Discrete service flow predictive digital triggering method and system under time delay constraint

Abstract

The invention discloses a discrete service flow predictive digital triggering method and system under time delay constraint, and relates to the field of network service quality assurance, and the method comprises the steps: collecting an arrival time sequence, a packet length sequence, network state data and service level protocol parameters of a plurality of discrete service flows; performing multi-scale dependency decomposition on the arrival time sequence, and predicting an arrival amount and a queuing length in a future time window; calculating and predicting a residual time delay budget in combination with the maximum tolerant time delay, and mapping the residual time delay budget into a digital triggering probability representing the scheduling urgency degree; and further fusing the stream importance index to determine the scheduling weight, and dynamically adjusting the service sequence and resource allocation of the edge nodes according to the scheduling weight. Therefore, advanced scheduling triggering of multiple paths of service flows is realized, and the time delay guarantee capability and the resource scheduling efficiency are improved.

Description

Technical Field

[0001] This application relates to the field of network service quality assurance technology, and in particular to a predictive digital triggering method and system for discrete service flows under time delay constraints. Background Technology

[0002] With the development of technologies such as cloud computing, edge computing, and the Industrial Internet of Things (IIoT), the types of services carried in modern network systems are becoming increasingly diverse, with a large number of services appearing in the form of discrete traffic flows. These traffic flows are typically characterized by strong bursts, uneven arrival times, and significant differences in service priorities, thus placing higher demands on network systems in terms of latency control, throughput assurance, and transmission reliability. Against this backdrop, how to achieve more refined and real-time scheduling and control for discrete traffic flows has become a crucial technical issue in network service assurance.

[0003] Currently, common practices for ensuring service quality (SQW) for business flows typically include scheduling mechanisms based on static thresholds, fixed priorities, or preset rules. While these solutions can achieve basic resource allocation and service control in scenarios with relatively stable business loads, they often rely primarily on the current state to respond to highly volatile and discrete business flows. This lack of proactive awareness of short-term traffic changes and potential queuing delays can easily lead to scheduling lags, impacting the effectiveness of SQW under latency constraints. Furthermore, such passive response mechanisms often struggle to consider overall system energy consumption and lack the ability to finely differentiate the dynamic value of business flows when resources are scarce.

[0004] In addition, some solutions attempt to introduce machine learning or deep learning methods to predict changes in business flow, thereby improving the foresight of scheduling decisions. However, one type of method relies heavily on manual feature construction and has limited ability to adapt to complex dynamic scenarios; another type of method, while possessing strong data fitting capabilities, has relatively high model complexity and often struggles to balance real-time inference requirements with practical application costs in deployment environments with limited computing power and timeliness (such as edge devices). Summary of the Invention

[0005] This application provides a predictive digital triggering method, system, storage medium, computer program product, and electronic device for discrete service flows under time-delay constraints, in order to at least solve the problems of insufficient forward-looking network scheduling and insufficient differentiated service guarantee capabilities in discrete service flow scenarios in the prior art.

[0006] In a first aspect, embodiments of this application provide a predictive digital triggering method for discrete service flows under latency constraints. The method includes: acquiring multiple discrete service flows in an edge network, and real-time collecting flow feature data, network status data, and service level protocol parameters corresponding to the multiple discrete service flows; the flow feature data includes arrival time series and packet length series, the network status data includes current queue length and available bandwidth, and the service level protocol parameters include maximum tolerable latency; performing multi-scale dependency decomposition on the arrival time series in the flow feature data of each discrete service flow to separate arrival sequence components at different time scales; based on the separated arrival sequence components and the packet length series, predicting the predicted arrival amount of each discrete service flow within a future preset time window, and estimating the arrival amount of each discrete service flow in conjunction with the current queue length and the available bandwidth. The predicted queue length of the flow within the preset future time window; for each discrete service flow, the predicted remaining latency budget of the discrete service flow is calculated based on the predicted queue length and the maximum tolerable latency in the service level agreement parameters, and the predicted remaining latency budget is mapped to a digital trigger probability; the digital trigger probability is used to characterize the scheduling urgency of each discrete service flow; the flow importance index of each discrete service flow is obtained, and the scheduling weight of each discrete service flow is determined by combining the digital trigger probability and the flow importance index corresponding to each discrete service flow; the flow importance index is used to characterize the service importance of the corresponding discrete service flow; based on the scheduling weight corresponding to each discrete service flow, the service order of the scheduling queue in the edge node is dynamically adjusted and corresponding local resources are allocated to trigger the scheduling operation for the multiple discrete service flows in advance.

[0007] Secondly, embodiments of this application provide a predictive digital triggering system for discrete service flows under latency constraints. The system includes: a service flow sensing unit, used to acquire multiple discrete service flows in an edge network and collect in real time the flow feature data, network status data, and service level protocol parameters corresponding to the multiple discrete service flows; the flow feature data includes arrival time series and packet length series, the network status data includes current queue length and available bandwidth, and the service level protocol parameters include maximum tolerable latency; a multi-scale decomposition unit, used to perform multi-scale dependency decomposition on the arrival time series in the flow feature data of each discrete service flow to separate arrival sequence components at different time scales; and an arrival prediction and queuing estimation unit, used to predict the predicted arrival amount of each discrete service flow within a future preset time window based on the separated arrival sequence components and the packet length series, and estimate the arrival amount of each discrete service flow in conjunction with the current queue length and the available bandwidth. The system comprises: a predicted queuing length for each discrete service flow within a future preset time window; a latency budget triggering unit, used to calculate the predicted remaining latency budget for each discrete service flow based on the predicted queuing length and the maximum tolerable latency in the service level agreement parameters, and mapping the predicted remaining latency budget to a digital trigger probability; the digital trigger probability is used to characterize the scheduling urgency of each discrete service flow; a scheduling weight determination unit, used to obtain the flow importance index of each discrete service flow, and determine the scheduling weight of each discrete service flow by combining the digital trigger probability and the flow importance index; the flow importance index is used to characterize the service importance of the corresponding discrete service flow; and a resource scheduling execution unit, used to dynamically adjust the service order of the scheduling queue within the edge node and allocate corresponding local resources based on the scheduling weight corresponding to each discrete service flow, so as to trigger the scheduling operation for the multiple discrete service flows in advance.

[0008] Thirdly, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the delay-constrained discrete traffic flow predictive digital triggering method of any embodiment of the present application.

[0009] Fourthly, embodiments of this application provide a storage medium storing a computer program thereon, characterized in that, when the program is executed by a processor, it implements the steps of the delay-constrained predictive digital triggering method for discrete service flows according to any embodiment of this application.

[0010] Fifthly, embodiments of this application provide a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the delay-constrained predictive digital triggering method for discrete service flows according to any embodiment of this application.

[0011] The predictive digital triggering method and system for discrete traffic flows under time-delay constraints provided in this application can achieve at least the following technical effects:

[0012] (1) Based on the short-term dynamic evolution of discrete business flows, this scheme constructs a forward-looking analysis link that connects multi-scale dependency decomposition, future arrival volume prediction, predicted queue length estimation, and predicted remaining delay budget. By separating the different time-scale change characteristics in the arrival time series of business flows and combining the group length sequence, current queue length, and available bandwidth, this scheme achieves a joint characterization of the queuing situation within the future window, extending the system's understanding of the business flow status from instantaneous observation at the current moment to the evolution trend in the future time domain. On this basis, the maximum tolerable delay is further used to constrain the predicted queuing results, forming a predicted remaining delay budget directly related to the business delay boundary. Thus, the delay risk of discrete business flows can be identified in advance before the queuing process actually deteriorates, thereby giving scheduling control stronger predictability and time-shifting capability, and significantly improving the service assurance stability and response effectiveness in delay-sensitive scenarios.

[0013] (2) The predicted remaining latency budget is mapped to a digital trigger probability and synergistically integrated with the flow importance index, which represents the difference in business value, to generate scheduling weights for queue service order adjustment and local resource allocation. This breaks the limitation of triggering scheduling from only the single dimension of future latency pressure. Instead, it uses the digital trigger probability to represent the urgency within the future window and the flow importance index to reflect the relative importance of different businesses in the system service objectives. As a result, the final scheduling basis can simultaneously cover the two core dimensions of "the strength of latency constraints" and "the level of business value". Based on this comprehensive weight, dynamic queue rearrangement and targeted resource allocation are implemented. Edge nodes can organize service order more effectively under the complex conditions of multiple discrete business concurrency, so that limited resources are prioritized for business flows with greater guarantee significance. This significantly improves the refinement of resource allocation and the matching degree between scheduling strategies and business objectives.

[0014] This technical solution establishes a predictive digital triggering scheduling mechanism comprised of future queuing trend perception, digital representation of latency budget, and integrated decision-making based on business importance. This elevates the scheduling and control of discrete service flows beyond simple state responses or static rule allocation, forming an integrated technical path of proactive triggering, quantitative judgment, and dynamic adjustment oriented towards future windows. Consequently, it achieves a joint characterization of latency risk and business value for discrete service flows, fundamentally enhancing the foresight, differentiated guarantee capabilities, and adaptability of service scheduling in edge network environments. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart is shown as an example of a predictive digital triggering method for discrete traffic flows under delay constraints according to an embodiment of this application.

[0017] Figure 2 A flowchart illustrating an example of calculating the probability of digital triggering in a method according to an embodiment of this application is shown.

[0018] Figure 3 A flowchart illustrating an example of performing a parameter closed-loop dynamic correction operation in a method according to an embodiment of this application is shown.

[0019] Figure 4 A schematic diagram illustrating the system operation mechanism of an example of a delay-constrained predictive digital triggering method for discrete traffic flows according to an embodiment of this application is shown.

[0020] Figure 5 A schematic diagram of the comparative experimental simulation results of the cumulative distribution probability of queuing delay for different methods is shown.

[0021] Figure 6 A schematic diagram of the experimental simulation results comparing the microscopic transient response of different methods to the bursts of discrete traffic flows is shown;

[0022] Figure 7 The diagram shows the experimental simulation results comparing the system energy efficiency and computational cost of different methods.

[0023] Figure 8 A structural block diagram of an example of a delay-constrained predictive digital triggering system for discrete traffic flows according to an embodiment of this application is shown. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] It should be noted that, in the specific implementation of business traffic prediction, the linear prediction models widely used in related technologies (such as autoregressive integral moving average models, exponential smoothing, etc.) perform reasonably well when processing data with linear and stationary characteristics. However, some studies have pointed out that such methods often perform poorly when dealing with highly nonlinear and bursty complex discrete traffic in the IoT environment, and they heavily rely on human experience to extract time-frequency domain features, resulting in weak generalization ability. Meanwhile, although various deep neural networks have been introduced to attempt to automatically extract complex traffic patterns, they all face bottlenecks in latency-sensitive scenarios with extremely stringent real-time requirements. For example, some solutions based on recurrent neural networks suffer from the drawback of non-parallel computation; while models based on self-attention mechanisms can capture long-term dependencies well, their quadratic computational complexity greatly limits their deployment on edge devices. Even though some cutting-edge research attempts to introduce approximation algorithms to reduce computational complexity, this inevitably leads to a significant decrease in traffic prediction accuracy, failing to provide reliable forward-looking support for microsecond-level latency constraints.

[0026] At the level of specific business scheduling execution, where effective predictive methods are lacking, some research in related technologies has proposed scheduling schemes based on latency and service quality awareness. These schemes typically analyze the underlying queue buffer state of each business flow and use queue latency to calculate scheduling weights, aiming to balance latency growth with total system throughput. However, these schemes essentially still rely on fixed weight evaluation logic and hard thresholds. Due to the lack of predictive triggering mechanisms for the evolution trends of discrete business flows, their scheduling performance is highly susceptible to drastic traffic fluctuations, making proactive intervention before queuing deteriorates is difficult.

[0027] Furthermore, in order to control system energy consumption while ensuring service quality, some studies have begun to explore the application of intelligent methods such as reinforcement learning to resource scheduling in open network architectures. However, in these solutions, the focus of intelligent dynamic adjustment is mostly concentrated on the macro-level placement of virtual network functions and cross-node migration, rather than fine-grained scheduling at the micro-level of business flows. At the underlying actual business processing level, it still relies on traditional queue scheduling logic, and cannot implement predictive latency control for the bursty characteristics of individual discrete business flows.

[0028] In summary, current technologies generally lack predictive digital triggering mechanisms for discrete business flows, making it difficult to accurately perceive future queuing trends with low computational overhead in edge scenarios where computing power is limited. At the same time, existing mechanisms have also failed to establish a comprehensive scheduling and control strategy that integrates latency evolution budget, node energy consumption costs, and business differentiation value.

[0029] It should be understood that the above description of the relevant technologies is intended only to help the public better understand the inventive spirit and motivation of this application, and is not intended to limit this application. Furthermore, the technical solutions described in the above-mentioned relevant technologies are not prior art, and may also be undisclosed technical solutions, such as those under research or in the laboratory stage.

[0030] The technical solutions in this application, including the collection, storage, use, processing, transmission, provision, and disclosure of users' personal information, comply with relevant laws and regulations and do not violate public order and good morals.

[0031] Figure 1 A flowchart illustrating an example of a delay-constrained predictive digital triggering method for discrete traffic flows according to an embodiment of this application is shown.

[0032] Regarding the execution subject of the method in the embodiments of this application, it can be any controller or processor with computing or processing capabilities, such as a local scheduling controller or edge processing unit deployed on the edge network side. It can implement the various data flow prediction and scheduling steps in the embodiments of this application by calling and executing computer program instructions stored in memory, or by relying on internally integrated dedicated logic acceleration circuits.

[0033] In some examples, it can be integrated into a computing device or network node through software, hardware, or a combination of both. The type of computing device or network node can be diverse. For example, the computing device can be an edge server, a base station baseband processing unit (BBU), a multi-access edge computing (MEC) device, an industrial IoT local gateway, or various edge-side devices with distributed network communication and near-end data processing capabilities.

[0034] like Figure 1 As shown, in step S110, multiple discrete service flows in the edge network are acquired, and the flow characteristic data, network status data, and service level protocol parameters corresponding to the multiple discrete service flows are collected in real time.

[0035] Here, the flow characteristic data includes arrival time series and packet length series, the network state data includes current queue length and available bandwidth, and the service level agreement parameters include maximum tolerable latency. It should be noted that these three types of data together constitute the multi-dimensional state-aware space of the predictive scheduling in this application. Specifically, the arrival time series in the flow characteristic data is used to characterize the burst and fluctuation patterns of discrete service flows at both micro and macro scales; the packet length series provides a computational benchmark for subsequent accurate conversion from packet count to absolute data volume; the current queue length in the network state data represents the historical backlog burden already existing in the edge node buffer, and the available bandwidth represents the estimated data release capacity of the physical link in the future time period. These two, as physical environmental constraints, jointly determine the initial base for congestion evolution within the node; the maximum tolerable latency in the service level agreement parameters establishes the insurmountable service quality (QoS) violation red line for various services, and is the core benchmark for subsequent calculation of the remaining latency budget and quantification of scheduling urgency.

[0036] In some implementations, edge nodes, acting as front-end processing nodes connecting terminal-side service requests with network-side processing resources, can concurrently handle multiple discrete service flows from different terminal devices, industrial field devices, or intelligent service units. For example, an edge node (such as a multi-access edge computing (MEC) server, a 5G base station, or an industrial edge gateway) serves as a core hub connecting wide-area terminal devices and the core network, and can concurrently handle a large number of discrete service flows from different industrial sensors, autonomous vehicles, or intelligent terminals.

[0037] To enable forward-looking scheduling, the system first establishes a basic state awareness mechanism for multiple service flows. Specifically, edge nodes can continuously extract the arrival timestamps and packet lengths of data packets from each service flow through traffic acquisition modules located at network interfaces, thereby forming arrival time sequences and packet length sequences that reflect the arrival behavior of discrete service flows.

[0038] Simultaneously, edge nodes also acquire the operational status of their internal network and resources, as well as service constraint parameters. For example, they can read the backlog length of each service processing queue from the queue management module, obtain available bandwidth from the link state management module, and parse service level agreement parameters from service request description information or pre-configured policy tables, including at least latency constraint information such as maximum tolerable latency. Thus, the system can simultaneously obtain the characteristics of the service flow itself, the current available resource status of the edge nodes, and service constraints at a unified time, providing a complete data input foundation for subsequent arrival behavior analysis, future queuing trend estimation, and scheduling decision generation.

[0039] In step S120, the arrival time series in the flow feature data of each discrete service flow is decomposed into multi-scale dependency to separate the arrival sequence components at different time scales.

[0040] It should be noted that discrete business flows typically exhibit significant non-stationarity in actual operation. Their arrival process may include local disturbances with sudden increases in a short period of time, as well as periodic fluctuations or slow trend changes over a longer time scale. If the original arrival time series is directly predicted as a whole, the change patterns at different scales are easily superimposed, thereby reducing the ability of subsequent prediction stages to identify key change features. In the steps of this embodiment, the edge nodes first perform multi-scale dependency decomposition on the original arrival time series, breaking down the complex and mixed arrival behavior into multiple arrival sequence components with different time scale characteristics.

[0041] During implementation, edge nodes can utilize time-domain or frequency-domain decomposition mechanisms to perform scale separation on the original arrival time series, obtaining components to characterize short-term sudden changes, components to characterize phased fluctuations, and components to characterize the overall slow trend of change. Thus, the system does not directly deal with the mixed original traffic flow arrival sequence, but rather obtains multiple component sequences with relatively clear characteristics. This facilitates subsequent prediction processing for different change patterns, reduces the analysis difficulty of complex discrete traffic sequences, and improves the targeting and stability of subsequent future arrival behavior estimation.

[0042] In step S130, based on the separated arrival sequence components and the group length sequence, the predicted arrival volume of each discrete service flow within a future preset time window is predicted, and combined with the current queue length and available bandwidth, the predicted queue length of each discrete service flow within a future preset time window is estimated.

[0043] Here, after obtaining the arrival sequence components at different time scales, the edge nodes further perform a forward-looking estimate of the service flow evolution within a preset future time window. Specifically, based on the change characteristics reflected by each component sequence, the system predicts the arrival intensity within the future window and converts the prediction results into the expected data arrival volume entering the edge nodes within the future time window by combining the group length sequence. Through this processing, the system can obtain future service input information that is more meaningful than simply the current instantaneous traffic.

[0044] Furthermore, predicting arrival volume alone is insufficient to reflect whether queuing congestion will occur in the future. Therefore, the edge node also combines the current queue length and available bandwidth to estimate the service processing and release capacity within the future time window, and thereby obtains the predicted queue length of the corresponding service flow within the future window. In other words, in this embodiment, the steps not only consider how many new services will arrive in the future, but also the currently backlogged service load and the processing capacity that the edge node can provide in the short term. As a result, the system can identify potential backlog trends before the queue grows significantly, shifting the scheduling trigger point from post-event response to pre-event prediction.

[0045] In step S140, for each discrete service flow, the predicted remaining delay budget of the discrete service flow is calculated based on the predicted queue length and the maximum tolerable delay in the service level agreement parameters, and the predicted remaining delay budget is mapped to a digital trigger probability, which is used to characterize the scheduling urgency of each discrete service flow.

[0046] Here, the edge node further converts the predicted queuing length obtained in step S130 into an evaluation result directly related to service latency constraints. Specifically, based on the service capabilities that the edge node can provide for the corresponding service flow, the system converts the predicted queuing length into the potential future queuing time and compares it with the maximum tolerable latency specified in the service level agreement parameters to obtain the predicted remaining latency budget for the service flow within the future window. This budget value characterizes how much buffer space the service flow has left before reaching the latency constraint boundary; the smaller the budget, the closer the service flow is to the risk state of latency default.

[0047] To avoid the problems of abrupt scheduling switches and frequent resource preemption that are easily caused by traditional fixed threshold determination methods, this embodiment further maps the predicted remaining latency budget to a continuously changing digital trigger probability. This probability value can be used to smoothly characterize the scheduling urgency, enabling the system to gradually increase the intensity of scheduling intervention as the remaining latency budget decreases, rather than suddenly triggering scheduling actions only after exceeding a certain threshold. Thus, edge nodes transform potential future latency risks into continuous digital quantities that can directly participate in scheduling decision calculations, thereby improving the continuity and controllability of the scheduling triggering process.

[0048] In step S150, the flow importance index of each discrete service flow is obtained, and the scheduling weight of each discrete service flow is determined by combining the digital trigger probability and the flow importance index corresponding to each discrete service flow. The flow importance index is used to characterize the service importance of the corresponding discrete service flow.

[0049] It should be noted that in multi-service concurrent scenarios, whether a service flow should be prioritized for scheduling depends not only on its latency risk but also on the importance of the service flow itself. In the steps of this embodiment, the edge node further obtains the flow importance index of each discrete service flow. This flow importance index can be comprehensively determined based on information such as the service level, service priority, default cost, and historical service quality status of the service flow, and is used to characterize the relative importance of different service flows to the overall service operation.

[0050] Based on this, the edge nodes jointly calculate the scheduling weights of each discrete service flow by combining the numerical trigger probabilities used to characterize scheduling urgency with the flow importance index used to characterize service importance. Specifically, this scheduling weight does not simply reflect whether a service flow is "urgent" or "important," but rather it is a ranking criterion that integrates both factors. Therefore, when multiple services simultaneously compete for limited resources, the system can simultaneously ensure both latency constraints and service priority, making the subsequent resource allocation results more aligned with the actual service needs of the edge network.

[0051] In step S160, based on the scheduling weights corresponding to each discrete service flow, the service order of the scheduling queues within the edge nodes is dynamically adjusted and corresponding local resources are allocated to trigger scheduling operations for multiple discrete service flows in advance.

[0052] As an execution step of the aforementioned analysis and calculation results, the edge node in this embodiment dynamically adjusts the service order of service flows in its internal scheduling queue based on the scheduling weights corresponding to each service flow. Specifically, service flows with higher scheduling weights will be given priority in service order, thus gaining priority in queue service opportunities within a limited processing window. This adjustment is no longer limited to simple processing based on arrival order, but rather dynamically reconstructs the processing order within the edge node based on the results of the aforementioned latency risk assessment and service importance assessment.

[0053] After adjusting the queue service order, edge nodes can further allocate corresponding local resources to the relevant service flows. These local resources may include computing resources, cache resources, and link-side resources within the edge node that can be used for service processing and forwarding. Since the scheduling weights in this scheme are based on predictions of future queuing trends and latency budgets, the resource allocation actions in this embodiment have significant foresight. That is, before the service flow actually experiences significant latency defaults, the service order reordering and resource configuration are completed in advance, enabling the edge node to truly implement the aforementioned predictive analysis results into executable scheduling control actions, thereby improving the service guarantee capability for discrete service flows under latency constraints.

[0054] Furthermore, it is worth noting that, addressing the technical shortcomings of current related technologies where complex prediction models struggle to balance real-time performance and application costs in edge computing environments, the predictive digital triggering method provided in this application possesses lightweight and localized characteristics highly suitable for edge node deployment. Specifically, at the algorithm architecture level, this solution adopts a "divide and conquer" design logic, effectively reducing the dimensionality of non-stationary and highly bursty complex traffic characteristics through multi-scale signal decomposition. This allows for high-precision trend prediction based solely on a time-series model with extremely low computational overhead and basic mathematical mapping. Thus, it avoids the dependence of large-scale end-to-end deep learning models on massive computing power, enabling edge nodes with relatively limited computing resources to independently undertake the entire inference and computation process, significantly reducing practical application costs.

[0055] On the other hand, compared to the conventional integrated architecture that transmits underlying traffic data back to a central node or control server for centralized processing, this solution directly pushes the core prediction and scheduling decision-making logic down to the edge of data access, bringing significant system-level advantages. Relying on a central node for traffic prediction and issuing scheduling commands inevitably introduces long network transmission latency (i.e., backhaul latency), causing scheduling actions to suffer fatal delays when facing highly sudden and discrete business flows. This solution, however, leverages the lightweight deployment of edge nodes to achieve an extremely short physical closed-loop control of "local data collection - local trend prediction - local trigger scheduling." This not only significantly reduces the bandwidth overhead of continuously reporting massive traffic status data to the upper-level network but also fundamentally eliminates the erosion of scheduling real-time performance by long-distance transmission latency. It ensures that scheduling intervention actions can be accurately triggered within a very narrow time window (i.e., before actual business congestion occurs), thereby maximizing the delivery of service quality commitments under stringent latency constraints with extremely low infrastructure overhead.

[0056] Regarding the implementation details of separating arrival sequence components at different time scales in step S120, in some examples of embodiments of this application, firstly, the arrival time sequence of each discrete service stream is discretized according to a preset discrete time sampling step size to obtain an arrival intensity sequence with equal time intervals.

[0057] It should be noted that because the raw service flow data captured at the network layer is usually asynchronous and irregular timestamp events, frequency domain or time-frequency domain analysis tools cannot be directly used. The data obtained by edge nodes from packet capture at the network interface is... The original arrival time series of discrete traffic flows is usually represented as a set of discrete points. ,in Indicates the first The system calculates the arrival timestamps of each data packet. To convert this into a standard time-series signal, a preset discrete-time sampling step size is introduced. Resampling and aggregation are performed using a sliding time window statistical method. For example, the first... Discrete traffic flow in the first Arrival intensity sequence at discrete time windows It can be calculated using the following formula:

[0058] Equation (1)

[0059] In the formula, This is an indicator function; it takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. This serves as the index for the discretized, equally spaced time series. Thus, the arrival events of bursty, asynchronous raw discrete data packets are transformed into a digital signal sequence with a fixed sampling frequency and a unit of measurement in packets per second (e.g., packets / second). This sequence can be characterized as arrival intensity in subsequent macroscopic queuing analysis in the continuous time domain. .

[0060] Then, a preset wavelet basis function is selected to perform multi-level discrete wavelet transform on the arriving intensity sequence to extract the low-frequency approximation coefficients of the highest decomposition level and the high-frequency detail coefficients of each decomposition level.

[0061] In practical implementation, considering the highly non-stationary characteristics of discrete traffic flows, traditional Fourier transforms cannot provide localized positioning capabilities in the time domain. Therefore, this embodiment preferably employs the Discrete Wavelet Transform (DWT), which possesses excellent time-frequency localization characteristics. For example, a preset wavelet basis function pair is selected for the arrival intensity sequence. conduct Layer discrete wavelet transform to extract the first layer The low-frequency approximation coefficients of the first layer and the first layer to the second layer The high-frequency detail coefficients of the layer, among which... It is a positive integer greater than 1.

[0062] For example, the system can select orthogonal wavelets such as Daubechies or Haar as preset wavelet basis functions, and use the Mallat pyramid algorithm to process the arrival intensity sequence. conduct Layered discrete wavelet transform. In each layer of iterative decomposition, the signal passes through orthogonal low-pass and high-pass filters, outputting approximate coefficients representing gradual signal changes and detail coefficients representing dramatic signal changes, respectively. After layer iteration, the system can extract the first... Low-frequency approximation coefficients of the layer (Contains the global basic outline of the signal), and layers 1 to 1. A series of high-frequency detail coefficients of the layer ( (This captures the details of local fluctuations in each frequency band). This allows the system to perform precise decoupling of the chaotic raw traffic signal in an orthogonal multi-resolution space, much like prism dispersion.

[0063] Furthermore, according to the preset hierarchical division rules, the high-frequency detail coefficients of the shallow decomposition level are reconstructed into short-term burst components to reflect the burst arrival characteristics at a shorter time scale, the high-frequency detail coefficients of the deep decomposition level are reconstructed into periodic components to characterize the periodic arrival change characteristics, and the low-frequency approximation coefficients of the highest decomposition level are reconstructed into long-term trend components to reflect the arrival change trend at a longer time scale, so that the short-term burst components, periodic components, and long-term trend components together constitute the arrival sequence components at different time scales.

[0064] It should be noted that, since the extracted wavelet coefficients are still in the wavelet domain, the system needs to project them back to the original time domain using inverse wavelet transform (IDWT) to restore their actual physical meaning. For example, the system sets a hierarchical division threshold based on prior knowledge of the frequency domain distribution of discrete service flows. ,in The system utilizes inverse wavelet transform to transform the first layer to the second layer. High-frequency detail coefficients of the layer (representing the highest frequency band) Reconstructed into short-term burst components ; will the first Layer to the first High-frequency detail coefficients of the layer (representing the mid-frequency band) Reconstructed into periodic components ; and will the Low-frequency approximation coefficients of the layer Reconstructed into long-term trend components .

[0065] Furthermore, due to the perfect energy preservation and linear reconstruction properties of the discrete wavelet transform, the reconstructed time-scale components are identical to the original arrival intensity sequence. It satisfies the following addition composition logic:

[0066] Equation (2)

[0067] In some implementations, this formula explicitly reveals the physical equivalence of signal reconstruction. Short-term burst components. To reflect the sudden arrival characteristics at shorter time scales, transient flow spikes at the microsecond or millisecond level were extracted; periodic components Used to characterize periodic arrival variations, it extracts rhythmic oscillations influenced by business tidal effects; long-term trend component Used to reflect arrival trends over longer timescales, it outlines the slow drift trajectory of network infrastructure load; short-term burst components Periodic components and long-term trend components Together they constitute arrival sequence components at different time scales, and are superimposed with a lossless correspondence arrival intensity sequence. .

[0068] This application's multi-scale dependency decomposition and reconstruction solves the problem of high-dimensional non-stationary network traffic being difficult to accurately fit with a single time-series model. Utilizing multi-resolution time-frequency analysis, it successfully extracts the inherent instantaneous burst characteristics, periodic fluctuation characteristics, and long-term trend characteristics of discrete service flows. By employing an orthogonal and lossless signal decoupling method, it not only preserves the dynamic evolution information of the original service flow at different time granularities to the maximum extent, but also enables subsequent systems to flexibly match the most suitable heterogeneous prediction algorithm for single pure components with different physical laws. This reduces the risk of feature dimension explosion and underfitting faced by single prediction models, providing a solid data support foundation for edge networks to achieve high-precision forward-looking perception and scheduling under stringent latency constraints.

[0069] Regarding the implementation details of estimating the predicted queuing length of each discrete service flow in step S130, in some examples of embodiments of this application, firstly, for a future preset time window, the short-term burst component, periodic component and long-term trend component of each discrete service flow are predicted using corresponding heterogeneous time series prediction models to generate a first predicted arrival intensity, a second predicted arrival intensity and a third predicted arrival intensity, respectively. The heterogeneous time series prediction models include an autoregressive prediction model, a periodic stationary prediction model and an exponentially weighted moving average model.

[0070] In practical implementation, since the aforementioned steps have decoupled the complex arrival sequence into three components with distinct physical characteristics, using a single model for prediction is highly prone to feature underfitting. Therefore, this embodiment adopts a heterogeneous prediction architecture of "dedicated model for dedicated purpose". Let the current time be... The duration of the future preset time window is For future preset time windows , respectively for the first Short-term burst components of discrete traffic flows Periodic components and long-term trend components Make predictions.

[0071] In some implementations, an autoregressive prediction model (AR) is used to predict short-term burst components (high-frequency oscillations). Leveraging its high sensitivity to recent historical data, the AR model captures transient changes at the millisecond / second level to generate a first predicted arrival intensity. For the periodic component, the system employs a periodic stationary prediction model (e.g., based on Fourier series expansion or seasonal autoregressive model) to accurately reproduce the tidal patterns of the business and generate a second predicted arrival intensity. For the low-frequency long-term trend component, the system employs an exponentially weighted moving average (EWMA) model. By assigning exponentially decaying weights to historical data, it effectively filters out high-frequency noise and smoothly tracks the evolution and drift of macroscopic flow, thereby generating a third predicted arrival intensity. Therefore, by matching the optimal heterogeneous mathematical model to signals in different frequency bands, the feature capture capability of the prediction engine is maximized, and the overall generalization error of a single prediction model is reduced.

[0072] Then, the first predicted arrival intensity, the second predicted arrival intensity, and the third predicted arrival intensity are combined in the time domain within a future preset time window, and the combined result is multiplied by the average group length determined based on the group length sequence to calculate the predicted arrival amount within the future preset time window.

[0073] In some implementations, the first step is combined with the step determined based on the group length sequence. Average packet length of discrete traffic flows Calculate the predicted arrival amount within a future preset time window. :

[0074] Equation (3)

[0075] In the formula, , and Representing time respectively The first predicted arrival intensity, the second predicted arrival intensity, and the third predicted arrival intensity. This is the time variable for integration.

[0076] In the above formula, the future time is recovered by adding the three predicted arrival intensities (e.g., in packets per second). The total predicted arrival rate, by using this total arrival rate in future time windows The system performs time-domain definite integral calculations (i.e., solves for the area enclosed by the arrival rate curve and the time axis) to accurately deduce the total number of data packets expected to arrive within that time period. Subsequently, the system multiplies this total number of data packets by the statistically obtained average packet length. (For example, in "bytes / packet" or "bits / packet"), thus completing the conversion from "number of packets" to "absolute data volume".

[0077] Furthermore, based on the fluid approximation queuing model, the current queue length, measured in terms of data volume, is added to the predicted arrival volume, and the cumulative service data volume estimated based on available bandwidth within a future preset time window is subtracted. The difference in data volume after subtraction is then multiplied by zero to obtain the larger value, and the result is used as the predicted queue length of the corresponding discrete service flow at the end of the future preset time window.

[0078] In some implementations, a fluid approximation queuing model is used, combined with the current queue length measured in data volume. And available bandwidth, and estimated predicted queue length at the end of the preset time window. This serves as the predicted queuing length for the corresponding discrete business flow within a future preset time window:

[0079] Equation (4)

[0080] In the formula, This is the instantaneous service rate estimated based on available bandwidth.

[0081] To transform flow prediction into a realistic congestion assessment within edge nodes, this step introduces a continuous-state fluid approximation queuing model. Equation (4) can be physically compared to a "water tank model": where... This represents the amount of data already accumulated in the buffer at the current moment (initial water level). This represents the expected influx of new data within a future time window (estimated inflow); while the integral term... This represents the total amount of data that an edge node can process and forward within a future time window based on available link bandwidth (estimated outflow). The system adds the "estimated inflow" to the "initial backlog" and subtracts the "estimated outflow" to calculate the future net backlog. Thus, it not only considers the impact of external traffic input but also incorporates the dynamic changes in the edge node's own service capabilities in a closed loop, achieving a high-fidelity physical simulation estimate of the future actual queuing backlog state.

[0082] Furthermore, through the outer nesting of equation (4) The function implements the rigid "non-negative queue constraint" in queuing theory—that is, when the network service capacity is extremely abundant (outflow is greater than the sum of inflow and backlog), the physical queue length cannot be negative, but will only be cleared to zero.

[0083] The heterogeneous time-series prediction and fluid queuing model provided in this application provides a cascaded extrapolation architecture from "multi-dimensional signal prediction" to "physical congestion evolution." The core advantages of this implementation are: firstly, it accurately matches corresponding autoregressive, periodically stationary, or exponentially smoothing models to a single flow component after multi-scale decoupling, completely breaking the generalization bottleneck of a single prediction model when fitting complex flow rates and achieving high-precision arrival rate prediction; secondly, it uses fluid approximate queuing theory and rigorous integral dimension transformation to upgrade flow scale prediction (how much data will arrive) to a more critical physical congestion prediction (how much data will be blocked). Thus, a dynamic non-negative queuing extrapolation mechanism considering "historical backlog + future bursts - future service capacity" is established, enabling edge nodes to anticipate buffer overflow risks before they occur, providing an absolutely reliable and rigorously physically supported judgment basis for subsequent pre-emptive anti-jitter scheduling.

[0084] Figure 2 A flowchart illustrating an example of calculating the probability of digital triggering in a method according to an embodiment of this application is shown.

[0085] like Figure 2 As shown, in step S210, the predicted queue length is divided by the average estimated service rate to calculate the predicted queue delay.

[0086] Here, the average estimated service rate is the estimated service rate that the edge node allocates to the corresponding discrete service flow within a future preset time window based on available bandwidth. In some implementations, the average estimated service rate is based on the edge node's allocation to the corresponding discrete service flow within a future preset time window based on available bandwidth. Average estimated service rate of discrete traffic flows And combined with predicted queue length Calculate the first Predicted queuing delay for discrete traffic flows :

[0087] Equation (5)

[0088] Specifically, although step S130 has already yielded the predicted queue length However, this metric belongs to the spatial dimension of data (e.g., bits or bytes), and cannot be directly compared with the time dimension (e.g., milliseconds) specified in the Service Level Agreement (SLA). Therefore, edge nodes need to work with the underlying network controller to assess their ability to provide services to the next node within a predetermined time window. Average estimated service rate of traffic flow allocation (e.g., bits per second). Through the division operation of Equation (5), the system accurately converts the data backlog in the spatial dimension into the queuing time in the time dimension, and also realizes the unification of physical dimensions between heterogeneous data.

[0089] In step S220, the predicted queuing delay and the dynamic safety margin in terms of time are successively subtracted from the maximum tolerable delay set in the service level agreement parameters to calculate the predicted remaining delay budget. The dynamic safety margin is used to compensate for prediction errors.

[0090] For example, calculate the first Predicting Remaining Delay Budget for Discrete Traffic Flows :

[0091] Equation (6)

[0092] In the formula, For the Service Level Agreement parameters regarding the first The maximum tolerable delay set for discrete traffic flows. This is a dynamic safety margin, expressed in terms of time, used to compensate for prediction errors.

[0093] In real-world edge network environments, due to fluctuations in wireless channels or transient preemption of underlying hardware resources, residuals inevitably exist between traffic prediction and service rate estimation. If the maximum tolerable delay is used directly... Subtract predicted queuing delay As a budget, in the event of unexpected congestion, the business flow will instantly exceed the default threshold. Therefore, this embodiment introduces a dynamic safety margin aligned with the time dimension. As a buffer zone. In Equation (6), the actual "safe activity space" of the business flow is strictly compressed, and there must still be a surplus after deducting queuing time and prediction fault tolerance costs. A relatively conservative and fault-tolerant delay budget calculation mechanism is constructed, which significantly enhances the robustness of the system in dealing with extremely uncertain network environments.

[0094] In step S230, the predicted remaining delay budget is used as the independent variable, and the trigger sensitivity coefficient used to control the steepness of the response is combined with the mapping calculation through a continuously smooth nonlinear probability activation function to obtain the digital trigger probability.

[0095] In some implementations, a continuously smooth nonlinear probabilistic activation function is used to predict the remaining time delay budget. Mapped to numerical trigger probability :

[0096] Equation (7)

[0097] In the formula, A trigger sensitivity coefficient greater than zero is used to control the steepness of the response of the digital trigger probability to changes in the predicted remaining delay budget. It is a natural exponential function. When predicting the remaining time delay budget... The probability of a number being triggered when it is less than zero and continues to decrease. The nonlinearity increases and approaches 1 to characterize the increased urgency of scheduling.

[0098] Here, when the predicted remaining delay budget is less than zero and continues to decrease, the digital trigger probability increases nonlinearly and approaches the upper limit of the value to characterize the increased scheduling urgency; among them, the continuously smooth nonlinear probability activation function is used to transform the hard-decision delay threshold into a flexible probability measure to suppress the scheduling jitter caused by frequent resource preemption by edge nodes when dealing with discrete traffic bursts.

[0099] Combining equation (7), it constructs a nonlinear mapping mechanism similar to Sigmoid. In traditional hard-decision mechanisms, it is usually set that when... Time-triggered scheduling, when If not triggered, this "either 0 or 1" step control can easily lead to a large number of business flows simultaneously preempting each other near the budget threshold, causing severe oscillations in the scheduling queue (i.e., the ping-pong effect). In the nonlinear mapping formula of this implementation, when predicting the remaining delay budget... When the number is sufficient (a large positive number), the exponent term in the denominator is amplified exponentially, increasing the probability of the number being triggered. Approaching 0; with Gradually decrease and approach zero. It begins to rise smoothly; when When the exponent term is less than zero and continues to decrease, The probability of digital triggering decreases rapidly and approaches zero. The nonlinear increase in value infinitely approaches the upper limit of 1, representing an extremely high level of scheduling urgency. Therefore, a continuously smooth nonlinear probabilistic activation function is employed to transform the hard-decision delay threshold into a flexible, continuous probability measure. This soft-decision mechanism not only eliminates the abrupt blind spots of the trigger boundary but also allows edge nodes to calmly and gradually implement scheduling interventions based on the gradual trend of probability, effectively suppressing global scheduling jitter caused by frequent resource preemption when dealing with bursts of discrete traffic.

[0100] This application successfully constructs a rigorous mapping chain from "physical space congestion" to "time-dimensional budget" and then to "system decision probability." It introduces a latency budget calculation model with dynamic fault-tolerant boundaries and further couples it with a nonlinear probabilistic activation function, transforming the absolute remaining time (milliseconds) into a relative urgency measure (a continuous probability between 0 and 1). This gives the system's scheduling decisions excellent smoothness and flexibility. As a result, edge nodes can not only accurately quantify the latency default risk of various business flows at different congestion stages, but also solve the resource switching oscillation problem under high-concurrency burst scenarios, providing crucial decision-level support for achieving highly stable, low-jitter deterministic edge network scheduling.

[0101] Regarding the implementation details of calculating the scheduling weight in step S150, in some examples of embodiments of this application, firstly, the service level agreement parameters are parsed to extract the basic service priority weight and default cost coefficient configured for the corresponding discrete service flow. For example, the service priority basic weight and default cost coefficient configured for the first discrete service flow are extracted. Basic weight of service priority in discrete service flow configuration and default cost coefficient .

[0102] It should be noted that in a multi-tenant edge network environment, different discrete service flows often carry tasks with vastly different commercial value and application scenarios (for example, the value of industrial control commands is far greater than that of ordinary background log feedback). Therefore, before allocating resources, the system must establish the basic commercial benchmarks for each service flow. Edge nodes extract the statically configured basic weights of service priorities by deeply analyzing the Service Level Agreement (SLA) context negotiated by the service flows when establishing connections. (Used to reflect basic tenant levels such as Gold and Silver) and default cost coefficient (Used to quantify the economic compensation or system performance penalty that the network operator needs to bear if the latency of the service flow exceeds the limit). Thus, the abstract business contract and service penalty cost are quantified into numerical variables that can directly participate in the underlying network scheduling, injecting a business value-driven dimension into pure technical scheduling.

[0103] Then, a sliding observation window is established with the current time as the baseline, and the historical latency default rate of the corresponding discrete service flows within the sliding observation window is statistically analyzed. The default cost coefficient is weighted using a preset default cost assessment coefficient, and the historical latency default rate is weighted using a preset service quality compensation coefficient. The weighted default cost coefficient, the historical latency default rate, and the basic weight of the service priority are then linearly summed to calculate the flow importance index. Here, the flow importance index is used to comprehensively characterize the degree of service importance determined by preset priority, default cost, and historical service quality status.

[0104] In some implementations, to accurately capture recent service quality fluctuations in business flows, edge nodes are established based on the current time. Based on, with a time span of A sliding observation window. For example, historical delay default rate. It can be dynamically calculated using the following formula:

[0105] Equation (8)

[0106] In the formula, Indicates the first The set of historical data packets that have been scheduled within the current sliding observation window for discrete traffic flows; This indicates the total number of data packets within the set; For historical data packets The actual latency experienced within the edge node; The characteristic function is defined when the actual delay of the packet is greater than the maximum tolerable delay. The output is 1 if the condition is met, and 0 otherwise.

[0107] Base weight of business priority Default cost coefficient and historical delay default rate Perform linear weighted fusion to calculate the flow importance index. :

[0108] Equation (9)

[0109] In the formula, The default cost assessment coefficient is the preset value. The preset service quality compensation coefficient; flow importance index. It is used to comprehensively characterize the importance of a business, which is determined by preset priority, cost of default, and historical service quality status.

[0110] It should be noted that a multi-dimensional linear evaluation system for the Flow Criticality Index (FCI) is constructed using Equation (9). The first term of Equation (9) establishes the static hierarchy of the business; the second term of Equation (9)... Amplify or reduce the system costs caused by default; the third term of equation (9) introduces the extremely important "historical service quality compensation mechanism"—when a certain business flow has recently encountered severe congestion leading to When it is too high, a compensation coefficient will be used. The overall importance of these flows is dynamically increased. This breaks away from the drawbacks of the traditional "static priority" approach in networks, allowing "weak, high-value flows" that have both high default costs and recent service quality degradation to receive dynamic weighting from the system, thus effectively preventing scheduling starvation.

[0111] Furthermore, after applying a power-law nonlinear adjustment to the digital trigger probability using a preset trigger adjustment index, the adjusted digital trigger probability is multiplied by the flow importance index to calculate the scheduling weight. Here, the scheduling weight is used to enhance the weight differentiation between discrete service flows with high scheduling urgency and those with low scheduling urgency, so as to jointly adjust the resource allocation priority according to scheduling urgency and service importance.

[0112] In some implementations, the digital trigger probability is... After power-law adjustment, the importance index of the flow Multiplication, calculation for the first Scheduling weights of discrete traffic flows , to serve as the scheduling weights for the corresponding discrete service flows:

[0113] Equation (10)

[0114] In the formula, To trigger the adjustment index, and Scheduling weight This is used to enhance the weight differentiation between discrete service flows with high scheduling urgency and discrete service flows with low scheduling urgency, so as to jointly adjust the priority of resource allocation according to scheduling urgency and service importance.

[0115] The aforementioned steps have normalized the scheduling urgency of each service flow into a numerical trigger probability between 0 and 1. In this embodiment, due to the base number Apply a trigger adjustment index greater than or equal to 1 to it. Performing exponentiation, mathematically speaking, acts as a form of "non-linear contrast enhancement." Specifically, for business flows with lower urgency (i.e.... Smaller values, such as 0.2, will be severely suppressed after exponentiation (e.g., ...). ); and for extremely urgent business flows that are on the verge of timeout (i.e. Approaching 1), its value is almost unaffected by exponential operations (e.g. This highly distinctive urgency indicator was then compared with the flow importance index, which represents business value. Multiply. Thus, the overall scheduling weight is only applied when the business flow is not only important but also genuinely time-sensitive. This will give it an overwhelming advantage and prevent high-priority businesses from ineffectively monopolizing underlying resources when there is no urgency.

[0116] By evaluating and calculating the comprehensive scheduling weights provided in this application, a joint scheduling decision-making mechanism driven by both "latency urgency (time constraints)" and "flow importance (business and service constraints)" has been successfully constructed. It not only utilizes a linear fusion formula incorporating historical default compensation to deeply extract multi-dimensional business attributes, but also employs a multiplicative coupling algorithm based on power-law nonlinear adjustment. This achieves "nonlinear amplification and low-priority suppression" of priority measurement, completely overcoming the dual industry pain points of the traditional earliest deadline first (EDF) algorithm ("easily leading to the discarding of high-value services") and the traditional strict priority (SP) algorithm ("easily causing low-priority services to be indefinitely starved"). By establishing comprehensive scheduling weights with high business differentiation, the system can maximize the preservation of SLA compliance for high-value services and minimize the global default cost under extremely congested conditions at the edge.

[0117] Regarding the implementation details of dynamically allocating local resources to edge nodes in step S160, in some examples of embodiments of this application, firstly, according to the scheduling weights corresponding to each discrete service flow... The multiple discrete service flows in the scheduling queue are sorted in descending order to update the resource allocation priority of each discrete service flow.

[0118] In practice, the resource management module within the edge nodes maintains a global scheduling waiting queue. During each scheduling decision cycle, the system no longer processes business flows solely based on their arrival order, but rather according to the comprehensive scheduling weight calculated in the preceding steps. The system reorders multiple discrete service flows currently in a waiting state. A higher scheduling weight indicates a higher scheduling priority for the corresponding service flow based on a comprehensive assessment of latency urgency and service importance; therefore, its position in the scheduling queue will be moved forward accordingly. Through this process, edge nodes can dynamically adjust the processing order of multiple service flows at the queue level before the actual resource allocation action occurs, thus providing a clear priority basis for subsequent local resource allocation.

[0119] Then, the local resource state within the edge node is analyzed, and a normalized energy consumption cost function is established for each candidate resource unit. This normalized energy consumption cost function is used to quantify the running energy consumption when assigning tasks to the corresponding candidate resource units for execution, as well as the state switching cost from a dormant state to an active state. For example, a normalized energy consumption cost function can be established for the first... Normalized energy consumption cost function of candidate resource units ,in, This is an index for candidate resource units. Used to quantify the assignment of tasks to the first The energy consumption of each candidate resource unit during execution and the state switching cost of switching the candidate resource unit from a dormant state to an active state.

[0120] At the physical implementation level, the local processing resources of edge nodes can be divided into multiple candidate resource units, such as different processing cores, dedicated signal processing units, or heterogeneous hardware acceleration units. Since the operating states of different candidate resource units may differ at any given time—for example, some resource units may be active while others are in a low-power sleep state—the system needs to uniformly quantify the available state and energy consumption characteristics of each candidate resource unit before performing resource allocation. In some implementations, a normalized energy consumption cost function is used. It can be represented as:

[0121] Equation (11)

[0122] In the formula, For the first The estimated basic operating energy consumption required for each candidate resource unit to perform the current task; For the state indication function, when the first The value is 1 when the candidate resource unit is currently in a dormant state and 0 when it is in an active state. For the first The state switching cost of a candidate resource unit switching from a dormant state to an active state; This is a preset upper bound for normalized energy consumption, used to uniformly map the energy consumption cost corresponding to different candidate resource units to... Within the dimensionless interval. Here, the comprehensive energy consumption cost of candidate resource units includes both the basic operating energy consumption required to execute the task itself and the additional energy consumption generated when they are woken up from their original dormant state. After unified normalization, it can provide comparable energy consumption quantification input for subsequent joint scheduling utility calculation.

[0123] Subsequently, when allocating local resources to each discrete service flow according to resource allocation priority, the scheduling weight of the corresponding discrete service flow is subtracted from the energy consumption penalty term obtained by multiplying the energy consumption penalty adjustment factor and the normalized energy consumption cost function, and the joint scheduling utility value corresponding to each candidate resource unit is calculated. For example, when allocating local resources to each discrete service flow according to resource allocation priority, the joint scheduling utility value corresponding to each candidate resource unit is calculated by subtracting the energy consumption penalty term obtained by multiplying the energy consumption penalty adjustment factor and the normalized energy consumption cost function from the scheduling weight of the corresponding discrete service flow. Discrete traffic flow allocation to the first Joint scheduling utility value of candidate resource units :

[0124] Equation (12)

[0125] In the formula, This is a preset energy consumption penalty adjustment factor used to adjust the system's sensitivity to energy consumption costs. In equation (12), the first term... This indicates that resources will be allocated preferentially to the first... The scheduling benefits brought by discrete traffic flows have comprehensively reflected the latency urgency and importance of the corresponding traffic flows; the second item This indicates that the task will be mounted to the first... The normalized energy consumption cost required for each candidate resource unit. Therefore, the joint scheduling utility value... This reflects a comprehensive trade-off between "service benefits" and "energy costs." Through this process, edge nodes no longer make decisions solely based on business priorities or minimum energy consumption when allocating local resources, but rather consider both factors within a unified utility framework.

[0126] Furthermore, under the condition of satisfying the delay constraints of the corresponding discrete service flows, the target candidate resource unit that maximizes the joint scheduling utility value is selected as the corresponding local resource to trigger the scheduling operation in advance. For example, under the condition of satisfying the... Under the delay constraint corresponding to the maximum tolerable delay of discrete traffic flows, select the value that maximizes the joint scheduling utility. The largest target candidate resource unit is used as the corresponding local resource to trigger the target in advance. Scheduling operations for discrete traffic flows.

[0127] In some implementations, after calculating the joint scheduling utility value, the system can also combine the available processing capacity of the candidate resource units to determine the optimal scheduling value. Discrete traffic flows are allocated to the first The predicted remaining latency budget is updated after each candidate resource unit. For example, the updated predicted remaining latency budget can be recorded as... It can be represented as:

[0128] Equation (13)

[0129] In the formula, For the first The maximum tolerable latency for discrete traffic flows; To make the first Discrete traffic flow allocation to the first The predicted queuing and processing latency corresponding to each candidate resource unit; This provides a dynamic safety margin. Therefore, the system can satisfy... The candidate resource units are considered as a set of feasible resource units that satisfy the delay constraint. Within this set of feasible resource units, the system further selects those that maximize the joint scheduling utility value. The largest candidate resource unit, as the first The system targets local resources for discrete service flows. In other words, the system first eliminates resource allocation schemes that may cause service flow delay violations, and then selects the resource unit with the highest joint scheduling utility from the remaining feasible schemes to ensure that the resource allocation result takes into account both delay compliance and energy consumption rationality.

[0130] In the specific execution process, when the scheduling weight of a certain discrete service flow... When the scheduling weight is high, the scheduling benefit item in the joint scheduling utility will account for a larger proportion. At this time, the system tends to prioritize allocating candidate resource units that can complete processing faster to this service flow to ensure that its latency constraints are met; while when the scheduling weight of a certain service flow is high... When the delay is relatively small and its remaining prediction delay budget is sufficient (i.e., there is still some margin in the delay constraint), the energy consumption penalty adjustment factor is used. The negative penalty effect of normalized energy consumption cost will make its impact on joint scheduling utility more pronounced. In this case, edge nodes prioritize assigning corresponding processing tasks to active candidate resource units or postpone waking up dormant candidate resource units to minimize the global energy consumption cost of edge nodes within latency constraints. In this way, edge nodes can establish a dynamic matching relationship between the scheduling requirements of different service flows and the energy consumption characteristics of different resource units, thereby making the local resource allocation process more adaptable.

[0131] The local resource dynamic allocation mechanism based on joint scheduling utility in this embodiment integrates scheduling weights and the normalized energy consumption cost of candidate resource units into the same utility evaluation framework, and selects target resource units while meeting service latency constraints. Compared to resource allocation based solely on service priority or resource scheduling based solely on hardware energy consumption, this embodiment can simultaneously consider the service guarantee requirements of service flows and the actual energy consumption of underlying hardware within the edge node, thereby achieving a more coordinated resource allocation result in dynamically changing discrete service flow scenarios. Therefore, when the service flow latency risk is high or the service importance is high, the system can overwhelmingly prioritize the processing resources of the corresponding service flow; when the service flow latency margin is relatively sufficient, the system can appropriately suppress unnecessary high-energy-consuming resource wake-ups and state switching, thereby improving the resource utilization efficiency and global energy consumption control capability of the edge node under stringent latency constraints.

[0132] In some examples of embodiments of this application, after the dynamic adjustment of the service order of the scheduling queue and the allocation of the corresponding local resources are completed, a closed-loop dynamic correction operation of parameters can also be performed. Figure 3 A flowchart illustrating an example of performing parameter closed-loop dynamic correction operations in a method according to an embodiment of this application is shown.

[0133] like Figure 3 As shown, in step S310, the actual queuing delay of each discrete service flow after dynamic resource allocation is monitored in real time, and the actual operating energy consumption of the edge node is monitored simultaneously.

[0134] It should be noted that although the preceding steps have estimated the future queuing state based on multi-scale prediction and a fluid approximation queuing model, in actual edge network environments, there may still be various disturbance factors that are difficult to model completely in advance, such as the additional handover overhead caused by underlying resource scheduling, the instantaneous fluctuations in wireless link states, and the uncertainty of execution latency of heterogeneous hardware. In this embodiment, after completing resource allocation and task execution, the system further introduces a closed-loop observation mechanism to collect the actual operating results in real time, so as to obtain the objective basis required for subsequent parameter correction.

[0135] Specifically, edge nodes can obtain the first queuing time by recording the enqueue and dequeue times in the protocol stack processing path or the local queue management module. Actual queuing delay of discrete traffic flows within the current scheduling period .in, Indicates the first Discrete traffic flow at time The corresponding actual queuing delay. Simultaneously, the system can also obtain the actual operating energy consumption of edge nodes during the corresponding observation period by reading local resource monitoring interfaces, power consumption sensors, or the operating system's energy consumption statistics module, and record it as... .in, Indicates the edge node at time... This could correspond to the actual operating energy consumption within the observation window. Through this step, the system transitions from the feedforward prediction stage to the feedback correction stage based on physical measurement results, providing a practical operational basis for subsequent parameter updates.

[0136] In step S320, the delay prediction residual between the actual queuing delay and the predicted queuing delay of the corresponding discrete service flow is calculated.

[0137] In some implementations, the edge node will monitor the actual queuing delay. The predicted queuing delay obtained in the preceding steps Perform time alignment and interpolation to obtain the first... Discrete traffic flow at time Delay prediction residual :

[0138] Equation (14)

[0139] In the formula, Indicates the first Discrete traffic flow at time The time delay prediction residual; This refers to the actual queuing delay. To predict queuing delay, we quantify the deviation between the current prediction model and the actual physical operating state by differentiating the actual measured delay from the predicted delay. When this occurs, it indicates that the system tends to underestimate queuing delay; when... This indicates that the system tends to overestimate queuing delay. Therefore, it can provide a directional basis for subsequent dynamic adjustments to trigger parameters and safety margins.

[0140] In some further implementations, to avoid excessive sensitivity to parameter adjustments caused by a single instantaneous disturbance, the system can statistically analyze instances of exceeding queuing delay limits within a preset sliding window. This is done by referring to the historical data packet set established above within the current sliding observation window. Then the first The queuing delay of discrete traffic flows exceeds the maximum tolerable delay within the sliding window. Number of times It can be represented as:

[0141] Equation (15)

[0142] In the formula, Indicates the first The number of times the latency of a discrete service flow exceeds the limit within the current sliding window; This represents the set of historical data packets that have been scheduled within the current sliding observation window; For historical data packets The actual queuing delay; It is an indicator function; For the first The maximum tolerable latency corresponding to discrete traffic flows. Using this statistic, the system can expand from judging single instantaneous errors to judging the overall latency compliance status over a period of time, thereby improving the stability of parameter updates.

[0143] In step S330, based on the time series distribution characteristics of the delay prediction residual and the actual operating energy consumption, the dynamic safety margin, trigger sensitivity coefficient, and energy consumption penalty adjustment factor are jointly updated using an adaptive feedback mechanism.

[0144] Specifically, on the one hand, when the actual queuing delay of the corresponding discrete business flow, calculated based on the delay prediction residual, exceeds the maximum tolerable delay within a preset sliding window... When the number of occurrences exceeds a preset threshold, increase the dynamic safety margin. The value of and trigger sensitivity coefficient The value of is determined to enhance the trigger response strength to the risk of latency exceeding limits.

[0145] In some implementations, when Greater than the preset number of times threshold If this occurs, it indicates that the current prediction mechanism or scheduling strategy is still too weak in responding to the latency risk of this service flow, and the system needs to increase its protection strength in subsequent cycles. In this case, the following update method can be used:

[0146] Equation (16)

[0147] Equation (17)

[0148] In the formula, and These are preset positive adjustment step sizes, and all are greater than zero; For the first Discrete traffic flow at time Dynamic safety margin; This corresponds to the trigger sensitivity coefficient. Here, increasing... This allows for a larger safety margin to be reserved in subsequent calculations of the remaining delay budget; increasing... This makes the digital trigger probability more sensitive to changes in the remaining latency budget. The synergistic effect of these two factors helps to identify potential latency exceedance risks earlier in subsequent scheduling cycles and trigger scheduling interventions in advance.

[0149] On the other hand, when the actual operating energy consumption is detected to exceed the preset energy consumption control threshold, and the actual queuing delay of the corresponding discrete service flow is determined to be lower than the maximum tolerable delay based on the delay prediction residual, At that time, increase the energy consumption penalty adjustment factor. The value of is determined, and the trigger sensitivity coefficient is reduced. The value of is chosen to reduce the energy consumption cost during resource scheduling while satisfying latency constraints.

[0150] In some implementations, if the actual operating energy consumption of the edge node Exceeding the preset energy consumption control threshold And the actual queuing delay of the corresponding business flow still meets the requirements. This indicates that while the current system meets the latency requirements, the overall scheduling strategy may be too aggressive, leading to excessively frequent resource wakeups or switches. In this case, the following update method can be used:

[0151] Equation (18)

[0152] Equation (19)

[0153] In the formula, This is the positive adjustment step size for the energy consumption penalty adjustment factor; A preset lower limit for triggering the sensitivity coefficient is used to prevent... Excessively reduced; and These are the corresponding parameters at the current time. Here, we increase... Subsequently, the impact of energy consumption cost in the joint scheduling utility calculation will be enhanced, thus prompting resource allocation to more favor the reuse of existing active resources; reducing Subsequently, the response of the digital trigger probability to changes in the remaining latency budget will be appropriately slowed down, thereby avoiding frequent high-intensity scheduling even when the system has sufficient latency margin. Through this joint update, the system can gradually suppress unnecessary high-energy-consumption operation while ensuring business latency compliance.

[0154] The parameter closed-loop dynamic correction mechanism provided in this application establishes an online feedback channel oriented towards actual operating results on the basis of the predictive scheduling framework. This allows the system to no longer rely on fixed safety margins, trigger sensitivity coefficients, and energy consumption penalty adjustment factors, but to adaptively adjust based on the deviation between actual queuing delay and predicted queuing delay, as well as the actual operating energy consumption status of nodes. On the one hand, when edge nodes frequently experience delay exceedance risks within a certain period, the system can improve the ability of subsequent scheduling to respond to risks in advance by increasing the safety margin and trigger sensitivity. On the other hand, when the system exhibits high operating energy consumption while meeting delay constraints, it can suppress unnecessary high-energy-consumption resource scheduling behavior by increasing energy consumption penalties and appropriately reducing trigger sensitivity. Thus, edge nodes can more stably balance delay guarantee requirements and energy consumption control requirements in dynamically changing discrete business flow scenarios, improving the actual adaptability and operational robustness of the entire predictive digital trigger scheduling mechanism.

[0155] Figure 4This paper illustrates a schematic diagram of the system operation mechanism of an example of a predictive digital triggering method for discrete service flows under time constraints according to an embodiment of this application. The mechanism architecture is mainly composed of four logical blocks: an input area, a core processing area, a control and constraint area, and an output area.

[0156] like Figure 4 As shown, firstly, the input area is responsible for non-intrusively collecting discrete service flows, network status data, and SLA parameters from the underlying network, and synchronously inputting them as multi-dimensional basic data into the core processing area, thereby providing a three-in-one panoramic data support of "traffic characteristics - system resources - service constraints" for subsequent forward-looking scheduling.

[0157] In the core processing area, the system sequentially performs multi-scale dependency decomposition to purify discrete traffic characteristics, followed by arrival and queuing prediction to assess congestion levels within a pre-defined time window. Subsequently, the system uses a trigger probability mapping module to transform rigid physical queuing delays into flexible scheduling urgency, and combines this with a service value metric generated by the FCI (Flow Importance Index) calculation module to establish a comprehensive scheduling weight for each service flow. Finally, in the triggering and resource allocation phase, the system pre-reorganizes and mounts local hardware resources based on these weights, and pushes the generated optimized scheduling decisions, digital trigger signals, and QoS guarantee indicators to the output area to guide the underlying hardware execution.

[0158] Furthermore, this architecture critically introduces a control and constraint region to construct a system-level adaptive evolution closed loop. This region incorporates latency constraint execution and energy consumption closed-loop control logic, which continuously feeds closed-loop feedback signals to the core processing area by monitoring the actual physical state after task execution in real time. Utilizing the extracted latency prediction residuals and actual energy consumption performance, the system dynamically corrects key control parameters in the core processing area. This bidirectional closed-loop architecture, coupled with "feedforward prediction and posterior feedback," ensures that edge nodes can continuously and stably operate at the optimal balance between service quality (latency guarantee) and global energy consumption minimization when facing highly uncertain sudden traffic surges.

[0159] To verify the effectiveness and advancement of the predictive digital triggering method proposed in this application in a real edge computing network environment, a simulation testbed was constructed based on the Discrete Event Network Simulator (NS-3). Regarding network topology and hardware configuration, the experiment simulated an aggregation network with edge scheduling nodes, setting the total link bandwidth to 10 Gbps, and physically limiting the computing power of the scheduling nodes to realistically reflect the hardware environment of the edge servers. In terms of traffic model settings, to accurately simulate the complex burst characteristics of discrete service flows, background traffic was generated using a Poisson process; while core discrete service flows were generated using a Markov Modulated Poisson Process (MMPP) to simulate high-intensity, unpredictable microsecond-level traffic bursts.

[0160] Regarding service classification and service level agreement (SLA) settings, the simulation environment is configured with three typical discrete service flows: high-priority flows (maximum tolerable latency). And comes with extremely high penalties for breach of contract), medium priority flow ( and low-priority background streams ( To highlight the advantages of the proposed solution, two baseline algorithms were introduced in the experiment: the first is the traditional Delay-Aware Scheduling Algorithm (DQAS, representing passive response scheduling based on the current queue congestion state), and the second is the Deep Reinforcement Learning-Based Scheduling Algorithm (representing mainstream black-box intelligent scheduling). These algorithms were then compared with the predictive digital triggering method proposed in this application under comprehensive and identical conditions.

[0161] To address the scheduling characteristics of discrete service flows and the physical constraints of edge nodes, this experiment constructs a multi-dimensional comprehensive evaluation index system. Key indicators include: 99th percentile tail latency, reflecting the system's ability to handle extreme bursts of traffic and a key metric for measuring the quality of high-reliability services; SLA Violation Rate, the probability that the actual queuing delay of a data packet exceeds its corresponding maximum tolerable latency; and system energy efficiency (in MB / J), used to comprehensively evaluate the global energy consumption performance of edge nodes under computation, resource allocation, and sleep / active state switching. Furthermore, computational overhead latency is introduced, specifically to quantify the CPU time consumed by the scheduling decision itself, objectively assessing the underlying physical computational cost brought about by the multi-scale dependency decomposition and heterogeneous timing prediction algorithms introduced in this application.

[0162] Figure 5A comparative simulation diagram of the cumulative distribution function (CDF) of queuing delays for different methods is shown. The horizontal axis of the comparative experimental diagram represents queuing delay, and a log-scale coordinate system is used to stretch the display to highlight tail features; the vertical axis represents the cumulative probability (CDF). The diagram includes three algorithm comparison curves: the method of the proposed embodiment, the reinforcement learning-based method, and the traditional delay-aware scheduling (DQAS) method. A vertical red dashed line at 10 ms is used as the maximum tolerable latency (SLA) red line scale for high-priority traffic flows.

[0163] like Figure 5 Simulation results show that when dealing with burst traffic of MMPP (Markov Modulated Poisson Process) with highly non-stationary characteristics, the traditional DQAS algorithm, due to its post-event response mechanism, exhibits a significant flattening at the tail of its curve, with the 99th percentile delay exceeding 45 ms, causing a large number of data packets to cross the 10 ms SLA threshold. Although the overall curve of the RL-Based method shifts relatively to the left, its tail convergence remains slow due to the state observation lag of the reinforcement learning model when facing unknown burst patterns.

[0164] In contrast, the CDF curve of the method (Proposed) provided in this application exhibits an extremely steep "truncation" characteristic at the tail. This is because this application, through multi-scale dependency decomposition and arrival queuing prediction, combined with a predicted remaining latency budget and a continuous digital trigger probability mapping mechanism, accurately quantifies the urgency level and triggers resource priority allocation in advance before actual congestion occurs. Thanks to this forward-looking flexible scheduling mechanism, the 99th percentile latency of high-priority nodes is strictly limited to within 12 ms, and the long-tail effect is significantly eliminated. At the same time, the SLA default rate of the method in this application is significantly reduced by approximately 58% compared to the traditional DQAS algorithm, which significantly improves the service reliability and determinism of edge nodes under extreme latency constraints.

[0165] Figure 6 This diagram illustrates the comparative simulation results of different methods for handling microscopic transient responses to discrete traffic bursts. The simulation captures the transient response process across two Y-axis axes within a 500 ms burst time window. The left Y-axis represents the instantaneous traffic arrival rate, represented by the gray shaded area plot; the right Y-axis represents the queue backlog length, represented by the three broken lines. The diagram visually demonstrates the... to During a sudden surge in traffic, the traditional passive scheduling (DQAS) method, the intelligent scheduling (RL) method, and the method of the present application embodiment (Proposed) are used to describe the dynamic queue backlog evolution process when dealing with typical discrete traffic surges.

[0166] like Figure 6 As shown, to deeply analyze the micro-level effectiveness mechanism of predictive digital triggering, a comparison of the transient performance of the three curves reveals that the queue length curves of the traditional DQAS and RL methods passively increase only after the instantaneous traffic surge begins, and even after the traffic subsides, it still takes a considerable amount of time to process the backlogged data packets. In contrast, the queue length curve of the method in this application's embodiment exhibits a highly significant "proactive suppression" characteristic. Specifically, in Before the actual traffic peak fully materializes, the multi-scale dependency decomposition mechanism in this application has already keenly captured the upward trend of short-term burst components and, through forward-looking queuing prediction and probability mapping, has increased the comprehensive scheduling weight of high-priority business flows in advance. Thanks to this mechanism of "pre-triggered" local resource allocation before actual congestion occurs, the peak queue backlog of this method is significantly suppressed when the traffic peak passes, reaching only about 40% of that of the traditional baseline algorithm, and quickly recedes and clears after the burst ends. These experimental results strongly demonstrate that the method in this application possesses extremely superior buffering resilience and advanced risk mitigation capabilities when dealing with microsecond / millisecond-level discrete bursts.

[0167] Figure 7 A comparative simulation diagram of the system energy efficiency and computational overhead of different methods is shown. The simulation diagram uses a scatter bubble chart to present the objective trade-off between computational power consumption and global performance of the scheduling mechanism. The horizontal axis represents the system load rate (%) of edge nodes, and the vertical axis represents the system energy efficiency (MB / J, i.e., the amount of data processed per joule of energy). The size of the scatter bubbles in the diagram visually maps the computational overhead (in microseconds) of a single decision by different scheduling algorithms. The diagram comprehensively compares the overall energy efficiency performance of the proposed method, the traditional passive scheduling (DQAS) method, and the intelligent scheduling (RL) method under different network loads.

[0168] like Figure 7Simulation analysis shows that in areas with extremely low system load and stable traffic (e.g., when the system load rate is below 30%), the computational overhead of a single scheduling decision (represented by a large bubble, approximately 150 μs) is objectively higher than that of the traditional DQAS method (represented by a small square, approximately 20 μs) which uses a simple first-come, first-served logic, due to the need for fine-grained data processing by the continuously running multi-scale dependency decomposition module and heterogeneous time-series prediction model in the method of this application. This physical computational overhead at the algorithm level results in a slightly higher basic standby power consumption of the system under extremely low load conditions compared to traditional solutions.

[0169] However, when the system enters a high-load or extremely congested state (e.g., when the system load rate is greater than 70%), the traditional DQAS algorithm, lacking congestion prediction, frequently falls into a vicious cycle of packet loss and retransmission and ineffective traversal of large backlog queues, leading to a precipitous deterioration in its system energy efficiency. In contrast, the method in this application, with its forward-looking smooth probability triggering mechanism, fundamentally avoids queue collapse. More importantly, by introducing a normalized energy consumption cost function and a joint scheduling utility optimization mechanism, the system intelligently suppresses frequent wake-ups and state switching of high-energy-consuming hardware during resource allocation. Therefore, under high-load burst scenarios, the global system energy efficiency of the method in this application exhibits a significant advantage, improving by approximately 25% compared to the traditional DQAS algorithm and by approximately 5% compared to the RL intelligent method. These experimental results fully demonstrate that the method provided in this application successfully achieves high network reliability and optimal overall energy efficiency in high-load complex edge scenarios at the cost of a slight increase in computational overhead during low-load periods.

[0170] In summary, this application addresses the technical pain points of existing network traffic scheduling in discrete service flow scenarios, namely the lack of predictability and difficulty in balancing latency and energy consumption. It proposes a predictive digital triggering scheduling method based on multi-scale dependency decomposition, a predictive remaining latency budget and trigger probability mapping mechanism, and a flow importance index-driven approach. Simulation data shows that the method of this application effectively overcomes the lag of traditional passive scheduling, strictly limiting the 99th percentile tail latency of high-priority service flows to within the tolerable extreme value, significantly reducing the SLA default rate by approximately 58% compared to traditional solutions, fundamentally eliminating the extreme long-tail effect. Furthermore, when facing microsecond-level high-intensity bursts, the proactive triggering mechanism of this application can respond in advance before the traffic peak arrives, reducing the queue backlog peak to 40% of the baseline algorithm, achieving excellent micro-transient suppression and "peak shaving and valley filling." Furthermore, the method in this application objectively follows the physical trade-off between computing power and performance. It successfully achieves a global system energy efficiency improvement of up to 25% during high-load congestion periods by sacrificing extremely small single-decision computation overhead (approximately 150 μs) during low-load periods, thus realizing the optimal trade-off between computing power and energy efficiency.

[0171] Therefore, this application provides highly deterministic QoS guarantees for latency-sensitive discrete service flows while ensuring overall controllable energy consumption and computational costs. It has significant engineering application value in high-reliability network scenarios such as edge computing and industrial IoT. Furthermore, as an extended application scenario, this predictive digital triggering scheduling mechanism is not limited to independent deployment of a single edge node. Its core architecture is also suitable for distributed collaborative scheduling in complex multi-hop network topologies and can be deeply integrated with underlying protocols such as Time-Sensitive Networking (TSN). In addition, considering the limited computing power and data silos of edge nodes in real-world industrial scenarios, the architecture of this application naturally supports the introduction of federated learning and lightweight privacy protection mechanisms. This allows edge nodes to collaboratively train and update multi-scale time-series prediction models without sharing raw traffic data, thereby further reducing the inference computation overhead of a single node and perfectly adapting to a wider range of demanding industrial application environments.

[0172] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of combined actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Secondly, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application. In the above embodiments, the descriptions of each embodiment have their own emphasis; for parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0173] Figure 8 A structural block diagram of an example of a delay-constrained predictive digital triggering system for discrete traffic flows is shown according to an embodiment of this application.

[0174] like Figure 8 As shown, the discrete traffic flow predictive digital triggering system 800 under time delay constraints includes a traffic flow sensing unit 810, a multi-scale decomposition unit 820, an arrival prediction and queuing estimation unit 830, a time delay budget triggering unit 840, a scheduling weight determination unit 850, and a resource scheduling execution unit 860.

[0175] The service flow awareness unit 810 is used to acquire multiple discrete service flows in the edge network and collect the flow feature data, network status data and service level protocol parameters corresponding to the multiple discrete service flows in real time. The flow feature data includes arrival time sequence and packet length sequence, the network status data includes current queue length and available bandwidth, and the service level protocol parameters include maximum tolerable latency.

[0176] The multi-scale decomposition unit 820 is used to perform multi-scale dependency decomposition on the arrival time series in the flow feature data of each discrete service flow, so as to separate the arrival sequence components at different time scales.

[0177] The arrival prediction and queuing estimation unit 830 is used to predict the predicted arrival volume of each discrete service flow within a future preset time window based on the separated arrival sequence components and the group length sequence, and to estimate the predicted queuing length of each discrete service flow within the future preset time window by combining the current queue length and the available bandwidth.

[0178] The delay budget triggering unit 840 is used to calculate the predicted remaining delay budget of each discrete service flow based on the predicted queue length and the maximum tolerable delay in the service level agreement parameters, and to map the predicted remaining delay budget to a digital triggering probability; the digital triggering probability is used to characterize the scheduling urgency of each discrete service flow.

[0179] The scheduling weight determination unit 850 is used to obtain the flow importance index of each discrete service flow, and to determine the scheduling weight of each discrete service flow by combining the digital trigger probability and the flow importance index corresponding to each discrete service flow; the flow importance index is used to characterize the service importance of the corresponding discrete service flow.

[0180] The resource scheduling execution unit 860 is used to dynamically adjust the service order of the scheduling queue in the edge node and allocate corresponding local resources based on the scheduling weights corresponding to each discrete service flow, so as to trigger the scheduling operation for the multiple discrete service flows in advance.

[0181] In some embodiments, this application provides a non-volatile computer-readable storage medium storing one or more programs including execution instructions. The execution instructions can be read and executed by an electronic device (including but not limited to a computer, server, or network device) to perform the steps of the discrete traffic flow predictive digital triggering method under any of the delay constraints described above.

[0182] In some embodiments, this application also provides a computer program product, the computer program product including a computer program stored on a non-volatile computer-readable storage medium, the computer program including program instructions, which, when executed by a computer, cause the computer to perform the steps of the discrete traffic flow predictive digital triggering method under any of the above-described delay constraints.

[0183] In some embodiments, this application also provides an electronic device comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of a delay-constrained discrete traffic flow predictive digital triggering method.

[0184] The above-described product can perform the methods provided in the embodiments of this application, and has the corresponding functional modules and beneficial effects for performing the methods. Technical details not described in detail in this embodiment can be found in the methods provided in the embodiments of this application.

[0185] The electronic devices in this application can exist in various forms, including but not limited to: mobile communication devices, ultra-mobile personal computer devices, portable entertainment devices, or other airborne electronic devices with data interaction functions.

[0186] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0187] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0188] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A predictive digital triggering method for discrete traffic flows under time-delay constraints, applied to edge nodes, characterized in that, The method includes: The system acquires multiple discrete service flows in the edge network and collects the corresponding flow feature data, network status data, and service level protocol parameters in real time. The flow feature data includes arrival time series and packet length series, the network status data includes current queue length and available bandwidth, and the service level protocol parameters include maximum tolerable latency. Multi-scale dependency decomposition is performed on the arrival time series in the flow feature data of each discrete service flow to separate the arrival sequence components at different time scales. Based on the separated arrival sequence components and the packet length sequence, the predicted arrival volume of each discrete service flow within a future preset time window is predicted, and combined with the current queue length and the available bandwidth, the predicted queue length of each discrete service flow within the future preset time window is estimated. For each discrete service flow, the predicted remaining latency budget of the discrete service flow is calculated based on the predicted queue length and the maximum tolerable latency in the service level agreement parameters, and the predicted remaining latency budget is mapped to a digital trigger probability; the digital trigger probability is used to characterize the scheduling urgency of each discrete service flow. The flow importance index of each discrete service flow is obtained, and the scheduling weight of each discrete service flow is determined by combining the digital trigger probability corresponding to each discrete service flow and the flow importance index; the flow importance index is used to characterize the service importance of the corresponding discrete service flow. Based on the scheduling weights corresponding to each discrete service flow, the service order of the scheduling queues within the edge nodes is dynamically adjusted and corresponding local resources are allocated to trigger scheduling operations for the multiple discrete service flows in advance.

2. The method according to claim 1, characterized in that, The multi-scale dependency decomposition of the arrival time series in the flow feature data of each discrete service flow to separate the arrival sequence components at different time scales includes: According to the preset discrete time sampling step size, the arrival time sequence of each discrete service stream is discretized to obtain an arrival intensity sequence with equal time intervals. A preset wavelet basis function is selected to perform multi-level discrete wavelet transform on the arrival intensity sequence to extract the low-frequency approximation coefficients of the highest decomposition level and the high-frequency detail coefficients of each decomposition level. According to the preset hierarchical division rules, the high-frequency detail coefficients of the shallow decomposition level are reconstructed into short-term burst components to reflect the burst arrival characteristics at a shorter time scale, the high-frequency detail coefficients of the deep decomposition level are reconstructed into periodic components to characterize the periodic arrival change characteristics, and the low-frequency approximation coefficients of the highest decomposition level are reconstructed into long-term trend components to reflect the arrival change trend at a longer time scale, so that the short-term burst components, the periodic components, and the long-term trend components together constitute arrival sequence components at different time scales.

3. The method according to claim 2, characterized in that, Based on the separated arrival sequence components and the packet length sequence, the predicted arrival volume of each discrete service flow within a future preset time window is predicted, and combined with the current queue length and the available bandwidth, the predicted queuing length of each discrete service flow within the future preset time window is estimated, including: For the aforementioned future preset time window, the short-term burst component, the periodic component, and the long-term trend component of each discrete business flow are predicted using corresponding heterogeneous time series prediction models to generate a first predicted arrival intensity, a second predicted arrival intensity, and a third predicted arrival intensity, respectively; wherein, the heterogeneous time series prediction model includes an autoregressive prediction model, a periodic stationary prediction model, and an exponentially weighted moving average model. The first predicted arrival intensity, the second predicted arrival intensity, and the third predicted arrival intensity are combined in the time domain within the future preset time window, and the combined result is multiplied by the average group length determined based on the group length sequence to calculate the predicted arrival amount within the future preset time window. Based on the fluid approximation queuing model, the current queue length, measured in data volume, is added to the predicted arrival volume, and the cumulative service data volume estimated based on the available bandwidth within the future preset time window is subtracted. The difference in data volume after subtraction is then multiplied by zero to obtain the result, which is used as the predicted queue length of the corresponding discrete service flow at the end of the future preset time window.

4. The method according to claim 3, characterized in that, For each discrete service flow, the predicted remaining latency budget is calculated based on the predicted queuing length and the maximum tolerable latency in the service level agreement parameters, and the predicted remaining latency budget is mapped to a digital trigger probability, including: The predicted queuing length is divided by the average estimated service rate to calculate the predicted queuing delay; wherein, the average estimated service rate is the estimated service rate that the edge node allocates to the corresponding discrete service flow within the future preset time window based on the available bandwidth. The predicted remaining latency budget is calculated by subtracting the predicted queuing latency and the dynamic safety margin (in terms of time) from the maximum tolerable latency set in the service level agreement parameters; the dynamic safety margin is used to compensate for prediction errors. Using the predicted remaining latency budget as the independent variable, combined with the trigger sensitivity coefficient used to control the steepness of the response, a mapping calculation is performed through a continuously smooth nonlinear probabilistic activation function to obtain the digital trigger probability. When the predicted remaining latency budget is less than zero and continues to decrease, the digital trigger probability increases nonlinearly and approaches the upper limit of the value to characterize the increased scheduling urgency. The continuously smooth nonlinear probabilistic activation function is used to transform the hard-decision latency threshold into a flexible probability metric to suppress the scheduling jitter caused by frequent resource preemption by edge nodes when dealing with discrete traffic bursts.

5. The method according to claim 4, characterized in that, The step of obtaining the flow importance index of each discrete service flow and determining the scheduling weight of each discrete service flow by combining the digital trigger probability corresponding to each discrete service flow and the flow importance index includes: Parse the service level agreement parameters to extract the basic weight of service priority and the default cost coefficient configured for the corresponding discrete service flow; Establish a sliding observation window based on the current time and statistically analyze the historical delay default rate of the corresponding discrete service flow within the sliding observation window; The default cost coefficient is weighted using a preset default cost assessment coefficient, and the historical latency default rate is weighted using a preset service quality compensation coefficient. The weighted default cost coefficient, the historical latency default rate, and the basic weight of the business priority are then linearly summed to calculate the flow importance index. The flow importance index is used to comprehensively characterize the degree of business importance determined by the preset priority, default cost, and historical service quality status. After applying a power-law nonlinear adjustment to the digital trigger probability using a preset trigger adjustment index, the adjusted digital trigger probability is multiplied by the flow importance index to calculate the scheduling weight. The scheduling weight is used to enhance the weight differentiation between discrete business flows with high scheduling urgency and discrete business flows with low scheduling urgency, so as to jointly adjust the resource allocation priority according to the scheduling urgency and business importance.

6. The method according to claim 5, characterized in that, The step of dynamically adjusting the service order of the scheduling queue within the edge node and allocating corresponding local resources based on the scheduling weights corresponding to each discrete service flow includes: The multiple discrete service flows in the scheduling queue are sorted in descending order according to the scheduling weights corresponding to each discrete service flow, so as to update the resource allocation priority of each discrete service flow. The local resource status within the edge node is analyzed, and a normalized energy consumption cost function is established for each candidate resource unit. The normalized energy consumption cost function is used to quantify the running energy consumption when assigning tasks to the corresponding candidate resource units for execution, as well as the state switching cost from dormant state to active state. When allocating local resources to each discrete service flow according to the resource allocation priority, the scheduling weight of the corresponding discrete service flow is subtracted from the energy consumption penalty term obtained by multiplying the energy consumption penalty adjustment factor and the normalized energy consumption cost function, and the joint scheduling utility value corresponding to each candidate resource unit is calculated. Under the condition of meeting the latency constraints of the corresponding discrete service flows, the target candidate resource unit that maximizes the joint scheduling utility value is selected as the corresponding local resource to trigger the scheduling operation in advance; wherein, when the scheduling weight is small and the predicted remaining latency budget is sufficient, the corresponding processing task is preferentially allocated to the candidate resource unit in the active state or the wake-up of the candidate resource unit in the dormant state is postponed through the negative penalty effect of the energy consumption penalty adjustment factor, so as to minimize the global energy consumption cost of the edge node within the latency constraint range.

7. The method according to claim 6, characterized in that, After completing the dynamic adjustment of the service order of the scheduling queue and the corresponding allocation of local resources, the method further includes performing a closed-loop dynamic correction operation on parameters, specifically including: The actual queuing delay of each discrete service flow after dynamic resource allocation is monitored in real time, and the actual operating energy consumption of the edge node is monitored simultaneously. Calculate the delay prediction residual between the actual queuing delay and the predicted queuing delay for the corresponding discrete service flow; Based on the time series distribution characteristics of the predicted delay residuals and the actual operating energy consumption, an adaptive feedback mechanism is used to jointly update the dynamic safety margin, the trigger sensitivity coefficient, and the energy consumption penalty adjustment factor, specifically including: When the actual queuing delay of the corresponding discrete service flow exceeds the maximum tolerable delay more than a preset number threshold within a preset sliding window, based on the delay prediction residual, the value of the dynamic safety margin and the value of the trigger sensitivity coefficient are increased to improve the trigger response strength to the delay exceedance risk. When the actual operating energy consumption is detected to exceed the preset energy consumption control threshold, and the actual queuing delay of the corresponding discrete service flow is determined to be lower than the maximum tolerable delay based on the delay prediction residual, the value of the energy consumption penalty adjustment factor is increased and the value of the trigger sensitivity coefficient is decreased, so as to reduce the energy consumption cost in the resource scheduling process under the condition of satisfying the delay constraint.

8. A predictive digital triggering system for discrete traffic flows under time-delay constraints, characterized in that, The system includes: The service flow awareness unit is used to acquire multiple discrete service flows in the edge network and collect the flow feature data, network status data and service level protocol parameters corresponding to the multiple discrete service flows in real time. The flow feature data includes arrival time sequence and packet length sequence, the network status data includes current queue length and available bandwidth, and the service level protocol parameters include maximum tolerable latency. A multi-scale decomposition unit is used to perform multi-scale dependency decomposition on the arrival time series in the flow feature data of each discrete service flow, so as to separate the arrival sequence components at different time scales. The arrival prediction and queuing estimation unit is used to predict the predicted arrival volume of each discrete service flow within a future preset time window based on the separated arrival sequence components and the group length sequence, and to estimate the predicted queuing length of each discrete service flow within the future preset time window in combination with the current queue length and the available bandwidth. The latency budget triggering unit is used to calculate the predicted remaining latency budget for each discrete service flow based on the predicted queue length and the maximum tolerable latency in the service level agreement parameters, and to map the predicted remaining latency budget to a digital triggering probability; the digital triggering probability is used to characterize the scheduling urgency of each discrete service flow. The scheduling weight determination unit is used to obtain the flow importance index of each discrete service flow, and to determine the scheduling weight of each discrete service flow by combining the digital trigger probability corresponding to each discrete service flow and the flow importance index; the flow importance index is used to characterize the service importance of the corresponding discrete service flow. The resource scheduling execution unit is used to dynamically adjust the service order of the scheduling queue within the edge node and allocate corresponding local resources based on the scheduling weights corresponding to each discrete service flow, so as to trigger the scheduling operation for the multiple discrete service flows in advance.

Images