring the accurate capture of genuine sudden demands.

[0067] Correspondingly, the emergency scheduling logic in the resource scheduling execution module has also been upgraded. When an emergency scheduling instruction triggered by the aforementioned enhanced anomaly detection mechanism is received, the module will not only activate the fast channel mechanism, but also simultaneously execute an operation called resource pre-borrowing.

[0068] Resource pre-lending refers to the module's right to temporarily reclaim a portion of resources from non-critical tasks or low-priority users whose current utilization rate is below 20%, and allocate these resources to the emergency resource pool. Pre-lending operations are subject to strict limitations, including a single pre-lending limit of no more than 15% of the system's total resources, a pre-lending time not exceeding 15 minutes, and ensuring that the current core tasks of the users whose resources are pre-lending are not interrupted.

[0069] In this way, the system can dynamically explore its internal potential resource supply when facing sudden peak demand, thereby more effectively meeting urgent needs and ensuring the smooth progress of key research and study activities. The system records the entire process of each resource pre-borrowing operation in detail, including the type, quantity, source, and return time of the pre-borrowed resources, for post-event auditing and strategy optimization.

[0070] Furthermore, this embodiment refines the multi-objective evaluation function of the feedback evaluation module. In the user satisfaction dimension, in addition to indirectly calculating the score, it also attempts to integrate the platform's real-time feedback system. When users actively provide a five-star rating or written review after using resources, this module can capture these explicit feedback data in real time, normalize them, and then directly incorporate them into the user satisfaction score calculation with a 20% weight, making the satisfaction evaluation more direct and accurate.

[0071] Meanwhile, in terms of resource utilization efficiency, a new evaluation index for the degree of resource fragmentation has been added. This index is measured by calculating the proportion of unused small and scattered resource units in the resource pool to the total resources. The system encourages resource allocation strategies to reduce fragmentation as much as possible while improving utilization, so as to improve the overall efficiency of resource scheduling.

[0072] The policy caching and rollback mechanisms are also enhanced in this embodiment. In addition to storing policy versions and performance data, the policy caching unit also records a snapshot of the specific system feature vectors triggered when a policy rollback is initiated. When the system runs to a similar feature state as the historical rollback point, the policy generation module receives a warning signal, making it more cautious when updating policies and avoiding repeating past mistakes, thereby improving the long-term stability and security of the learning algorithm.

[0073] This embodiment, through the aforementioned targeted enhancement design, is particularly suitable for study tour scenarios with rapidly changing activity rhythms, numerous unexpected tasks, and extremely high requirements for resource response timeliness, further expanding the application boundaries and robustness of the system.

Claims

1. A dynamic feedback resource management system for an education research and learning platform, characterized in that, Comprise: Real-time data acquisition module for continuous acquisition of multi-dimensional data stream from user terminal, environmental sensor and platform operation database, the multi-dimensional data stream includes student behavior log, device usage status, environmental parameters, resource inventory and reservation queue data; Dynamic feature extraction module connected to the output end of the real-time data acquisition module, for time window division and feature transformation of the input data stream, extracting key feature vectors representing dynamic changes in resource demand, wherein the time window length is set as a configurable parameter of 5 to 30 minutes; Strategy generation module receives the feature vectors output by the dynamic feature extraction module, and updates the resource allocation strategy model in real time through an online learning algorithm, the model is built using a deep reinforcement learning framework, the state space is defined as the current resource demand distribution and available resource capacity, and the action space is defined as the allocation ratio adjustment instruction of each type of resource; Feedback evaluation module monitors the system state change after strategy execution, calculates resource utilization rate, response delay and user satisfaction index, and generates strategy effect evaluation report; Resource scheduling execution module generates specific resource allocation instructions according to the latest strategy output by the strategy generation module and the correction signal provided by the feedback evaluation module, and issues them to the platform bottom control system.

2. The dynamic feedback type resource management system for the education and research platform according to claim 1, characterized in that, The real-time data acquisition module includes 3 types of data acquisition subunits; User behavior acquisition subunit records the interactive operation sequence of students on the research platform through the burying point technology, including page stay time, content click frequency and task completion progress; Device state acquisition subunit establishes communication connection with physical devices through Internet of Things protocol, and periodically reports device working state, geographic position and remaining available time; Environmental perception acquisition subunit integrates temperature, humidity and illumination intensity sensors, collects environmental data every 10 seconds and uploads them to the platform data hub. 3.The education and research platform-oriented dynamic feedback resource management system of claim 1, wherein, The feature extraction process executed by the dynamic feature extraction module includes sliding window statistics, time series pattern recognition and abnormal fluctuation detection; Sliding window statistics calculates the mean, variance and trend of resource request quantity in each time window; Time series pattern recognition uses an adaptive piecewise aggregation approximation algorithm to reduce dimensionality and extract patterns from long time series data; Abnormal fluctuation detection identifies sudden demand signals that exceed the normal fluctuation range based on the isolation forest algorithm, and marks these signals as high priority features.

4. The education and research platform-oriented dynamic feedback resource management system according to claim 1, characterized in that, The online learning algorithm of the strategy generation module uses the proximal policy optimization method, and its learning process includes two alternating stages of policy evaluation and policy improvement; Policy evaluation stage estimates the value function of the current policy through Monte Carlo sampling; Policy improvement stage updates policy network parameters through gradient ascent method; The policy network performs model update every 5 minutes, and importance sampling technique is used to ensure the stability of the learning process during the update process.

5. The education and research platform-oriented dynamic feedback resource management system according to claim 4, characterized in that, The loss function of the proximal policy optimization method is defined as: ； where, is the expected operation at time step , represents the parameters of the policy network, denotes the probability ratio of taking action under the new policy and the old policy at time step , is the advantage value estimated by the advantage function, which measures the goodness of taking a certain action relative to the average action in a particular state, is a hyperparameter. The function limits the probability ratio to the interval , thereby achieving the stability constraint of policy update.

6. The education and research platform-oriented dynamic feedback resource management system according to claim 1, wherein, The feedback evaluation module constructs a multi-objective evaluation function, which includes 3 core evaluation dimensions; Resource utilization efficiency dimension calculates the average usage rate and idle rate of each type of resource; Response time dimension measures the average time delay from resource request to resource allocation; User satisfaction dimension indirectly calculates the satisfaction score by analyzing user feedback data and behavioral data; The evaluation results of the three dimensions are fused into a single comprehensive evaluation score by weighted summation, and the weight coefficients are dynamically adjusted according to the platform operation strategy.

7. The education and research platform-oriented dynamic feedback resource management system according to claim 1, wherein, The resource scheduling execution module executes different scheduling logic according to the type of input instruction; For proportion adjustment instructions, the module reallocates resource quotas in the available resource pool according to the specified proportion; For priority adjustment instructions, the module reorders the resource request queue according to the preset priority rules; For emergency scheduling instructions, the module starts the fast channel mechanism to complete the emergency allocation and deployment of resources within 2 minutes.

8. The education and research platform-oriented dynamic feedback resource management system according to claim 1, characterized in that, The system also includes a policy cache and rollback mechanism; The policy cache unit stores all policy versions used within the last 24 hours and their performance evaluation data; When the comprehensive evaluation score of the new policy is lower than the average level of historical policies, the system automatically triggers the policy rollback operation and switches to the historical policy version with the best performance to continue running.

9. The education and research platform-oriented dynamic feedback resource management system according to claim 1, wherein, The system uses a distributed micro-service architecture for deployment, and the modules communicate asynchronously through a message queue; The real-time data acquisition module and the dynamic feature extraction module are deployed on edge computing nodes; The policy generation module and the feedback evaluation module are deployed in the cloud computing center; The resource scheduling execution module is directly embedded in the platform core business system, forming a collaborative computing paradigm of edge perception, cloud decision, and local execution.

10. The dynamic feedback resource management system for education and research platform according to claim 3, wherein, The abnormal fluctuation detection also includes a mutation point detection mechanism based on the exponential weighted moving average model; This mechanism calculates the exponential weighted moving average and standard deviation of the historical request volume time series for each type of resource in real time; For the request volume data points in the newly arrived time window, if the deviation from the current moving average exceeds 3 times the current standard deviation, the time window is marked as a potential mutation point; After confirmation, check whether the data of the next two consecutive time windows deviate continuously, and if confirmed, mark the event as a high-confidence abnormal fluctuation signal.

Images