Idle time bandwidth multiplexing control method and system based on LSTM

Abstract

The invention relates to the technical field of network bandwidth, and particularly discloses an idle time bandwidth multiplexing control method and system based on LSTM (Long Short Term Memory). According to the embodiment of the invention, the method comprises the steps: collecting bandwidth utilization rate data from a plurality of network nodes, carrying out the feature construction and data processing of the bandwidth utilization rate data, and obtaining standard feature data; obtaining historical flow data, training an LSTM model, processing the standard feature data, and outputting an idle time prediction time period; dynamically reserving a part of bandwidth resources according to an idle prediction time period, and constructing a virtual bandwidth pool; receiving bandwidth tasks, performing task classification and identification, selecting delayable tasks, and dynamically matching the delayable tasks to a virtual bandwidth pool for waiting multiplexing; and sampling influence evaluation data in real time, and optimizing the virtual bandwidth pool or task processing. According to the method, the LSTM model can be trained, the idle period can be predicted, the virtual bandwidth pool can be established, space-time dynamic matching of tasks and bandwidths is realized, and it is ensured that the multiplexing process does not affect core services.

Description

Technical Field

[0001] The present invention belongs to the technical field of network bandwidth, and in particular relates to a method and system for controlling off-peak bandwidth multiplexing based on LSTM. Background Art

[0002] Network bandwidth is an important metric for measuring a network's data transmission capacity. It typically refers to the amount of data that can be transmitted over the network per unit time, and is commonly measured in bits per second (bps), kilobits per second (kbps), megabits per second (Mbps), or gigabits per second (Gbps). Network bandwidth is not the actual transmission speed, but rather its "maximum capacity," similar to the diameter of a water pipe. The greater the bandwidth, the more data can theoretically pass through simultaneously. In practice, bandwidth is affected by a variety of factors, including network architecture, transmission protocols, hardware performance, and the number of concurrent connections.

[0003] In the prior art, the utilization of network bandwidth has the following drawbacks: (1) Network bandwidth utilization is insufficient during off-peak hours (usually less than 30%), resulting in resource waste; (2) When bandwidth is tight during peak hours, there is a lack of flexible scheduling capabilities, and the bandwidth reserved during off-peak hours cannot be dynamically reused; (3) The reuse of network bandwidth can easily affect core business. Summary of the Invention

[0004] The purpose of the embodiments of the present invention is to provide an LSTM-based off-peak bandwidth multiplexing control method and system, aiming to solve the problems raised in the background technology.

[0005] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions: The LSTM-based off-peak bandwidth reuse control method specifically includes the following steps: Collecting bandwidth usage data from multiple network nodes, and performing feature construction and data processing on the bandwidth usage data to obtain standard feature data; Obtain historical traffic data, train an LSTM model, process the standard feature data, and output the off-peak forecast time period; Dynamically reserving some bandwidth resources according to the off-peak forecast time period to build a virtual bandwidth pool; Receive bandwidth tasks, classify and identify tasks, select deferrable tasks, and dynamically match them to the virtual bandwidth pool for waiting for reuse; The impact assessment data is sampled in real time to optimize the virtual bandwidth pool or task processing.

[0006] An LSTM-based off-peak bandwidth reuse control system, which is applied to the above-mentioned LSTM-based off-peak bandwidth reuse control method, includes a data acquisition and processing unit, an off-peak prediction and processing unit, a virtual bandwidth pool construction unit, a multiplexing dynamic matching unit, and an impact assessment and feedback unit, wherein: A data collection and processing unit is used to collect bandwidth usage data from multiple network nodes, and perform feature construction and data processing on the bandwidth usage data to obtain standard feature data; An off-peak prediction processing unit is used to obtain historical traffic data, train an LSTM model, process the standard feature data, and output an off-peak prediction time period; A virtual bandwidth pool construction unit, configured to dynamically reserve some bandwidth resources according to the off-time prediction time period to construct a virtual bandwidth pool; A multiplexing dynamic matching unit is used to receive bandwidth tasks, classify and identify tasks, select tasks that can be delayed, and dynamically match them to the virtual bandwidth pool to wait for multiplexing; The impact assessment feedback unit is used to sample the impact assessment data in real time and optimize the virtual bandwidth pool or task processing.

[0007] Compared with the prior art, the present invention has the following beneficial effects: The embodiment of the present invention collects bandwidth utilization data from multiple network nodes, constructs features and processes the bandwidth utilization data to obtain standard feature data; obtains historical traffic data, trains an LSTM model, processes the standard feature data, and outputs a predicted off-peak time period; dynamically reserves some bandwidth resources based on the predicted off-peak time period to build a virtual bandwidth pool; receives bandwidth tasks, classifies and identifies them, selects deferrable tasks, and dynamically matches them to the virtual bandwidth pool for reuse; and samples impact assessment data in real time to optimize the virtual bandwidth pool or task processing. This system can train an LSTM model, predict off-peak periods, establish a virtual bandwidth pool, and achieve dynamic spatiotemporal matching of tasks and bandwidth, ensuring that the reuse process does not affect core services. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention.

[0009] Figure 1 A flow chart of a method provided by an embodiment of the present invention is shown.

[0010] Figure 2 The application architecture diagram of the system provided by the embodiment of the present invention is shown. DETAILED DESCRIPTION

[0011] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0012] It is understandable that in the prior art, the utilization of network bandwidth has the following drawbacks: (1) Network bandwidth utilization is insufficient during off-peak hours (usually less than 30%), resulting in resource waste; (2) When bandwidth is tight during peak hours, there is a lack of flexible scheduling capabilities, and the bandwidth reserved during off-peak hours cannot be dynamically reused; (3) The reuse of network bandwidth can easily affect core business.

[0013] To address the above issues, the embodiments of the present invention collect bandwidth utilization data from multiple network nodes, perform feature construction and data processing on the bandwidth utilization data, and obtain standard feature data; obtain historical traffic data, train an LSTM model, process the standard feature data, and output a predicted off-peak time period; dynamically reserve some bandwidth resources based on the predicted off-peak time period to build a virtual bandwidth pool; receive bandwidth tasks, classify and identify tasks, select deferrable tasks, and dynamically match them to the virtual bandwidth pool to await reuse; and sample impact assessment data in real time to optimize the virtual bandwidth pool or task processing. This system can train an LSTM model, predict off-peak periods, and establish a virtual bandwidth pool, achieving dynamic spatiotemporal matching of tasks and bandwidth to ensure that the reuse process does not affect core services.

[0014] Figure 1 A flow chart of a method provided by an embodiment of the present invention is shown.

[0015] Specifically, the LSTM-based off-peak bandwidth reuse control method includes the following steps: Step S101 : collecting bandwidth usage data from multiple network nodes, and performing feature construction and data processing on the bandwidth usage data to obtain standard feature data.

[0016] In an embodiment of the present invention, a data sampling period is determined, and bandwidth usage data is periodically collected from multiple network nodes according to the data sampling period. The corresponding timestamps, average values, peak values, service tags, and other data are recorded. Then, from the multiple bandwidth usage data, time feature data (including hour, day of the week, holiday signs, etc.), flow feature data (including moving average, maximum value, periodic index, etc.), and service type feature data (labeled by protocol port, such as HTTP, FTP, etc.) are extracted. Then, data cleaning and normalization processing are performed on the time feature data, flow feature data, and service type feature data to obtain standard feature data.

[0017] Specifically, in a preferred embodiment of the present invention, collecting bandwidth usage data from multiple network nodes, and performing feature construction and data processing on the bandwidth usage data to obtain standard feature data specifically includes the following steps: Determine the data sampling period; Periodically collecting bandwidth usage data from multiple network nodes according to the data sampling period; Extracting time characteristic data, flow characteristic data, service type characteristic data and periodic comprehensive vector from the plurality of bandwidth usage rate data; The time characteristic data, the flow characteristic data, the service type characteristic data and the periodic comprehensive vector are cleaned and normalized to obtain standard characteristic data.

[0018] In a preferred embodiment of the present invention, extracting a periodic integrated vector from the plurality of bandwidth usage data specifically includes the following steps: Based on the current time t, bandwidth usage data of L consecutive sampling points are intercepted forward to form a historical time window data sequence with a length of L; Perform fast Fourier transform on the historical time window data sequence to obtain the amplitude of each frequency component in the spectrum; According to the amplitude of each frequency component in the spectrum, K preset characteristic frequency points are identified, and the spectrum amplitude of the corresponding frequency points is extracted to obtain a characteristic frequency amplitude set; Determine the fundamental frequency according to the data sampling period, and calculate the relative deviation between each characteristic frequency and the fundamental frequency in the characteristic frequency amplitude set; The relative deviation is subjected to a Gaussian function to generate a weight coefficient for each frequency point to obtain a first dynamic weight set; Performing a weighted summation on the characteristic frequency amplitude set and the first dynamic weight set to characterize the dominant period intensity of the current window, thereby obtaining a weighted spectrum total value; Calculate the window standard deviation of all sample values ​​in the window of the historical time window data sequence, and use the window standard deviation to normalize the total value of the weighted spectrum to obtain the periodic intensity index; Calculate the moving average and maximum value of the historical time window data series, concatenate the periodic strength index with the moving average and maximum value to obtain a periodic comprehensive vector.

[0019] In this embodiment of the present invention, by dynamically adjusting the weight distribution of the Gaussian function, the system automatically focuses on the characteristic frequencies most relevant to the service cycle. This improves the accuracy of periodic feature recognition in scenarios where traffic patterns suddenly change during holidays. Furthermore, using the sliding window standard deviation as the denominator establishes a dynamic balance between noise intensity and periodic signals, ensuring the stability of subsequent LSTM model inputs. Using a periodic comprehensive vector as an LSTM input feature can help the LSTM identify traffic periodicity, particularly improving predictions during special periods such as holidays. Furthermore, by complementing the periodic embedded end-point indicators with traditional statistical features (such as moving mean and maximum), it can effectively accelerate LSTM model training convergence and reduce the incidence of overfitting.

[0020] Furthermore, the LSTM-based idle bandwidth multiplexing control method further includes the following steps: Step S102: Obtain historical traffic data, train an LSTM model, process the standard feature data, and output a predicted off-peak time period.

[0021] In an embodiment of the present invention, historical traffic data is obtained (which may be traffic data of the past 30 days), and is divided into a training set, a test set, and a validation set according to a preset division ratio. A loss function, an optimizer, and a training strategy are selected (Early Stopping may be used to prevent overfitting and save the optimal weight model). An LSTM model is trained using the training set, the test set, the validation set, the loss function, the optimizer, and the training strategy. The LSTM model is then used to perform off-peak prediction on the standard feature data, and the off-peak prediction time period is output.

[0022] Specifically, in a preferred embodiment of the present invention, the steps of obtaining historical traffic data, training an LSTM model, processing the standard feature data, and outputting the off-peak prediction time period specifically include the following steps: Get historical traffic data; Dividing the historical traffic data into a training set, a test set, and a validation set; Choose a loss function, optimizer, and training strategy; Training the LSTM model using the training set, test set, validation set, loss function, optimizer, and training strategy; The LSTM model is used to predict the off-peak hours of the standard feature data, and the off-peak hour prediction time period is output.

[0023] In the preferred embodiment of the present invention, the selected loss function is a dynamic perception loss function, and the construction process of the dynamic perception loss function specifically includes the following steps: Extract the holiday flag of the current timestamp t in the time feature data. If the holiday flag is a holiday, it is marked as 1, otherwise it is 0; Obtain the cycle strength index at the current time point, combine the holiday sign with the cycle strength index to obtain the basic weighting factor; Introducing learnable parameters, the basic weighting factor is fused with the learnable parameters, and 1 is added to the fusion result to ensure that the basic weight is maintained during non-holiday periods, thereby obtaining the second dynamic weight coefficient. Calculate the squared prediction error at a single time point based on the LSTM model prediction value and the actual traffic value; The second dynamic weight coefficient is used to weight the square value of the prediction error to obtain a weighted error value, and the weighted errors of all time points are averaged to obtain a time period weighted mean square error term; Count the proportions of each business type in the training set and the business type in the prediction results, calculate the relative proportion difference between the distributions of each business type in the training set and the business type in the prediction results, and obtain the relative distribution deviation of each business type; Obtain the three validation set evaluation indicators and calculate the sliding average change of the three validation set evaluation indicators; The sliding average change is transformed nonlinearly using the sigmoid function, and then the inversion operation is performed to obtain the adaptive adjustment coefficient; Sum the relative distribution deviations of all business types to obtain the total distribution deviation; The total distribution deviation is fused with the adaptive adjustment coefficient to obtain the adjusted penalty term. The time period weighted mean square error term and the adjusted penalty term constitute the dynamic perception loss function.

[0024] In this embodiment of the present invention, a periodic strength indicator can quantify the intensity of periodic fluctuations in traffic data, and a holiday flag can identify special periods such as statutory holidays and weekends. By coupling the product of the holiday flag (IsHoliday(t)) and the periodic strength indicator (C(t)), the prediction error weights for different time periods are dynamically adjusted. The service type distribution of the model's prediction results (e.g., the proportion of video, voice, and IoT data) is forced to be consistent with the distribution in the training set. By calculating relative proportional differences rather than absolute differences, the penalty term is avoided from being dominated by high-traffic services. An adaptive adjustment mechanism is also designed to dynamically adjust the strength of the service distribution penalty term based on the performance change of the validation set. When the validation index drops below zero, the adaptive adjustment coefficient increases to strengthen the constraint and prevent overfitting.

[0025] Furthermore, the LSTM-based idle bandwidth multiplexing control method further includes the following steps: Step S103: dynamically reserve some bandwidth resources according to the off-peak predicted time period to build a virtual bandwidth pool.

[0026] In an embodiment of the present invention, by analyzing the task congestion rate and task type structure of historical traffic data, recording historical impact data, and then determining a dynamic reservation ratio (between 30% and 70%) based on the historical impact data, part of the bandwidth resources are dynamically reserved according to the dynamic reservation ratio during the predicted off-peak time period. A virtual bandwidth pool is constructed using these bandwidth resources (different task channels are isolated by VLANs and tags), and the remaining resources and task usage status of the virtual bandwidth pool are regularly updated.

[0027] Specifically, in a preferred embodiment of the present invention, dynamically reserving part of the bandwidth resources according to the off-time predicted time period and constructing the virtual bandwidth pool specifically includes the following steps: Performing task congestion rate and task type structure analysis on the historical traffic data, recording historical impact data, and determining a dynamic reservation ratio based on the historical impact data; Dynamically reserving part of the bandwidth resources according to the dynamic reservation ratio during the predicted off-peak time period; A virtual bandwidth pool is constructed using the portion of bandwidth resources.

[0028] In a preferred embodiment of the present invention, performing task congestion rate and task type structure analysis on the historical traffic data, recording historical impact data, and determining the dynamic reservation ratio based on the historical impact data specifically include the following steps: Count the total number of all historical tasks in the current period in the historical traffic data; Compare the delay data of each task in the total number of historical tasks with the network delay threshold. If the task delay data exceeds the network delay threshold, it is marked as a congested task; Calculate the ratio of the number of congested tasks to the total number of tasks to obtain the time period congestion rate; Count the service type distribution of each service type in the historical traffic data, and calculate the degree of chaos of the service type distribution based on information entropy to obtain the original entropy value; The original entropy value is normalized using the theoretical maximum entropy value to obtain the normalized business entropy value; The cyclical strength index is transformed nonlinearly using the Sigmoid function, and then the nonlinear transformation result is adjusted according to the time period type of the cyclical strength index to obtain the normalized cyclical strength; The 95th percentile is extracted from historical traffic data as the upper threshold. The LSTM model prediction result is compared with the 95th percentile. If the LSTM model prediction result exceeds the 95th percentile, it is forcibly set to the 95th percentile to obtain the corrected off-peak bandwidth baseline value. Calculate the gradient of the loss function with respect to the reserved ratio during model training, determine the adjustment amplitude according to the learning rate, determine the adjustment direction according to the gradient direction, adjust the learnable parameters according to the adjustment direction and adjustment amplitude, and obtain the adjusted learnable parameters; The gradient is mapped through the Sigmoid function, and the adaptive adjustment coefficient is adjusted based on the mapping result to obtain the adjusted adaptive adjustment coefficient; The congestion rate of the time period is combined with the adjusted learnable parameters to calculate the congestion attenuation term; The normalized business entropy value is combined with the cycle intensity to calculate the business entropy cycle term; The corrected off-peak bandwidth baseline value is combined with the adjusted adaptive adjustment coefficient to obtain the baseline guarantee value; The congestion attenuation term is combined with the service entropy period term and superimposed with the baseline guarantee value to obtain the final dynamic reservation ratio.

[0029] In this embodiment, a dynamic resource allocation model is constructed by integrating four types of data: congestion rate analysis, service type entropy, cycle intensity index, and predicted baseline values. These four types of data correspond to the four dimensions of network status, service structure, temporal patterns, and predictive capability, forming an orthogonal feature space that effectively avoids decision-making bias.

[0030] The congestion decay term and the service entropy period term are used to achieve nonlinear coupling: when the congestion rate increases, the exponential decay term responds quickly, giving priority to key services during strong periodic periods (such as weekday peak hours), automatically increasing the weight of service diversity, and preventing periodic tasks from crowding out resources.

[0031] It also shares learnable parameters and adaptive adjustment coefficients with the LSTM training process, forming cross-module parameter transfer, so that the resource allocation strategy and the prediction model are optimized collaboratively.

[0032] Furthermore, the LSTM-based idle bandwidth multiplexing control method further includes the following steps: Step S104: receiving bandwidth tasks, performing task classification and identification, selecting deferrable tasks, and dynamically matching them to the virtual bandwidth pool to wait for reuse.

[0033] In an embodiment of the present invention, bandwidth tasks are received, classified, identified, and judged. If the bandwidth task is a real-time task (such as a video conference), it is processed through a preset high-priority channel of the main link. If the bandwidth task is a deferrable task (such as a system update or data backup), it is dynamically matched to a virtual bandwidth pool and added to a scheduling queue for reuse.

[0034] Specifically, in a preferred embodiment of the present invention, receiving bandwidth tasks, classifying and identifying tasks, selecting deferrable tasks, and dynamically matching them to the virtual bandwidth pool for waiting for reuse specifically include the following steps: Receive bandwidth tasks; Performing task classification, identification, and judgment on the bandwidth task; If the bandwidth task is a real-time task, it is processed through the preset main link high priority channel; If the bandwidth task is a deferrable task, dynamically matching it to the virtual bandwidth pool; The delayable task is added to the scheduling queue to wait for reuse.

[0035] In a preferred embodiment of the present invention, if the bandwidth task is a deferrable task, dynamically matching it to the virtual bandwidth pool specifically includes the following steps: Preset the maximum allowable delay time for each type of task, and calculate the remaining delay time based on the current task delay and the maximum allowable delay time for each type of task, to obtain the remaining delay time for the current task; Calculate the remaining delay time ratio based on the remaining delay time of the current task and the maximum allowable delay time of each task; The adaptive adjustment coefficient is used as the exponential base to perform exponential operation on the remaining delay time ratio. If the exponential operation result exceeds 1, it is forced to be truncated to 1; if the exponential operation result is less than 0, it is set to 0 to obtain the delay tolerance gain term; According to the current task type identifier, extract the proportion of the current task type in the historical traffic data, and calculate the entropy contribution of the current task type based on the proportion of the current task type in the historical traffic data; Normalize the entropy contribution of the current task type using the theoretical maximum entropy value to obtain the normalized business entropy value; The normalized service entropy value is combined with the final dynamic reservation ratio to obtain the preliminary synergy coefficient; Add 1 to the preliminary synergy coefficient to ensure that the basic weight is retained when the final dynamic reservation ratio is 0. If the calculated result exceeds 2, it is forcibly truncated to 2 to obtain the entropy pool synergy term; Obtain real-time monitoring data of the dispatch queue, calculate the queue load rate, and use the queue load rate as the queue pressure penalty item; The task matching degree is obtained by comprehensive calculation based on the delay tolerance gain term, entropy pool cooperation term, and queue pressure penalty term; The task matching degree is compared with the preset matching threshold. If the task matching degree is greater than the preset matching threshold, the task is added to the scheduling queue. If the task matching degree is less than the preset matching threshold, it is rejected from entering the virtual bandwidth pool and transferred to the backup link for processing, and a task scheduling instruction is generated.

[0036] In this embodiment of the present invention, a dynamic matching model is constructed by integrating four parameters: task delay tolerance, service type entropy, resource reservation status, and queue pressure. These parameters reflect task real-time requirements, service diversity requirements, resource availability, and system load, respectively, forming complementary decision-making criteria. Nonlinear coupling is achieved through exponential operation (delay tolerance gain term) and dynamic product (entropy pool coordination term). The exponential design causes task priority to decay nonlinearly with the remaining delay time. For example, the priority decreases more slowly when the remaining delay time decreases from 120s to 60s than from 60s to 0s, thus avoiding sudden changes in thresholds. The product design of entropy value and reservation ratio amplifies the impact of service diversity when resources are abundant (ultimately high dynamic reservation ratio) and automatically converges to the basic weight when resources are scarce (ultimately low dynamic reservation ratio). Queue pressure data is used to adjust the resource reservation policy in real time. When the queue load exceeds 80%, a penalty term significantly reduces matching efficiency, reducing the injection of new tasks. Consecutive task rejection events trigger dynamic adjustment of the virtual bandwidth pool capacity, forming a complete closed loop from task scheduling to resource optimization.

[0037] Furthermore, the LSTM-based idle bandwidth multiplexing control method further includes the following steps: Step S105: Sampling impact assessment data in real time to optimize the virtual bandwidth pool or task processing.

[0038] In an embodiment of the present invention, impact assessment data, including sampling delay, packet loss rate, and jitter, is sampled in real time. Based on preset standard assessment data, the impact assessment data is compared to determine whether there is an indicator deterioration. If it is determined that there is an indicator deterioration, the virtual bandwidth pool is reduced or the task execution is canceled, and the data is fed back to the LSTM model for training and optimization.

[0039] Specifically, in a preferred embodiment of the present invention, the real-time sampling of impact assessment data and the optimization of the virtual bandwidth pool and task processing specifically include the following steps: Real-time sampling impact assessment data, including sampling delay, packet loss rate, and jitter; Comparing the impact assessment data based on preset standard assessment data to determine whether there is an indicator deterioration; When it is determined that the indicator has deteriorated, the virtual bandwidth pool is reduced or the task execution is canceled.

[0040] Further, Figure 2 The application architecture diagram of the system provided by the embodiment of the present invention is shown.

[0041] Among them, in another preferred embodiment provided by the present invention, an LSTM-based idle bandwidth reuse control system is applied to the above-mentioned LSTM-based idle bandwidth reuse control method, and the system includes: The data collection and processing unit 101 is configured to collect bandwidth usage data from a plurality of network nodes, and perform feature construction and data processing on the bandwidth usage data to obtain standard feature data.

[0042] In an embodiment of the present invention, the data acquisition and processing unit 101 determines a data sampling period, periodically collects bandwidth usage data from multiple network nodes according to the data sampling period, and records corresponding timestamps, average values, peak values, service tags, and other data. Time feature data (including hour, day of the week, holiday signs, etc.), flow feature data (including moving average, maximum value, periodic index, etc.), and service type feature data (labeled by protocol port, such as HTTP, FTP, etc.) are then extracted from the multiple bandwidth usage data. Data cleaning and normalization processing is then performed on the time feature data, flow feature data, and service type feature data to obtain standard feature data.

[0043] The off-peak prediction processing unit 102 is used to obtain historical traffic data, train the LSTM model, process the standard feature data, and output the off-peak prediction time period.

[0044] In an embodiment of the present invention, the off-peak prediction processing unit 102 obtains historical traffic data (which may be traffic data of the past 30 days), divides the historical traffic data into a training set, a test set and a validation set according to a preset division ratio, and selects a loss function, an optimizer and a training strategy (which may be to use Early Stopping to prevent overfitting and save the optimal weight model). The LSTM model is trained through the training set, the test set, the validation set, the loss function, the optimizer and the training strategy, and then the standard feature data is predicted during off-peak hours through the LSTM model to output the off-peak prediction time period.

[0045] The virtual bandwidth pool building unit 103 is configured to dynamically reserve some bandwidth resources according to the off-time prediction time period to build a virtual bandwidth pool.

[0046] In the embodiment of the present invention, the virtual bandwidth pool construction unit 103 analyzes the task congestion rate and task type structure of historical traffic data, records historical impact data, and then determines a dynamic reservation ratio (between 30% and 70%) based on the historical impact data. Furthermore, during the predicted off-peak time period, the unit dynamically reserves some bandwidth resources according to the dynamic reservation ratio. Using these bandwidth resources, the unit constructs a virtual bandwidth pool (using VLANs and tags to isolate different task channels). The unit also regularly updates the remaining resources and task usage status of the virtual bandwidth pool.

[0047] The multiplexing dynamic matching unit 104 is configured to receive bandwidth tasks, perform task classification and identification, select deferrable tasks, and dynamically match them to the virtual bandwidth pool for waiting for multiplexing.

[0048] In an embodiment of the present invention, the multiplexing dynamic matching unit 104 receives bandwidth tasks, classifies and identifies the bandwidth tasks, and processes them through a preset high-priority channel of the main link when the bandwidth tasks are real-time tasks (such as video conferencing). When the bandwidth tasks are deferrable tasks (such as system updates, data backups, etc.), they are dynamically matched to a virtual bandwidth pool and added to a scheduling queue for multiplexing.

[0049] The impact assessment feedback unit 105 is configured to sample the impact assessment data in real time and optimize the virtual bandwidth pool or task processing.

[0050] In an embodiment of the present invention, the impact assessment feedback unit 105 samples the impact assessment data in real time, including sampling delay, packet loss rate, and jitter, and compares the impact assessment data based on preset standard assessment data to determine whether there is an indicator deterioration. When it is determined that there is an indicator deterioration, the virtual bandwidth pool is reduced or the task execution is canceled, and the feedback is fed back to the LSTM model to train and optimize the LSTM model.

[0051] It should be understood that, although the various steps in the flow chart of each embodiment of the present invention are shown in sequence according to the indication of the arrows, these steps are not necessarily performed in sequence according to the order indicated by the arrows. Unless otherwise specified herein, the execution of these steps is not strictly limited in order, and these steps can be performed in other orders. Moreover, at least a portion of the steps in each embodiment may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these sub-steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a portion of other steps or sub-steps or stages of other steps.

[0052] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When executed, the program can include the processes of the above-described method embodiments. Any reference to memory, storage, database, or other media used in the various embodiments provided herein may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM).

[0053] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0054] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

[0055] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. The off-peak bandwidth reuse control method based on LSTM is characterized by: The method specifically comprises the following steps: Collecting bandwidth usage data from multiple network nodes, and performing feature construction and data processing on the bandwidth usage data to obtain standard feature data; Obtain historical traffic data, train an LSTM model, process the standard feature data, and output the off-peak forecast time period; Dynamically reserving some bandwidth resources according to the off-peak forecast time period to build a virtual bandwidth pool; Receive bandwidth tasks, classify and identify tasks, select deferrable tasks, and dynamically match them to the virtual bandwidth pool for waiting for reuse; Sampling impact assessment data in real time to optimize the virtual bandwidth pool or task processing; The collecting of bandwidth usage data from multiple network nodes, and performing feature construction and data processing on the bandwidth usage data to obtain standard feature data specifically includes the following steps: Determine the data sampling period; Periodically collecting bandwidth usage data from multiple network nodes according to the data sampling period; Extracting time characteristic data, flow characteristic data, service type characteristic data and periodic comprehensive vector from the plurality of bandwidth usage rate data; The time characteristic data, the flow characteristic data, the service type characteristic data and the periodic comprehensive vector are cleaned and normalized to obtain standard characteristic data.

2. The LSTM-based idle time bandwidth multiplexing control method according to claim 1 is characterized in that: Extracting a periodic integrated vector from the plurality of bandwidth usage data specifically includes the following steps: Based on the current time t, bandwidth usage data of L consecutive sampling points are intercepted forward to form a historical time window data sequence with a length of L; Perform fast Fourier transform on the historical time window data sequence to obtain the amplitude of each frequency component in the spectrum; According to the amplitude of each frequency component in the spectrum, K preset characteristic frequency points are identified, and the spectrum amplitude of the corresponding frequency points is extracted to obtain a characteristic frequency amplitude set; Determine the fundamental frequency according to the data sampling period, and calculate the relative deviation between each characteristic frequency and the fundamental frequency in the characteristic frequency amplitude set; The relative deviation is subjected to a Gaussian function to generate a weight coefficient for each frequency point to obtain a first dynamic weight set; Performing a weighted summation on the characteristic frequency amplitude set and the first dynamic weight set to characterize the dominant period intensity of the current window, thereby obtaining a weighted spectrum total value; Calculate the window standard deviation of all sample values ​​in the window of the historical time window data sequence, and use the window standard deviation to normalize the total value of the weighted spectrum to obtain the periodic intensity index; Calculate the moving average and maximum value of the historical time window data series, concatenate the periodic strength index with the moving average and maximum value to obtain a periodic comprehensive vector.

3. The LSTM-based idle time bandwidth multiplexing control method according to claim 2, characterized in that: The steps of acquiring historical traffic data, training the LSTM model, processing the standard feature data, and outputting the off-peak forecast time period specifically include the following steps: Get historical traffic data; Dividing the historical traffic data into a training set, a test set, and a validation set; Choose a loss function, optimizer, and training strategy; Training the LSTM model using the training set, test set, validation set, loss function, optimizer, and training strategy; The LSTM model is used to predict the off-peak hours of the standard feature data, and the off-peak hour prediction time period is output.

4. The LSTM-based idle time bandwidth multiplexing control method according to claim 3 is characterized in that: The loss function selected is the dynamic perception loss function. The construction process of the dynamic perception loss function specifically includes the following steps: Extract the holiday flag of the current timestamp t in the time feature data. If the holiday flag is a holiday, it is marked as 1, otherwise it is 0; Obtain the cycle strength index at the current time point, combine the holiday sign with the cycle strength index to obtain the basic weighting factor; Introducing learnable parameters, the basic weighting factor is fused with the learnable parameters, and 1 is added to the fusion result to ensure that the basic weight is maintained during non-holiday periods, thereby obtaining the second dynamic weight coefficient. Calculate the squared prediction error at a single time point based on the LSTM model prediction value and the actual traffic value; The second dynamic weight coefficient is used to weight the square value of the prediction error to obtain a weighted error value, and the weighted errors of all time points are averaged to obtain a time period weighted mean square error term; Count the proportions of each business type in the training set and the business type in the prediction results, calculate the relative proportion difference between the distributions of each business type in the training set and the business type in the prediction results, and obtain the relative distribution deviation of each business type; Obtain the three validation set evaluation indicators and calculate the sliding average changes of the three validation set evaluation indicators; perform a nonlinear transformation on the sliding average changes using the sigmoid function and then perform the inversion operation to obtain the adaptive adjustment coefficient; The relative distribution deviations of all business types are summed to obtain the total distribution deviation; the total distribution deviation is integrated with the adaptive adjustment coefficient to obtain the adjusted penalty term. The time period weighted mean square error term and the adjusted penalty term constitute the dynamic perception loss function.

5. The LSTM-based off-peak bandwidth multiplexing control method according to claim 4 is characterized in that: Dynamically reserving part of the bandwidth resources according to the off-peak predicted time period to build a virtual bandwidth pool specifically includes the following steps: Performing task congestion rate and task type structure analysis on the historical traffic data, and recording historical impact data; Determining a dynamic reservation ratio based on the historical impact data; Dynamically reserving part of the bandwidth resources according to the dynamic reservation ratio during the predicted off-peak time period; A virtual bandwidth pool is constructed using the portion of bandwidth resources.

6. The LSTM-based off-peak bandwidth multiplexing control method according to claim 5, characterized in that: The performing of task congestion rate and task type structure analysis on the historical traffic data, recording historical impact data, and determining the dynamic reservation ratio based on the historical impact data specifically includes the following steps: Count the total number of all historical tasks in the current period in the historical traffic data; Compare the delay data of each task in the total number of historical tasks with the network delay threshold. If the task delay data exceeds the network delay threshold, it is marked as a congested task. Calculate the ratio of the number of congested tasks to the total number of tasks to obtain the congestion rate of the period. The service type distribution of each service type in historical traffic data is statistically analyzed, and the degree of disorder of the service type distribution is calculated based on information entropy to obtain the original entropy value. The original entropy value is normalized using the theoretical maximum entropy value to obtain the normalized service entropy value. The periodic intensity index is nonlinearly transformed using the Sigmoid function, and the nonlinear transformation result is adjusted according to the time period type of the periodic intensity index to obtain the normalized periodic intensity. The 95th percentile is extracted from historical traffic data as the upper threshold. The LSTM model prediction result is compared with the 95th percentile. If the LSTM model prediction result exceeds the 95th percentile, it is forcibly set to the 95th percentile to obtain the corrected off-peak bandwidth baseline value. Calculate the gradient of the loss function with respect to the reserved ratio during model training, determine the adjustment amplitude according to the learning rate, determine the adjustment direction according to the gradient direction, adjust the learnable parameters according to the adjustment direction and adjustment amplitude, and obtain the adjusted learnable parameters; The gradient is mapped through the Sigmoid function, and the adaptive adjustment coefficient is adjusted based on the mapping result to obtain the adjusted adaptive adjustment coefficient; The congestion rate of the time period is combined with the adjusted learnable parameter to calculate the congestion attenuation term; the normalized service entropy value is combined with the normalized period intensity to calculate the service entropy period term; The corrected off-peak bandwidth baseline value is combined with the adjusted adaptive adjustment coefficient to obtain the baseline guarantee value; the congestion attenuation term is combined with the service entropy period term and superimposed with the baseline guarantee value to obtain the final dynamic reservation ratio.

7. The LSTM-based off-peak bandwidth multiplexing control method according to claim 6, characterized in that: The receiving of bandwidth tasks, performing task classification and identification, selecting deferrable tasks, and dynamically matching them to the virtual bandwidth pool for waiting for reuse specifically includes the following steps: Receive bandwidth tasks; Performing task classification, identification, and judgment on the bandwidth task; If the bandwidth task is a real-time task, it is processed through the preset main link high priority channel; If the bandwidth task is a deferrable task, dynamically matching it to the virtual bandwidth pool; The delayable task is added to the scheduling queue to wait for reuse.

8. The LSTM-based off-peak bandwidth multiplexing control method according to claim 7, characterized in that: If the bandwidth task is a deferrable task, dynamically matching it to the virtual bandwidth pool specifically includes the following steps: Preset the maximum allowable delay time for each type of task, and calculate the remaining delay time based on the current task delay and the maximum allowable delay time for each type of task, to obtain the remaining delay time for the current task; Calculate the remaining delay time ratio based on the remaining delay time of the current task and the maximum allowable delay time of each task; The adaptive adjustment coefficient is used as the exponential base to perform exponential operation on the remaining delay time ratio. If the exponential operation result exceeds 1, it is forced to be truncated to 1; if the exponential operation result is less than 0, it is set to 0 to obtain the delay tolerance gain term; According to the current task type identifier, extract the proportion of the current task type in the historical traffic data, and calculate the entropy contribution of the current task type based on the proportion of the current task type in the historical traffic data; Normalize the entropy contribution of the current task type using the theoretical maximum entropy value to obtain the normalized business entropy value; The normalized service entropy value is combined with the final dynamic reservation ratio to obtain the preliminary synergy coefficient; Add 1 to the preliminary synergy coefficient to ensure that the basic weight is retained when the final dynamic reservation ratio is 0. If the calculated result exceeds 2, it is forcibly truncated to 2 to obtain the entropy pool synergy term; Obtain real-time monitoring data of the dispatch queue, calculate the queue load rate, and use the queue load rate as the queue pressure penalty item; The task matching degree is obtained by comprehensive calculation based on the delay tolerance gain term, entropy pool cooperation term, and queue pressure penalty term; The task matching degree is compared with the preset matching threshold. If the task matching degree is greater than the preset matching threshold, the task is added to the scheduling queue; if the task matching degree is less than the preset matching threshold, it is rejected from entering the virtual bandwidth pool and transferred to the backup link for processing, and a task scheduling instruction is generated.

9. The LSTM-based off-peak bandwidth multiplexing control method according to claim 8, characterized in that: The real-time sampling impact assessment data and the optimization of the virtual bandwidth pool and task processing specifically include the following steps: Real-time sampling impact assessment data, including sampling delay, packet loss rate, and jitter; Comparing the impact assessment data based on preset standard assessment data to determine whether there is an indicator deterioration; When it is determined that the indicator has deteriorated, the virtual bandwidth pool is reduced or the task execution is canceled.

10. An LSTM-based off-peak bandwidth reuse control system, the system being applied to the LSTM-based off-peak bandwidth reuse control method according to any one of claims 1 to 9, characterized in that: The system includes a data acquisition and processing unit, an off-peak prediction and processing unit, a virtual bandwidth pool construction unit, a multiplexing dynamic matching unit, and an impact assessment and feedback unit, wherein: A data collection and processing unit is used to collect bandwidth usage data from multiple network nodes, and perform feature construction and data processing on the bandwidth usage data to obtain standard feature data; An off-peak prediction processing unit is used to obtain historical traffic data, train an LSTM model, process the standard feature data, and output an off-peak prediction time period; A virtual bandwidth pool construction unit, configured to dynamically reserve some bandwidth resources according to the off-time prediction time period to construct a virtual bandwidth pool; A multiplexing dynamic matching unit is used to receive bandwidth tasks, classify and identify tasks, select tasks that can be delayed, and dynamically match them to the virtual bandwidth pool to wait for multiplexing; The impact assessment feedback unit is used to sample the impact assessment data in real time and optimize the virtual bandwidth pool or task processing.

Images