AI prediction-based multi-scene bandwidth dynamic allocation system for colleges and universities

Abstract

This invention discloses an AI-based prediction-based dynamic bandwidth allocation system for universities across multiple scenarios, primarily relating to the field of local area network management technology. It includes: a multi-source data acquisition module for collecting multi-dimensional data and performing data cleaning, standardization, and tagging; a scenario-aware AI prediction module for constructing an AI prediction model fused with scenario tags based on the multi-dimensional data, outputting predicted bandwidth demand values ​​for each physical region within a specified future time period; a dynamic bandwidth scheduling module for generating elastic bandwidth allocation schemes for each region based on the predicted bandwidth demand values, combined with a bandwidth pool mechanism and multi-level priority scheduling logic; and an SDN controller execution module for real-time distribution and execution of the elastic bandwidth allocation schemes, dynamically adjusting the allocation strategy based on execution feedback. The beneficial effects of this invention are: it achieves accurate scenario-based bandwidth prediction for universities while improving resource utilization and ensuring the smooth operation of critical services.

Description

Technical Field

[0001] This invention relates to the field of local area network management technology, specifically to a dynamic bandwidth allocation system for multiple scenarios in universities based on AI prediction. Background Technology

[0002] With the continuous development of smart campus construction, university campus networks have become the core information infrastructure supporting teaching, research, management, and the daily lives of teachers and students. Currently, campus network bandwidth resource management technology has mainly gone through two stages: initially, a static fixed allocation mode was commonly used, which preset fixed bandwidth quotas for different areas (such as teaching buildings and dormitory areas); subsequently, some dynamic allocation schemes based on historical traffic statistics were developed, which can be adjusted to a certain extent according to past usage patterns.

[0003] However, these existing technologies have revealed the following major drawbacks in practical applications:

[0004] 1. Rigid resource allocation, unable to adapt to the dynamic needs of university scenarios: The reason is that the fixed allocation mode and simple historical traffic prediction have failed to effectively perceive and integrate the unique business scenario information of universities that changes rapidly over time (such as course schedules and research task cycles), resulting in insufficient bandwidth preparation in sudden high-load scenarios such as online courses and large-scale scientific research data transmission, while resources are idle during off-peak hours.

[0005] 2. Lack of a refined priority guarantee mechanism makes key business operations susceptible to interference: This is mainly because the existing system fails to effectively distinguish and identify traffic of different natures, such as "teaching needs", "scientific research tasks" and "life and entertainment". As a result, when bandwidth competition occurs, low-priority traffic (such as video streaming) may crowd out the bandwidth required by high-priority business (such as online examinations and national-level project data transmission), affecting the conduct of core teaching and scientific research activities.

[0006] 3. Limited prediction accuracy and delayed scheduling response, leaving the system in a passive response state: This defect is caused by the fact that traditional prediction models do not incorporate advanced algorithms such as artificial intelligence to deeply explore the correlation between multi-source data and ignore key features such as scene labels, resulting in large prediction deviations for special events such as the start of the school season and exam week; at the same time, scheduling strategies are mostly triggered and adjusted "after congestion occurs", with long response delays, which cannot meet the real-time requirements of interactive teaching, online experiments and other scenarios.

[0007] Therefore, there is an urgent need for an AI-based predictive system for dynamic bandwidth allocation in multiple scenarios at universities to solve the above problems. Summary of the Invention

[0008] The purpose of this invention is to provide an AI-based prediction-based dynamic bandwidth allocation system for universities across multiple scenarios. This system enables accurate prediction, proactive allocation, and priority guarantee of bandwidth for universities based on specific scenarios, while improving resource utilization and ensuring the smooth operation of critical services.

[0009] To achieve the above objectives, the present invention employs the following technical solution:

[0010] This invention provides an AI-based prediction-based dynamic bandwidth allocation system for multiple scenarios in universities, comprising:

[0011] The multi-source data acquisition module is used to collect multi-dimensional data on teaching, research and living scenarios from university academic affairs systems, scientific research management platforms, dormitory management systems and network devices, and to clean, standardize and label the data.

[0012] The scene-aware AI prediction module constructs an AI prediction model that integrates scene tags based on the multi-dimensional data, and outputs the predicted bandwidth demand values ​​for each physical area within a specified future time period.

[0013] The dynamic bandwidth scheduling module generates an elastic bandwidth allocation scheme for each region based on the predicted bandwidth demand, combined with the bandwidth pool mechanism and multi-level priority scheduling logic.

[0014] The SDN controller execution module communicates with network devices through software-defined network protocols, issues and executes the elastic bandwidth allocation scheme in real time, and dynamically adjusts the allocation strategy based on execution feedback.

[0015] Preferably, the multi-source data acquisition module specifically includes:

[0016] The data interface unit is used to collect data from heterogeneous data sources;

[0017] The data preprocessing unit cleans, standardizes, and labels the collected data. The timestamp format unification algorithm supports multiple input formats and converts them to a standard format. Numerical data is normalized using the Z-Score standardization method, with the formula:

[0018] ;

[0019] in, These are the original eigenvalues. This is the mean of the feature across all samples. Standard deviation;

[0020] The data acquisition frequency dynamic adjustment unit sets hierarchical acquisition frequencies based on scenario priority and time period characteristics, and temporarily adjusts the acquisition frequency when a specific event is triggered.

[0021] Preferably, the scene-aware AI prediction module employs a hybrid model combining LSTM and an attention mechanism, specifically including:

[0022] Input feature vectors, including time features, scene features, and historical features;

[0023] The LSTM base layer is used to capture the temporal dependencies of traffic data.

[0024] The attention layer assigns higher weights to high-priority scene features, thereby enhancing the prediction accuracy of key scenes.

[0025] The output layer outputs the predicted bandwidth demand for each region over the next 30 minutes.

[0026] The hyperparameters of the model were determined through Bayesian optimization and cross-validation, including a hidden layer dimension of 128, a time step of 30 minutes, a dropout probability of 0.3, a number of attention heads of 4, and a learning rate of [missing information]. Batch size is 64.

[0027] Preferably, the AI ​​prediction module further includes a model training and optimization unit, specifically including:

[0028] The training dataset construction unit collects multi-source data from at least three full semesters and supplements it with a special event dataset, with a total data volume of no less than 1 million records;

[0029] The hyperparameter optimization unit uses Bayesian optimization and cross-validation methods to search for the optimal hyperparameter combination.

[0030] The cold start processing unit employs a transfer learning and rule completion strategy for scenarios lacking historical data. The transfer learning uses models with a scenario similarity of ≥80% from the pre-trained model library for parameter fine-tuning.

[0031] The model drift correction unit performs drift evaluation every two weeks. When the RMSE on the test set increases by more than 15% or the prediction deviation of key scenarios exceeds 20%, incremental training or feature update is triggered.

[0032] Preferably, the dynamic bandwidth scheduling module specifically includes:

[0033] The bandwidth pool is divided into a basic bandwidth pool and an elastic bandwidth pool. The basic bandwidth pool is used to guarantee the minimum bandwidth requirements of each region, while the elastic bandwidth pool accounts for 30%-50% of the total bandwidth.

[0034] The priority scheduling logic unit divides the scenarios into four priorities: the first priority is teaching-related scenarios and national-level scientific research projects; the second priority is provincial-level scientific research projects and school-level teaching platforms; the third priority is daily office work for teachers and students; and the fourth priority is entertainment traffic.

[0035] The scheduling execution unit allocates elastic bandwidth according to priority based on the difference between the predicted demand and the basic bandwidth, and imposes bandwidth restrictions on low-priority areas when the elastic bandwidth is insufficient.

[0036] Preferably, the SDN controller execution module specifically includes:

[0037] The hardware interface unit is used to communicate with the core switch and edge router, and updates the bandwidth allocation command every 5 minutes.

[0038] The status feedback unit collects the execution results of the device in real time. If the actual bandwidth deviates from the predicted demand by more than 10%, it triggers the model to re-predict and reschedule.

[0039] The dual-machine hot standby unit adopts a master-slave controller architecture and achieves fault switching through the VRRP protocol, with a switching latency of no more than 3 seconds.

[0040] Preferably, it also includes a security and privacy protection module, specifically including:

[0041] The data anonymization unit employs a tiered anonymization strategy for student and faculty identity information, accommodation data, research project data, and traffic logs, including k-anonymization, field deletion, data generalization, and encryption.

[0042] The access control unit sets data access and operation permissions for four roles: AI model trainer, network administrator, researcher, and ordinary teachers and students, based on the RBAC model.

[0043] The transmission and storage encryption unit uses the TLS 1.3 protocol to encrypt the data transmission link, and the database uses TDE and AES-256 encryption to store sensitive data;

[0044] The audit and traceability unit records all data access and operation logs, and uses blockchain for evidence storage to ensure that the logs are tamper-proof.

[0045] Preferably, it also includes an edge node collaborative scheduling module, specifically including:

[0046] The edge node deployment unit deploys three types of edge gateways in teaching buildings, laboratory clusters, and dormitory areas, which are respectively responsible for local data collection, lightweight prediction, and scheduling execution.

[0047] The core-edge collaborative scheduling unit divides local fine-tuning and global scheduling permissions according to the adjustment magnitude and scope of impact. Local fine-tuning is executed by edge nodes, while global scheduling is coordinated by the core controller.

[0048] The edge node fault handling unit takes over the scheduling authority of an edge node when it is offline for more than 30 minutes, and synchronizes the latest rules after the node recovers.

[0049] Preferably, it also includes a bandwidth burst over-limit processing module, specifically including:

[0050] The over-limit detection unit compares the actual flow rate with the predicted value every 10 seconds. When the actual flow rate exceeds the predicted value by 20% and continues for 30 seconds, it is determined to be a sudden over-limit.

[0051] The fast elastic allocation unit can search for idle elastic bandwidth in low-priority areas within 1 second and allocate it according to the principle of proximity. The whole process takes no more than 5 seconds.

[0052] The borrowing and recycling unit has a default borrowing time of no more than 15 minutes. Borrowed bandwidth will be recycled within 5 seconds after the traffic in the over-limit area returns to normal.

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

[0054] 1. This invention achieves accurate prediction and proactive allocation of bandwidth resources, significantly improving scenario adaptability. By integrating multi-dimensional scenario data from teaching, research, and daily life, and employing an LSTM model with an attention-based mechanism for prediction, this invention overcomes the shortcomings of traditional solutions that rely solely on historical traffic and ignore business logic. This enables the system to accurately predict the needs of special events such as "back-to-school season," "online exams," and "research data transmission," achieving a fundamental shift from "passive response after congestion" to "proactive preparation before demand." Experimental verification shows a significant improvement in prediction accuracy.

[0055] 2. A multi-level priority guarantee mechanism was established to ensure the smoothness and stability of critical services. By introducing a clear four-level priority scheduling logic and combining it with the rapid execution capability of the SDN controller, the system can prioritize the bandwidth needs of high-priority scenarios when there is competition for campus network resources. This effectively solves the long-standing pain points of university network management, such as online exam lag and errors in scientific research data transmission, fundamentally ensuring the smoothness of services in critical scenarios and significantly reducing the lag rate.

[0056] 3. Significantly improves overall bandwidth resource utilization and reduces expansion costs. This invention establishes a "basic bandwidth pool" and an "elastic bandwidth pool," and dynamically schedules bandwidth based on prediction results, enabling previously idle bandwidth resources to flow efficiently across different regions and time periods. This refined resource management model breaks down resource barriers caused by fixed quotas, significantly increasing the average utilization rate of the campus network's total bandwidth, thereby delaying or reducing the need for hardware expansion to cope with peak traffic and saving construction costs.

[0057] 4. A complete technical solution integrating prediction, scheduling, and security has been constructed, resulting in high system intelligence and reliability. This invention is not merely an improvement on a single algorithm, but rather the construction of a complete system encompassing data acquisition, AI prediction, dynamic scheduling, security protection, and edge collaboration. Its mechanisms, including model drift correction, emergency response, and fault redundancy, ensure long-term stable operation and adaptability to changes in university operations, forming a self-optimizing, robust, and reliable intelligent bandwidth allocation closed loop for the campus network. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0059] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.

[0060] In this invention, terms such as "upper," "lower," "left," "right," "front," "back," "vertical," "horizontal," "side," and "bottom" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used only to facilitate the description of the structural relationships of the various components or elements of this invention and do not specifically refer to any component or element in this invention. They should not be construed as limiting the invention.

[0061] Example:

[0062] like Figure 1 As shown, this embodiment provides a dynamic bandwidth allocation system for multiple scenarios in universities based on AI prediction, including: a multi-source data acquisition module, a scenario-aware AI prediction module, a dynamic bandwidth scheduling module, an SDN controller execution module, an emergency response module, a security and privacy protection module, an edge node collaborative scheduling module, and a bandwidth burst over-limit handling module.

[0063] The multi-source data acquisition module also includes: a data interface unit, a data preprocessing unit, and a data acquisition frequency dynamic adjustment unit;

[0064] The scene-aware AI prediction module adopts a hybrid model combining LSTM and attention mechanism. The hybrid model includes: input feature vector, LSTM base layer, attention layer and output layer. The scene-aware AI prediction module also includes: model training and optimization unit, specifically including: training dataset construction unit, hyperparameter optimization unit, cold start processing unit and model drift correction unit.

[0065] The dynamic bandwidth scheduling module also includes: a bandwidth pool partitioning unit, a priority scheduling logic unit, and a scheduling execution unit;

[0066] The SDN controller execution module also includes: a hardware interface unit, a status feedback unit, and a dual-machine hot standby unit;

[0067] The emergency response module also includes: an emergency event classification unit, an emergency triggering and execution unit, and an emergency status disclosure unit;

[0068] The security and privacy protection module also includes: a data anonymization unit, an access control unit, a transmission and storage encryption unit, and an audit and traceability unit;

[0069] The edge node collaborative scheduling module also includes: an edge node deployment unit, a core-edge collaborative scheduling unit, and an edge node fault handling unit;

[0070] The bandwidth burst over-limit processing module also includes: an over-limit detection unit, a fast and flexible allocation unit, and a borrowing and recycling unit.

[0071] The core implementation process of the system in this embodiment is as follows:

[0072] First, the multi-source data acquisition module is put into operation: the acquisition agent deployed in this module simultaneously extracts data from the RESTful API of the academic affairs system, the database of the scientific research platform, the RabbitMQ message queue of the dormitory management system, and the NetFlow stream of the core switch; for example, at 7:50 am, the system retrieves a course record from the academic affairs API:

[0073] {"course_id":"CS101";

[0074] "class_time":"2024-09-0108:00";

[0075] "online_flag":1;

[0076] "student_count":100};

[0077] At the same time, a real-time headcount data was received from the dormitory system message queue:

[0078] {"dorm_building":"Building 1";}

[0079] "active_people": 350;

[0080] "timestamp":"2024-09-0107:50"};

[0081] Subsequently, the data preprocessing unit in the multi-source data acquisition module was activated: it performed format unification recognition on the "class_time" field, successfully converting it from "YYYY-MM-DDHH:MM" to the standard format "2024-09-01 08:00:00"; at the same time, it tagged the course data with "teaching scenario=1, online teaching=1", and calculated the Z-Score standardized value for the dormitory data, assuming the average number of active users in all dormitory areas of the campus. Standard deviation The standardized value of "Building No. 1" is then... ;

[0082] Next, the scene-aware AI prediction module is triggered: the preprocessed feature vector (containing time, scene, and historical features) is input into the trained LSTM-attention hybrid model. The model runs its attention mechanism and finds that there is a high-weight "online teaching" scene in the current time period (08:00), so it significantly increases the weight of neurons related to this scene. The model finally outputs the prediction result: the bandwidth demand of the first teaching building area is predicted to be 850Mbps in the next 30 minutes.

[0083] Then, the dynamic bandwidth scheduling module calculates based on the predicted value. It finds that the basic bandwidth of the first teaching building is 500Mbps, therefore the elastic bandwidth requirement is 850-500=350Mbps. The scheduler searches the elastic bandwidth pool and finds that the currently available resources are 400Mbps. Since the teaching scenario is the first priority, the system immediately approves the allocation of 350Mbps of elastic bandwidth to the first teaching building area.

[0084] Finally, the SDN controller execution module, via the OpenFlow protocol, issued flow table modification commands to the core switch responsible for the first teaching building area, completing the reallocation of bandwidth resources within 5 seconds. Simultaneously, the controller continuously monitors the actual traffic in this area. If, after 5 minutes, the actual traffic is found to be consistently 20% lower than the predicted value, a feedback mechanism is triggered, requiring the AI ​​prediction module to re-predict and re-schedule.

[0085] The data interface unit operates as follows:

[0086] For the RESTful API of the academic affairs system, the system uses Java's HttpClient library to construct HTTPSGET requests and carries a JWT token in the header for authentication. The system sets the connection timeout to 30 seconds. If the timeout occurs, it will retry according to the "exponential backoff" strategy.

[0087] For the database of the scientific research management platform, the system uses the OracleReader plugin of the DataX middleware, configures the JDBC connection string, uses the readonly account readonly_user and limits the source IP to execute SQL;

[0088] For the message queue of the dormitory management system, the system establishes a RabbitMQ consumer, binds it to a Direct type exchange named dormitory.direct, subscribes using the routing key dormitory.flow, and deserializes received ProtocolBuffers format messages into Java objects for processing.

[0089] The core algorithm implementation of the data preprocessing unit is as follows:

[0090] 1. Unified timestamp format algorithm:

[0091] {

[0092] / / Input the raw time string

[0093] String rawTime="202409010800"; / / Example from a research platform

[0094] String[]possibleFormats={"yyyyMMddHHmm","MM / dd / yyyyHH:mm","yyyy-MM-ddHH:mm"};

[0095] StringstandardFormat="yyyy-MM-ddHH:mm:ss";

[0096] for(Stringformat:possibleFormats){

[0097] try{

[0098] SimpleDateFormatsdf=newSimpleDateFormat(format);

[0099] Datedate = sdf.parse(rawTime);

[0100] SimpleDateFormatoutputSdf=newSimpleDateFormat(standardFormat);

[0101] StringstandardizedTime=outputSdf.format(date); / / Output: 2024-09-0108:00:00;

[0102] break

[0103] }catch(ParseExceptione){

[0104] / / Try another format

[0105] }

[0106] }

[0107] },

[0108] If all format matches fail, the string length and character characteristics are used to infer the format. For example, if the string is a 10-digit pure number, it is considered a Unix timestamp and needs to be multiplied by 1000 before being converted to the standard format.

[0109] 2. Data cleaning and standardization:

[0110] Outlier cleaning: For the number of active people in the dormitory, the system presets the maximum capacity of each building to be 500 people. If "active_people":600 is collected, it will be automatically corrected to 500 and marked as "cleaned".

[0111] Z-Score standardization: This process handles continuous features such as "historical bandwidth usage," and the formula is as follows:

[0112] ;

[0113] in, These are the original eigenvalues. It is the mean of this feature on the training set. It is the standard deviation on the training set, for example: the average bandwidth usage of the entire school. Mbps, standard deviation If the current usage in a certain area is 2500 Mbps, then the standardized value is... ;

[0114] 3. The system maintains a scene tag mapping dictionary. When processing a scientific research project data, if the project_level field is 3, the tag "scientific research scene=1, national-level project=1, priority=top level" will be automatically added to it.

[0115] For course data, if online_flag is 1, add the tag "Teaching scenario=1, Online teaching=1, Priority=Level 1".

[0116] The strategy for the dynamic adjustment unit of the data acquisition frequency is as follows:

[0117] During critical periods in the teaching scenario (1 hour before class to 30 minutes after class), the system collects data from the academic affairs system at a frequency of once per minute. If the academic affairs system pushes a "temporary additional class" notification at this time, the system will immediately increase the frequency to once per 30 seconds and continuously monitor the dynamics of the temporary class until it ends.

[0118] During off-peak hours in daily life (such as 2 a.m.), the data collection frequency is reduced to once every 15 minutes to save system resources.

[0119] In this embodiment, the model input feature vector is constructed as follows:

[0120] Time characteristics: [Weekday = 1 (Monday), Time period = 8, Holiday status = 0],

[0121] Scene characteristics: [Number of online courses = 5, Number of national-level projects = 2, Number of active users in dormitory area = 1200]

[0122] Historical characteristics: [Average bandwidth over the same period in the past 7 days = 1800, traffic growth rate = 0.05];

[0123] The forward propagation process of the LSTM and attention hybrid model is as follows:

[0124] 1. LSTM Base Layer: Input sequence data (such as features from the past 30 time steps), the LSTM unit calculates the hidden state through its gating mechanism (input gate, forget gate, output gate). For a time step Its calculation can be simplified to:

[0125] ;

[0126] in, yes The input feature vector at time t, This is the hidden state from the previous time step. The final output of this layer is a sequence of hidden states containing temporal information. ;

[0127] 2. Attention Layer: This layer receives the output of the LSTM. And calculate the attention weights at each time step. :

[0128] ;

[0129] in, For a scoring function, As a query vector, it focuses more on historical information most relevant to the current prediction, for example, when Compared with features including "online teaching" When highly correlated, the weight of this time step It will increase, and the final context vector is obtained by weighted summation:

[0130] ;

[0131] This allows the model to pay more "attention" to historical moments related to high-priority scenarios when making predictions;

[0132] 3. Output layer: This layer stores the context vector. Through a fully connected layer, the final output is a scalar value, which is the bandwidth prediction value for the next 30 minutes;

[0133] The hyperparameter optimization process adopts a Bayesian optimization framework (such as the Hyperopt library), and the optimization objective is the RMSE (root mean square error) on the validation set. After 100 rounds of iterative search, the optimal hyperparameter combination is finally determined as follows: hidden layer dimension = 128, time step = 30, dropout probability = 0.3, number of attention heads = 4, learning rate = 5e-4, batch size = 64.

[0134] In this embodiment, the decision logic of the dynamic bandwidth scheduling module is as follows:

[0135] Assuming the total outbound bandwidth of the campus network is 10Gbps, the system divides it into a basic bandwidth pool and an elastic bandwidth pool:

[0136] The basic bandwidth pool is 7Gbps, used to guarantee the minimum requirements of each region, for example:

[0137] Dormitory area: (Total number of residents: 5000) × 2Mbps / person = 10Gbps (Note: This is the sum of the basic bandwidth of each area, which may exceed 7G. The actual allocation will be scaled up proportionally or prioritized).

[0138] Teaching buildings: 500Mbps fixed speed for each building;

[0139] Research building: Each national-level project laboratory has a fixed 200Mbps connection;

[0140] The elastic bandwidth pool is 3Gbps. This is a globally shared resource pool used to cope with sudden demand in various regions.

[0141] Execution flow of priority scheduling logic:

[0142] Demand Calculation: The AI ​​module predicts that in the next 30 minutes, Area A (teaching building, priority 1) will need 1050Mbps, Area B (research, priority 2) will need 600Mbps, and Area C (dormitory, priority 4) will need 1200Mbps. Their base bandwidths are 500Mbps, 200Mbps, and 800Mbps, respectively. Therefore, the elastic demand is as follows: A needs 550Mbps, B needs 400Mbps, and C needs 400Mbps.

[0143] Resource allocation:

[0144] First, the 550Mbps requirement of the highest priority area A is met. After allocation, the elastic pool has 3000-550=2450Mbps remaining.

[0145] Secondly, to meet the 400Mbps requirement of region B, after allocation, the elastic pool has 2450-40=2050Mbps remaining;

[0146] Finally, the 400Mbps requirement of region C was processed. At this time, the elastic pool was sufficient, so it was fully allocated. In the end, the elastic pool had 2050-400=1650Mbps remaining.

[0147] Resource contention handling (assuming another scenario):

[0148] If the elastic pool is initially set at 800Mbps, after satisfying the 550Mbps requirement of region A, only 250Mbps remains, which is insufficient to meet the 400Mbps requirement of region B.

[0149] The system will allocate 250Mbps to region B proportionally.

[0150] Meanwhile, since region C has the lowest priority and the elastic pool is empty, the system will not allocate any elastic bandwidth to region C and may rate limit its entertainment traffic (such as data packets whose destination port is streaming media) to ensure that it does not exceed 80% of the basic bandwidth.

[0151] In this embodiment, the execution and high availability mechanism of the SDN controller execution module is specifically as follows:

[0152] 1. Workflow of the SDN controller execution module:

[0153] Instruction encapsulation: The scheduling module generates the allocation scheme.

[0154] {Region:"Building_A",

[0155] After Bandwidth:850Mbps, Priority:1}, the controller converts it into a Flow-Mod message according to the OpenFlow 1.5 protocol standard;

[0156] Command issuance: The controller issues a command to the core switch through a secure channel (such as a TLS-encrypted TCP connection). The command content is roughly as follows: "Add a rule to the flow table of port 1 that matches all IP packets on this port and sets the maximum rate of its queue to 850Mbps".

[0157] Status feedback: The controller queries the switch for port counter information every 30 seconds via OpenFlow's Multipart-Request message to obtain the actual traffic. If the actual traffic in area A is found to be below 700Mbps (82% of the predicted value of 850Mbps) for one minute, it is determined that the prediction deviation is too large, triggering the AI ​​model to re-evaluate.

[0158] 2. Fault switching process of the dual-machine hot standby unit:

[0159] The primary controller (Active) and the standby controller monitor each other via a heartbeat line. The primary controller synchronizes critical states in memory (such as flow table rules and priority mapping tables) with the standby controller every 2 seconds.

[0160] When the server where the main controller is located goes down due to a power failure, the backup controller will immediately determine that the main controller has failed after failing to receive a heartbeat for 3 consecutive times (3 seconds in total).

[0161] The backup controller then takes over ownership of the virtual IP via the VRRP protocol;

[0162] The network switch will detect the connection interruption to the main controller and automatically reconnect to a new virtual IP, i.e., the backup controller. The entire process, from the occurrence of the fault to the restoration of services, takes no more than 3 seconds, during which only 1-2 scheduling instructions may be lost, without affecting the established network connection.

[0163] In this embodiment, the emergency response module operates as follows:

[0164] Take major exams as an example:

[0165] 1. Trigger: The day before the exam, the academic affairs system sends an emergency command to the SDN controller via an encrypted REST API interface. The command body is in JSON format and digitally signed using an RSA private key.

[0166] {

[0167] "event_id":"EMERGENCY_EXAM_2024";

[0168] "event_type":"National-level examination";

[0169] "affected_areas":["First Teaching Building","Second Teaching Building"];

[0170] "start_time":"2024-12-2508:00:00";

[0171] "end_time":"2024-12-2512:00:00";

[0172] "action":"SET_PRIORITY_TO_SPECIAL"};

[0173] 2. Verification and Update: After receiving the instruction, the SDN controller first uses the pre-set public key of the academic affairs system to verify the digital signature. After successful verification, the regional priority of "First Teaching Building" and "Second Teaching Building" is upgraded from "Level 1" to "Special Level" within 10 seconds.

[0174] 3. Bandwidth preemption: The controller calculates that in order to ensure the smooth conduct of the exam, these two areas need an additional 500Mbps of elastic bandwidth. Therefore, it "borrows" 500Mbps from the allocated elastic bandwidth of the current low-priority areas (such as the dormitory area). The controller issues a flow table modification command to the dormitory area switch to reduce its total bandwidth limit by 500Mbps.

[0175] 4. Status Announcement: Simultaneously, the campus network management platform homepage has been updated with a prominent announcement: "[Urgent Notice] Due to the National Postgraduate Entrance Examination, the network priority of the first and second teaching buildings has been adjusted to the highest level. Network speed fluctuations may occur in some dormitory areas. We apologize for the inconvenience." Faculty and students can check the bandwidth allocation status of each area in real time.

[0176] In this embodiment, the specific implementation measures of the security and privacy protection module are as follows:

[0177] 1. Specific operation of the data anonymization unit:

[0178] Teacher and student identity information: For student ID 2024012345, k-anonymization (k=50) and field deletion are applied. First, the name is deleted, and then the student ID is processed to 202401****, so that at least 50 students have the same generic student ID, making it impossible to locate the individual.

[0179] Accommodation data: The specific room number "Room 301, Building 1" is generalized to "Building 1". Individual-level accommodation records are aggregated into building-level statistical data, such as "Total number of people in Building 1: 240".

[0180] Research project data: The project leader's name "Zhang San" is encrypted using the AES-256 algorithm. The encryption key is centrally managed by KMS (Key Management System), and the encrypted ciphertext is "U2FsdGVkX1 / ...v4sX6ZzQ==", which can only be decrypted and viewed by authorized administrators;

[0181] Traffic logs: The last segment of the terminal IP address is replaced with 0, which anonymizes the network segment while retaining information for regional scheduling;

[0182] 2. Implementation of access control policies:

[0183] After the AI ​​model trainer logs into the system, their SQL query will be automatically rewritten. When they want to query `SELECT * FROM flow_log`, the system actually executes:

[0184] SELECTtimestamp,area,bandwidth_usageFROMflow_log;

[0185] It does not contain any IP or identity information;

[0186] When a network administrator attempts to delete a historical traffic record, the system will display a message saying "Insufficient permissions: Deletion of historical data is prohibited." All of their operations, including this deletion attempt, will be recorded in the audit log.

[0187] 3. Operation of the audit and traceability mechanism:

[0188] Every hour, the system packages the operation logs generated in the past hour, calculates their SHA-256 hash values, and writes these hash values ​​into a private blockchain network. For example, the 1001st block contains the hash value a1b2c3d4 of the log file log_2024090110.zip. Any tampering with the log file will cause its hash value to differ from the record on the chain, thus being detected immediately.

[0189] In this embodiment, the specific implementation of the edge node collaborative scheduling module is as follows:

[0190] 1. Edge node deployment:

[0191] In a teaching building called "Yifu Building", a teaching edge gateway is deployed. Its hardware configuration is: Intel Core i7-13700H CPU, 32GB DDR5 memory, 512GB SSD, and equipped with 4 10 Gigabit optical ports. It is responsible for the network data of all 20 classrooms and 5 teachers' offices in Yifu Building.

[0192] 2. Core-edge collaborative scheduling rules:

[0193] (1) Local fine-tuning: One afternoon, an online lecture was held in classroom 203 of the Yifu Building, while classroom 205 was idle. The local prediction module of the edge gateway predicted that the demand of classroom 203 would increase by 50Mbps in the next 15 minutes. Since the adjustment range (50Mbps) < 100Mbps and the adjustment was only made within the Yifu Building, the local scheduling module directly issued an instruction to dynamically allocate part of the bandwidth of classroom 205 to classroom 203. The whole process was completed within 1 second and the result was reported to the core controller.

[0194] (2) Global scheduling: On another day, multiple online exams were conducted simultaneously across the school. The core controller needed to perform large-scale bandwidth allocation across regions (>100Mbps). It issued a global instruction to the Yifu Building edge gateway: "Yifu Building area, the overall bandwidth limit is increased by 200Mbps, and the exam system is given priority." After receiving the instruction, the edge gateway immediately executed it and suspended any local fine-tuning that might conflict with it.

[0195] 3. Edge node fault handling:

[0196] The core controller sends a UDP heartbeat packet to the Yifu Building gateway every 10 seconds. On one occasion, the network optical cable was cut during construction, resulting in three consecutive heartbeat timeouts.

[0197] The core controller determines that the node is offline. During the 30-minute offline period, the controller only caches scheduling instructions for that area; after 30 minutes, the controller automatically takes over the bandwidth scheduling rights of the Yifu Building and continues to allocate bandwidth to it based on the historical data reported by the Yifu Building and global predictions.

[0198] Two hours later, the fiber optic cable was repaired, and the edge gateway came back online. It re-established its connection with the core controller, which synchronized all rule changes from the past two hours to the gateway and then returned scheduling authority to the edge node.

[0199] In this embodiment, the bandwidth burst over-limit processing module processes the data as follows:

[0200] 1. Over-limit detection:

[0201] The system reads the port counter of the core switch every 10 seconds. One evening at 20:00:00, it detected that the actual traffic in the dormitory area was 1.2Gbps, while the AI ​​predicted 1.0Gbps.

[0202] The system started timing, and the actual traffic was 1.22Gbps and 1.25Gbps at 20:00:10 and 20:00:20 respectively.

[0203] At 20:00:30, the system determined that the actual traffic exceeded the predicted value by more than 20% for 30 consecutive seconds (the over-limit range was 25%), which was considered a moderate over-limit, and immediately triggered the "rapid elastic allocation" process.

[0204] 2. Fast and flexible allocation:

[0205] (1) Gap calculation: The controller calculates the bandwidth gap in the dormitory area as 1.25Gbps (actual) - 1.0Gbps (predicted) - 0.2Gbps (currently allocated elastic bandwidth) = 0.05Gbps (i.e. 50Mbps);

[0206] (2) Resource retrieval: The controller retrieves the usage of the school's flexible bandwidth within 1 second and finds that 100Mbps of flexible bandwidth in the teaching building area is not being fully utilized;

[0207] (3) Allocation: According to the "proximity principle", the controller selects the teaching building area under the same aggregation switch as the dormitory area for allocation. It sends flow table instructions to the core switch and aggregation switch to temporarily allocate 100Mbps elastic bandwidth of the teaching building area to the dormitory area. The whole process takes 4.5 seconds from detection to allocation completion.

[0208] (4) Borrowing and Reclaiming: 15 minutes later (20:15:00), the system detected that the traffic in the dormitory area had dropped to 1.02Gbps. The controller issued an instruction within 5 seconds to reclaim the temporarily allocated 100Mbps bandwidth, and the elastic bandwidth in the teaching building area was restored to its original state.

[0209] The above is a detailed description of the preferred embodiments of the present invention, but the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A dynamic bandwidth allocation system for multiple scenarios in universities based on AI prediction, characterized in that, include: The multi-source data acquisition module is used to collect multi-dimensional data on teaching, research and living scenarios from university academic affairs systems, scientific research management platforms, dormitory management systems and network devices, and to clean, standardize and label the data. The scene-aware AI prediction module constructs an AI prediction model that integrates scene tags based on the multi-dimensional data, and outputs the predicted bandwidth demand values ​​for each physical area within a specified future time period. The dynamic bandwidth scheduling module generates an elastic bandwidth allocation scheme for each region based on the predicted bandwidth demand, combined with the bandwidth pool mechanism and multi-level priority scheduling logic. The SDN controller execution module communicates with network devices through a software-defined network protocol, issues and executes the elastic bandwidth allocation scheme in real time, and dynamically adjusts the allocation strategy based on execution feedback. The multi-source data acquisition module specifically includes: The data interface unit is used to collect data from heterogeneous data sources; The data preprocessing unit cleans, standardizes, and labels the collected data. The timestamp format unification algorithm supports multiple input formats and converts them to a standard format. Numerical data is normalized using the Z-Score standardization method, with the formula: ； in, These are the original eigenvalues. This is the mean of the feature across all samples. Standard deviation; The data acquisition frequency dynamic adjustment unit sets hierarchical acquisition frequencies according to scene priority and time period characteristics, and temporarily adjusts the acquisition frequency when an event is triggered. The scene-aware AI prediction module employs a hybrid model combining LSTM and an attention mechanism, specifically including: Input feature vectors, including time features, scene features, and historical features; The LSTM base layer is used to capture the temporal dependencies of traffic data. The attention layer assigns higher weights to high-priority scene features, thereby enhancing the prediction accuracy of key scenes. The output layer outputs the predicted bandwidth demand for each region over the next 30 minutes. The hyperparameters of the model were determined through Bayesian optimization and cross-validation, including a hidden layer dimension of 128, a time step of 30 minutes, a dropout probability of 0.3, a number of attention heads of 4, and a learning rate of [missing information]. Batch size is 64; The AI ​​prediction module also includes a model training and optimization unit, specifically including: The training dataset construction unit collects multi-source data from at least three full semesters and supplements it with a special event dataset, with a total data volume of no less than 1 million records; The hyperparameter optimization unit uses Bayesian optimization and cross-validation methods to search for the optimal hyperparameter combination. The cold start processing unit employs a transfer learning and rule completion strategy for scenarios lacking historical data. The transfer learning uses models with a scenario similarity of ≥80% from the pre-trained model library for parameter fine-tuning. The model drift correction unit performs drift evaluation every two weeks. When the RMSE on the test set increases by more than 15% or the prediction deviation of key scenarios exceeds 20%, incremental training or feature update is triggered. The dynamic bandwidth scheduling module specifically includes: The bandwidth pool is divided into a basic bandwidth pool and an elastic bandwidth pool. The basic bandwidth pool is used to guarantee the minimum bandwidth requirements of each region, while the elastic bandwidth pool accounts for 30%-50% of the total bandwidth. The priority scheduling logic unit divides the scenarios into four priorities: the first priority is teaching-related scenarios and national-level scientific research projects; the second priority is provincial-level scientific research projects and school-level teaching platforms; the third priority is daily office work for teachers and students; and the fourth priority is entertainment traffic. The scheduling execution unit allocates elastic bandwidth according to priority based on the difference between the predicted demand and the basic bandwidth, and imposes bandwidth restrictions on low-priority areas when the elastic bandwidth is insufficient. The SDN controller execution module specifically includes: The hardware interface unit is used to communicate with the core switch and edge router, and updates the bandwidth allocation command every 5 minutes. The status feedback unit collects the execution results of the equipment in real time. If the actual bandwidth deviates from the predicted demand by more than 10%, it triggers the model to re-predict and reschedule. The dual-machine hot standby unit adopts a master-slave controller architecture and achieves fault switching through the VRRP protocol, with a switching latency of no more than 3 seconds.

2. The AI-based prediction-based dynamic bandwidth allocation system for multiple scenarios in universities according to claim 1, characterized in that, It also includes a security and privacy protection module, specifically including: The data anonymization unit employs a tiered anonymization strategy for student and faculty identity information, accommodation data, research project data, and traffic logs, including k-anonymization, field deletion, data generalization, and encryption. Access control unit, based on RBAC model, sets data access and operation permissions for four roles: AI model trainer, network administrator, researcher and ordinary teacher and student; The transmission and storage encryption unit uses the TLS 1.3 protocol to encrypt the data transmission link, and the database uses TDE and AES-256 encryption to store sensitive data; The audit and traceability unit records all data access and operation logs, and uses blockchain for evidence storage to ensure that the logs are tamper-proof.

3. The AI-based prediction-based dynamic bandwidth allocation system for multiple scenarios in universities according to claim 1, characterized in that, It also includes an edge node collaborative scheduling module, specifically including: The edge node deployment unit deploys three types of edge gateways in teaching buildings, laboratory clusters, and dormitory areas, which are respectively responsible for local data collection, lightweight prediction, and scheduling execution. The core-edge collaborative scheduling unit divides local fine-tuning and global scheduling permissions according to the adjustment magnitude and scope of impact. Local fine-tuning is executed by edge nodes, while global scheduling is coordinated by the core controller. The edge node fault handling unit takes over the scheduling authority of an edge node when it is offline for more than 30 minutes, and synchronizes the latest rules after the node recovers.

4. The AI-predictive-based dynamic bandwidth allocation system for multiple scenarios in universities according to claim 1, characterized in that, It also includes a bandwidth burst overload handling module, specifically including: The over-limit detection unit compares the actual flow rate with the predicted value every 10 seconds. When the actual flow rate exceeds the predicted value by 20% and continues for 30 seconds, it is determined to be a sudden over-limit. The fast elastic allocation unit can search for idle elastic bandwidth in low-priority areas within 1 second and allocate it according to the principle of proximity. The whole process takes no more than 5 seconds. The borrowing and recycling unit has a default borrowing time of no more than 15 minutes. Borrowed bandwidth will be recycled within 5 seconds after the traffic in the over-limit area returns to normal.

Images