An AI-SD-WAN bandwidth dynamic aggregation method based on a multi-objective optimization algorithm

Abstract

The present application relates to the technical field of SD-WAN bandwidth dynamic aggregation, aiming at the core technical defects of the existing bandwidth aggregation scheme, such as insufficient scalability of centralized architecture, difficult to balance convergence speed and accuracy of multi-objective optimization, high consumption of fixed period calculation redundancy, strategy scheduling easy to cause link shock, and weight configuration dependent on artificial experience. The present application constructs a three-level hierarchical autonomous architecture of end-edge-cloud, completes network state perception through incremental data collection triggered by SLA threshold, completes dynamic pruning of solution space based on the reference feasible region library matched by business SLA, generates bandwidth aggregation strategy by using event-driven incremental multi-objective optimization algorithm, and after lightweight twin simulation verification, it is executed, and at the same time, through incremental explainability analysis, the adaptive adjustment of optimization target weight is completed, which can be widely applied to enterprise branch interconnection, industrial internet of things wide area networking, multi-cloud resource scheduling and other wide area network communication scenes.

Description

Technical Field

[0001] This invention relates to the field of software-defined wide area network (SDWAN) technology, and in particular to an AI-SDWAN bandwidth dynamic aggregation method based on a multi-objective optimization algorithm. Background Technology

[0002] The core of existing technologies is based on a centralized management and control architecture of software-defined wide area networks (SD-WAN), combined with multi-objective optimization algorithms to achieve dynamic allocation and path optimization of bandwidth resources. Mainstream solutions collect network-wide link status, traffic characteristics, and service SLA requirements data through an SD-WAN controller. This data is used to construct a mathematical model with bandwidth utilization, transmission latency, and packet loss rate as core optimization objectives. Intelligent optimization algorithms such as genetic algorithms, NSGA-II, and deep reinforcement learning are employed to solve the model, generating multi-link bandwidth aggregation and scheduling strategies. Simultaneously, differentiated traffic scheduling is achieved by prioritizing services. Through fixed-period data collection and model iteration updates, the system adapts to dynamic changes in network status, ensuring service transmission stability and network resource utilization efficiency. This aligns with the bandwidth scheduling needs of typical SD-WAN application scenarios such as enterprise branch interconnection and cross-domain service transmission.

[0003] Existing technologies often suffer from problems with centralized control and full-scale optimization modes. Current technologies generally employ a centralized control architecture, relying on a fixed-period full-data reporting mechanism to achieve network-wide status awareness. In this mode, the SD-WAN controller needs to continuously process all network data reported by all nodes, resulting in extremely high central node computing power overhead, a large amount of redundant control signaling interaction, consuming valuable link bandwidth resources, and exacerbating transmission latency for network status awareness and scheduling policy deployment. Simultaneously, the multi-objective optimization process in existing technologies generally adopts a full solution space traversal approach without targeted pruning and constraints on the optimization solution space. This directly leads to slow iterative convergence speed of the optimization algorithm, making it unable to quickly adapt to real-time fluctuations in network status. Ultimately, this results in lagging and unreasonable bandwidth resource allocation, not only lowering the overall utilization efficiency of network bandwidth resources but also failing to provide continuous and stable SLA guarantees for high-priority services, making it difficult to meet the transmission needs of differentiated services in complex network environments. Summary of the Invention

[0004] The technical problem this invention aims to solve is that the core issues of centralized management and full-scale optimization modes in existing SD-WAN bandwidth dynamic aggregation technologies are: centralized architecture and fixed-period full-scale data reporting lead to high computing power overhead and redundant control signaling; full-scale solution space traversal optimization leads to slow convergence speed, low bandwidth resource utilization, and insufficient service SLA guarantee capability. To address this, we propose an AI-SD-WAN bandwidth dynamic aggregation method based on a multi-objective optimization algorithm.

[0005] To achieve the above objectives, this application adopts the following technical solution: an AI-SD-WAN bandwidth dynamic aggregation method based on a multi-objective optimization algorithm, comprising the following steps:

[0006] S1. Construct a three-tiered autonomous architecture for edge-cloud. This architecture is a three-layer SD-WAN network architecture divided according to data permissions and decision-making permissions, possessing edge-domain autonomy capabilities. The network is divided into an edge node layer, an edge controller layer, and a global controller layer, each with differentiated data permissions and decision-making permissions. These differentiated data permissions and decision-making permissions are mutually isolated data collection, reporting, and storage permissions, as well as policy generation, distribution, and execution decision-making permissions set at different levels. The edge node layer only has local data collection and policy execution permissions, the edge controller layer has edge-domain data aggregation and autonomous decision-making permissions, and the global controller layer has cross-domain collaboration and global rule management permissions.

[0007] S2, the edge node performs incremental feature data collection based on the service SLA requirements. The incremental feature data is the feature dimension data in the link status and service traffic data collected by the edge node that has changed beyond the preset triggering conditions of the service SLA compared to the previous reported data, excluding the unchanging full original data. The edge node only reports the changed incremental feature data to its respective edge controller when the link status or service requirements change beyond the preset triggering conditions. The preset triggering conditions are a two-dimensional reporting triggering threshold constructed based on the service SLA level.

[0008] S3, the edge controller dynamically prunes the full solution space for multi-objective optimization based on hard constraints of the service SLA, and sets an adaptive robust buffer at the boundary of the narrow feasible region after pruning. The hard constraints of the service SLA are the critical indicators that the normal operation of the service flow must meet, including the maximum tolerable end-to-end latency, the minimum guaranteed bandwidth, the maximum acceptable packet loss rate, and the maximum transmission cost limit. The full solution space for multi-objective optimization is the complete set of feasible and infeasible solutions corresponding to all link combinations, bandwidth allocation weights, and traffic scheduling rules in the SD-WAN bandwidth aggregation scenario. The narrow feasible region is the set of low-dimensional effective solutions that match the current service requirements after the full solution space is dynamically pruned based on the hard constraints of the service SLA, eliminating all infeasible solutions that do not meet the hard constraints. The adaptive robust buffer is a solution interval set outside the boundary of the narrow feasible region with an adaptively adjustable width. Solutions within the buffer are included in the optimization iteration range, which avoids frequent full pruning operations caused by small fluctuations in the link and prevents the loss of the optimal solution due to excessive pruning of the solution space.

[0009] S4. An event-triggered mechanism is used to initiate incremental multi-objective optimization, using the optimal solution converged in the previous cycle as the starting point for iteration, and generating a bandwidth aggregation strategy within the narrow feasible domain. The event-triggered mechanism is a multi-objective optimization mechanism that is only initiated when a preset network state or service demand change event occurs; no full-scale iterative calculation is performed under network steady state. The incremental multi-objective optimization is a multi-objective optimization method that uses the optimal solution converged in the previous cycle as the starting point for iteration and performs iterative calculations only within the pruned narrow feasible domain.

[0010] S5. Perform lightweight pre-simulation verification on the generated bandwidth aggregation strategy. After the verification passes, it is sent to the edge node for execution, and a multi-level automatic fallback mechanism is bound to the strategy. The lightweight pre-simulation verification is a process based on the edge lightweight digital twin image to complete the pre-verification of the compliance, stability and optimization effect of the bandwidth aggregation strategy before the strategy is sent. The multi-level automatic fallback mechanism is a closed-loop guarantee mechanism that automatically switches to the corresponding level baseline strategy based on the deviation between the actual execution effect of the strategy and the pre-simulation result.

[0011] S6. Based on the interpretability analysis framework, the target weights of multi-objective optimization are adaptively adjusted, and a traceable audit link that is tamper-proof throughout the entire process is generated simultaneously. The interpretability analysis framework is a feature importance analysis framework built on the SHAP interpretability algorithm, which is used to quantify the marginal contribution of each optimization objective to the business SLA compliance rate. The tamper-proof traceable audit link is a full-process operation log with chained storage, which fully records all operations from incremental data reporting, feasible domain pruning, incremental optimization solution, strategy pre-playing, weight adjustment, strategy issuance, and execution rollback. It is tamper-proof and traceable.

[0012] Preferably, in step S2, incremental feature data acquisition and reporting are implemented through a three-level linkage closed-loop mechanism of edge nodes, domain controllers, and global controllers. Specifically, this includes: edge nodes constructing dual-dimensional reporting trigger thresholds based on service SLA levels. These thresholds include a normal fluctuation adaptive threshold to accommodate normal link fluctuations and an extreme value fallback trigger threshold to accommodate extreme link jumps. A data change exceeding either threshold satisfies the reporting trigger condition. Edge nodes only send a reporting request to their respective domain controllers when the data change exceeds either threshold, filtering the changed incremental feature data. If no threshold is triggered, only the baseline data is updated locally, and no reporting operation is performed. Upon receiving a reporting request, the domain controller allocates differentiated reporting time slots to each edge node based on the current reporting node scale and network load status. It aggregates and compresses incremental feature data of the same feature dimension and performs batch staggered scheduling according to service SLA priority when the reported traffic exceeds a preset threshold. The global controller performs time-series synchronization of cross-domain incremental feature data and uses the synchronized unified data as the input baseline for dynamic pruning of the solution space and incremental multi-objective optimization.

[0013] Preferably, in step S3, a multi-objective optimization baseline feasible domain library matching the business SLA level is pre-constructed. This baseline feasible domain library is a standardized set of feasible domains pre-constructed according to three dimensions: business type, SLA level, and link scenario. Each baseline feasible domain corresponds to a clear business SLA hard constraint boundary and a Pareto optimal solution interval under the corresponding constraint. The Pareto optimal solution interval is the range of optimal solution values ​​obtained from multi-objective optimization that cannot optimize any other objective without degrading at least one optimization objective. The edge controller parses the SLA hard constraint requirements of the business flow, matches the corresponding baseline feasible domain from the baseline feasible domain library, and, combined with the real-time link status data of its jurisdiction, eliminates all invalid solution intervals in the baseline feasible domain that do not meet the real-time SLA hard constraints. This completes the dynamic pruning of the solution space, compressing the high-dimensional full solution space into a low-dimensional narrow feasible domain that matches the current business requirements. An adaptive robustness buffer zone is set outside the boundary of the narrow feasible region after pruning. The solutions within the buffer zone are included in the optimization iteration range. The width of the robustness buffer zone is adaptively adjusted based on the coefficient of variation and kurtosis coefficient of the link state index. Under steady-state network conditions, multi-objective optimization is performed only within the narrow feasible region. In extreme scenarios, it automatically switches to the full solution space to complete global optimization.

[0014] Preferably, in step S4, the incremental multi-objective optimization solution adopts an improved NSGA-II algorithm. The improved NSGA-II algorithm is a non-dominated sorting genetic algorithm customized and improved for the incremental optimization scenario of SD-WAN bandwidth aggregation. The Pareto optimal solution set that converged in the previous cycle is used as the core of the initial population. The Pareto optimal solution set is the set of optimal solutions obtained by multi-objective optimization that cannot optimize any other objective without degrading at least one optimization objective. The remaining individuals in the initial population are randomly generated within the narrow feasible domain after pruning. The non-dominated sorting process of the algorithm embeds SLA hard constraint verification. The non-dominated sorting is a method for ranking the individuals in the population in the multi-objective optimization algorithm, which is used to screen Pareto optimal solutions so that the non-dominated sorting level of feasible individuals is always higher than that of infeasible individuals.

[0015] Furthermore, in the improved NSGA-II algorithm, the relative change rate of the hypervolume between successive iterations, which is the relative difference between the hypervolume value of the Pareto optimal solution set of the current iteration and the hypervolume value of the previous generation, is used to quantify the convergence degree of the algorithm, adaptively adjust the execution intensity of the crossover operator and the mutation operator, and dynamically adjust the convergence judgment condition based on the incremental convergence feature.

[0016] Preferably, in S5, the lightweight pre-test verification of the bandwidth aggregation strategy adopts a three-level progressive process. The three-level progressive process is a sequential process of hard constraint compliance verification, link oscillation risk quantification verification, and global optimization effect verification. The next level of verification is executed only after the current level of verification passes. The sequence is hard constraint compliance verification, link oscillation risk quantification verification, and global optimization effect verification. The bandwidth aggregation strategy is issued only after all verifications pass.

[0017] Furthermore, in the three-level progressive process, if any level of verification fails, the bandwidth aggregation strategy is returned to the multi-objective optimization solution stage, and the constraints that failed the verification are simultaneously updated as optimization boundaries to the solution space pruning stage, terminating the subsequent verification process.

[0018] Furthermore, the lightweight pre-simulation verification is performed based on the lightweight digital twin image corresponding to the edge domain. The lightweight digital twin image is a lightweight simulation model of the edge domain network built into the edge domain controller. It retains only the core elements necessary for bandwidth aggregation strategy pre-simulation and does not require full network protocol stack modeling. The lightweight digital twin image is constructed in a layered manner, including a static topology layer, a dynamic link feature layer, and a service SLA rule layer. Incremental updates are performed on the corresponding link features in the image only when edge nodes report incremental feature data. When the core features of the lightweight digital twin image and the real network deviate from a preset threshold, a full synchronization operation is automatically triggered. After synchronization is completed, the pre-simulation verification is re-executed.

[0019] Preferably, in step S5, a multi-level automatic fallback mechanism is bound to the bandwidth aggregation strategy, corresponding to the setting of multi-level baseline strategies. The multi-level baseline strategies are pre-set compliant bandwidth aggregation strategies with different protection levels, including a primary baseline strategy, a backup baseline strategy, and an emergency backup strategy. The primary baseline strategy is a compliant strategy that has been running stably in the previous cycle, the backup baseline strategy is a historically optimal strategy that has been running stably within the preset cycle, and the emergency backup strategy is a single-link optimal path strategy. Based on the deviation between the actual execution effect of the strategy and the pre-simulation result, the automatic fallback operation of the corresponding level is triggered, and the deviation data is simultaneously fed back to the multi-objective optimization model.

[0020] Preferably, the adaptive adjustment of optimization target weights and the generation of a traceable audit link specifically include: based on an interpretability analysis framework and combined with incremental feature data after time-series synchronization, incrementally calculating the marginal contribution of each optimization target to the business SLA compliance rate, and generating an optimization target weight benchmark. Based on the real-time default risk of business SLA indicators, identifying business SLA sensitivity drift events, which are phenomena of dynamic changes in the sensitivity of business flows to different SLA indicators, the judgment criterion is that the default risk of a certain SLA indicator exceeds 80% for three consecutive sliding windows, and the weight ratio of the optimization target to which the drifting indicator belongs is increased accordingly. Upper and lower limit constraints are set for the weight of individual optimization targets. The bandwidth aggregation strategy generated after weight adjustment needs to be pre-verified; if the verification fails, the weight adjustment is revoked. Operational data of the entire weight adjustment process and all stages of strategy execution are stored in a chain to generate an immutable traceable audit link.

[0021] The technical effects and advantages of this invention are as follows: Addressing the pain points of existing SD-WAN bandwidth dynamic aggregation technologies, this invention overcomes the centralized management and control computing power bottleneck through a three-tiered autonomous architecture (end-edge-cloud). It reduces control signaling redundancy through an incremental data collection and reporting mechanism, improves scheduling convergence speed and bandwidth utilization through dynamic solution space pruning and incremental multi-objective optimization methods, and strengthens network stability and service SLA assurance capabilities through supporting policy pre-playback, adaptive weight adjustment, and full-process auditing mechanisms. This method achieves efficient aggregation and scheduling of WAN bandwidth resources, comprehensively improving the operational performance of SD-WAN networks. Attached Figure Description

[0022] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts:

[0023] Figure 1 This is a flowchart illustrating the overall process of software-defined wide area network bandwidth dynamic aggregation in this invention. Figure 2This is a diagram of the three-level hierarchical autonomous architecture of the software-defined wide area network bandwidth dynamic aggregation system of this invention; Figure 3 This is a flowchart illustrating the incremental feature data acquisition and reporting process of the present invention. Figure 4 This invention provides an improved NSGA-II incremental iterative optimization execution flowchart; Figure 5 This is a curve comparing the convergence speed of the multi-objective optimization algorithm of this invention; Figure 6 This invention provides a multi-objective optimization of the three-dimensional Pareto front surface plot; Figure 7 This is a comparison curve of the Pareto front of end-to-end latency and bandwidth utilization in this invention. Figure 8 This is the closed-loop diagram of the three-level progressive strategy pre-test verification and hierarchical automatic backoff of the present invention; Figure 9 This is a flowchart illustrating the adaptive adjustment of the target weights for interpretability optimization in this invention. Detailed Implementation

[0024] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0025] Specifically, when implementing the method, the first step is to complete the construction of a three-tiered autonomous architecture encompassing the endpoint, edge, and cloud, such as... Figure 1 , 2 As shown, the architecture is divided into three layers: edge node layer, edge controller layer, and global controller layer. Each layer is configured with different data reporting permissions and decision-making permissions.

[0026] The edge node layer, deployed on SD-WAN client CPE devices, is responsible for collecting local access link status data, local service traffic characteristic data, and local execution effect data of bandwidth aggregation policies. Link status data includes real-time bandwidth utilization, end-to-end latency, packet loss rate, and jitter. Service traffic characteristic data includes service type, bandwidth requirements, SLA level, and priority. The edge controller layer divides the SD-WAN network into multiple independent edge autonomous units based on at least one dimension: region, tenant, or service type. Each edge corresponds to one edge controller, responsible for aggregating incremental characteristic data reported by its edge nodes, maintaining the local network topology of its edge, and independently generating and distributing bandwidth aggregation policies within the edge, achieving complete autonomy within the edge. The global controller layer is responsible for maintaining the topology skeleton of the entire SD-WAN network, the status characteristics of cross-domain links, tenant global SLA rules, and offline updates of the multi-objective optimization baseline feasible domain library. It does not participate in real-time policy decisions within the edge, but is only responsible for global collaborative constraints on cross-edge traffic and coordination of policy conflicts between edge regions.

[0027] Specifically, such as Figure 3 As shown, the edge node is configured with a trigger threshold based on the service SLA. Only when the collected link status data or service demand changes exceed the preset trigger conditions will the incremental feature data of the changed part be reported to the corresponding edge controller. When the data change does not exceed the threshold under network steady state, no data reporting operation is performed. The incremental feature data is the feature dimension data of the link status and service traffic data collected by the edge node that has changed beyond the preset trigger threshold of the service SLA compared to the previous reported data. It does not include the full original data that has not changed.

[0028] The reporting trigger threshold for incremental feature data is adaptively calibrated based on the SLA sensitivity coefficient of the business flow. The calculation logic of the SLA sensitivity coefficient is as follows: the SLA sensitivity coefficient equals the difference between the business SLA default threshold and the current indicator operating value, divided by the business SLA default threshold, with a fixed value range between 0 and 1. The calculation logic of the dynamic trigger threshold is as follows: the dynamic trigger threshold equals the difference between the baseline trigger threshold multiplied by 1 and the SLA sensitivity coefficient, where the baseline trigger threshold is fixed at 25%. In a typical scenario, the end-to-end latency default threshold for high-definition video conferencing services is 50ms, the current indicator operating value is 10ms, and the calculated SLA sensitivity coefficient is 0.8, corresponding to a dynamic trigger threshold of 5%. The bandwidth default threshold for nighttime data backup services is 50Mbps, the current indicator operating value is 48Mbps, and the calculated SLA sensitivity coefficient is 0.04, corresponding to a dynamic trigger threshold of 24%, matching the threshold settings for low-sensitivity services.

[0029] The edge node synchronous configuration reports a storm suppression mechanism. When the number of incremental reporting nodes received by the edge controller within 1 second exceeds 30% of the total number of nodes under its jurisdiction, the aggregation reporting rule is automatically triggered to aggregate and compress incremental data of nodes in the same network segment, with the same service type and the same trend of change, and then process them in a unified manner.

[0030] Specifically, after completing the three-tier architecture, a multi-objective optimization baseline feasible domain library matching the business SLA level is pre-built. The full solution space is dynamically trimmed based on real-time business SLA requirements. Robust buffers are set at the boundaries of the trimmed feasible domains, and the narrow feasible domain and the full solution space are adaptively switched according to changes in network status and business needs. The multi-objective optimization baseline feasible domain library is a standardized set of feasible domains pre-built according to three dimensions: business type, SLA level, and link scenario. Each baseline feasible domain corresponds to a clear hard constraint boundary of the business SLA, and the Pareto optimal solution interval under the corresponding constraint. The hard constraint boundary of the SLA is the critical indicator that the business flow must meet to operate normally, including the maximum tolerable end-to-end latency, minimum guaranteed bandwidth, maximum acceptable packet loss rate, and maximum transmission cost ceiling.

[0031] When a service flow accesses a corresponding edge domain, the edge domain controller first parses the service flow's SLA hard constraints, matches the corresponding baseline feasible domain from the baseline feasible domain library, and then, combined with the real-time link status data of the edge domain under its jurisdiction, eliminates all invalid solution intervals in the baseline feasible domain that do not meet the real-time SLA hard constraints. This completes the dynamic pruning of the solution space, compressing the high-dimensional full solution space into a low-dimensional narrow feasible domain that matches the current service requirements. The dynamic pruning of the full solution space adopts a progressive process. The pruning priority of hard constraints is determined based on the service SLA default penalty coefficient. The solution space is pruned level by level from high to low priority, and the remaining solution space after the previous level of pruning is used as the input for the next level of pruning. The default penalty coefficient is determined by a combination of the service's SLA level, default compensation amount, and service importance. The higher the penalty coefficient, the higher the pruning priority of the corresponding hard constraint.

[0032] A robust buffer band is set outside the boundary of the narrow feasible region after pruning. The solutions within the buffer are included in the optimization iteration range, which avoids frequent full pruning operations caused by small fluctuations in the link and prevents the loss of the optimal solution due to excessive pruning of the solution space. The width of the robust buffer band is adaptively adjusted based on the coefficient of variation of the link state index. The coefficient of variation is obtained by dividing the standard deviation of the link state index by the mean. The higher the coefficient of variation, the greater the fluctuation of the link, and the larger the corresponding buffer band width.

[0033] To address the dual extreme value problem of existing fixed buffer bands being unable to adapt to extreme link fluctuations, prone to over-pruning leading to loss of optimal solutions, and under-pruning resulting in computational redundancy, this embodiment, based on the classical theory of robust optimization ε-constraints and combined with the thick-tailed distribution characteristics of extreme fluctuations in communication links, designs a core calculation formula for adaptive robust buffer band width: ;in, The normalized width of the robust buffer band corresponding to a single link is the final actual width, which is the feasible region boundary value multiplied by [the value]. ; The lower limit of the buffer band width is fixed at 0.05 to avoid the risk of the optimal solution being pruned when the extreme value of the link fluctuates due to an excessively narrow buffer band. The upper limit of the buffer band width is fixed at 0.35 to avoid the extreme risk of redundancy in solution space and waste of computing power due to an excessively wide buffer band; The coefficient of variation of the link status index is equal to the ratio of the standard deviation of the link index to the mean, and it characterizes the degree of normal fluctuation of the link. This is the kurtosis coefficient of the link status index time series, with a normal distribution kurtosis baseline of 3. >3 indicates that the link exhibits extreme fluctuations with a heavy-tailed distribution, i.e., extreme value jumps. Used to quantify the degree of extreme value deviation of the link, the buffer band is amplified only for scenarios where extreme value fluctuations exist; The upper limit of the kurtosis coefficient is fixed at 10 to avoid excessive amplification of the buffer band due to extreme outliers. , Assign coefficients to the weights, satisfying Based on the business SLA level settings, for extremely sensitive businesses, such as video conferencing, [the following information is missing]. =0.3、 =0.7, prioritizing adaptation to extreme fluctuations; for low-sensitivity businesses, such as data backup, take... =0.7、 =0.3, prioritizing control over the solution space size.

[0034] In typical scenarios, the kurtosis coefficient of leased line links =3.2, coefficient of variation =0.05, calculated by substituting the video conferencing service into the core calculation formula for adaptive robust buffer width. =0.054, corresponding to a buffer band width of ±5%, which perfectly matches the parameters in the embodiment; 5G public network link kurtosis coefficient =8.5, coefficient of variation =0.2, substituting into the calculation yields =0.293, corresponding to a buffer band width of ±30%.

[0035] Under steady-state network conditions, multi-objective optimization is performed only within the pruned narrow feasible domain. In extreme scenarios such as drastic changes in link status, changes in service SLA levels, or large-scale cross-domain scheduling requirements, the system automatically switches from the narrow feasible domain to the full solution space to perform global optimization, ensuring global optimality in extreme scenarios. After the extreme scenario ends, the system automatically switches back from the full solution space to the narrow feasible domain after the link status has been running stably for 5 minutes, resuming incremental optimization mode.

[0036] Specifically, after the solution space processing is completed, an event-triggered mechanism is used to initiate multi-objective optimization. The multi-objective optimization operation is only initiated when a preset trigger event occurs. No full-scale iterative calculation is performed when the network is in steady state. The preset trigger events include changes in link status data reported by edge nodes exceeding the SLA trigger threshold, changes in the SLA requirements or priorities of service flows, cross-domain traffic scheduling requirements, the current optimal solution shifting to the robust buffer zone boundary, and the deviation between the policy execution effect and the expected value exceeding the preset threshold.

[0037] After triggering optimization, the previously converged optimal solution is used as the starting point for iteration. Incremental iterative solutions are performed within the pruned feasible region to generate a bandwidth aggregation strategy. The incremental iterative solution employs an improved NSGA-II algorithm, which incorporates three customized improvements for SD-WAN incremental optimization scenarios: incremental population initialization, SLA hard-constraint embedded non-dominated sorting, and an adaptive crossover and mutation operator. The complete execution flow of the algorithm can be found in [link to relevant documentation]. Figure 4The incremental population initialization is implemented as follows: after triggering optimization, the Pareto optimal solution set obtained from the previous convergence is used as the core of the initial population, accounting for 60% of the total population. The remaining 40% of the initial population individuals are randomly generated within the pruned narrow feasible region. The SLA hard constraint embedded non-dominated sorting is implemented as follows: first, the population individuals are divided into feasible individuals and infeasible individuals. Feasible individuals are those that satisfy all SLA hard constraints, message out-of-order constraints, and link load constraints. Infeasible individuals are those that do not satisfy any hard constraint. The non-dominated sorting of feasible individuals... Individuals ranked higher than infeasible individuals are always ranked higher. Infeasible individuals are directly removed during the ranking process and do not participate in subsequent crowding calculations and iterations. The adaptive crossover and mutation operators are implemented as follows: the initial crossover probability is set to 0.9 and the initial mutation probability is set to 0.1. When the Pareto front change is less than 1% for two consecutive generations, the crossover probability is reduced to 0.6 and the mutation probability is increased to 0.3 to increase population diversity and avoid getting trapped in local optima. When the Pareto front change is greater than 5%, the crossover probability is increased to 0.9 and the mutation probability is reduced to 0.1 to accelerate the convergence speed.

[0038] The iteration termination condition is set as follows: the iteration terminates when any one of the following conditions is met: the convergence criterion is satisfied, the number of iterations reaches 20, or the iteration time reaches 1 millisecond. To address the problems of premature termination and getting trapped in local optima due to existing fixed convergence thresholds, and the problem of excessive iteration and redundant computational power leading to dual extremes, this embodiment, based on the classic evaluation index of multi-objective optimization hypervolume (HV), designs an incremental optimization convergence extreme value termination criterion formula specifically for the characteristics of incremental population initialization scenarios: ;in, For the first The relative rate of change of the hypervolume HV value in each iteration is used to quantify the convergence of the algorithm. For the first The hypervolume value of the Pareto optimal solution set after iteration. For the first The excess volume value of the generation; The baseline convergence threshold is fixed at 0.01, meaning that in normal scenarios, a change rate of less than 1% in the HV value is considered convergence. The incremental convergence extremum factor is calculated using the following formula: ,in This represents the number of Pareto optimal solutions that converged in the previous iteration. In this embodiment, the total population size is... =50; Extreme value constraint rule: when When γ > 0.6, γ < 0.4, the convergence threshold is tightened, avoiding the risk of premature termination of iterations when using the previous local optimum as the starting point during incremental optimization; when When γ < 0.2, γ > 0.8, the convergence threshold is relaxed, which accelerates the convergence speed in scenarios with drastic changes in the link and avoids redundant computing power.

[0039] For a comparison of the convergence speed of this embodiment with the traditional NSGA-II full optimization algorithm, please refer to [link / reference]. Figure 5 The experiment shown in the figure was completed on the NS3 network simulation platform. The experimental environment was completely consistent with that of this embodiment. The convergence judgment threshold was set to HV≥0.94. The experimental results show that the improved algorithm of this invention can meet the convergence judgment requirements of the incremental optimization convergence extreme value termination judgment formula in the 8th generation, while the traditional NSGA-II algorithm needs to be iterated to the 45th generation to converge. This verifies the significant advantage of the algorithm of this embodiment in terms of convergence speed.

[0040] After iteration terminates, the optimal solution in the Pareto optimal solution set is output, generating the bandwidth aggregation strategy for the corresponding service flow. The bandwidth aggregation strategy includes the optimal link combination for the service flow, bandwidth allocation weights, and traffic scheduling rules. The multi-objective optimization model sets four optimization objectives: maximizing link bandwidth utilization, minimizing end-to-end service latency, minimizing transmission cost, and maximizing link reliability. All optimization objectives are solved using a unified minimization framework. The model also sets four hard constraints: bandwidth aggregation efficiency constraint, packet out-of-order constraint, service SLA hard constraint, and link load constraint. The bandwidth aggregation efficiency constraint requires that the effective bandwidth after aggregation is not less than 1.2 times the effective bandwidth of a single optimal link. The packet out-of-order constraint requires that the maximum one-way latency difference between aggregated links not exceed 20 milliseconds. The link load constraint requires that the bandwidth utilization of a single link not exceed 85%. The three-dimensional Pareto front surface corresponding to the four optimization objectives in this embodiment can be found in [reference needed]. Figure 6 The figure shows the end-to-end one-way delay on the X-axis, the transmission cost per unit bandwidth on the Y-axis, and the link bandwidth utilization on the Z-axis. It comprehensively illustrates the distribution of Pareto optimal solutions under different constraints. The global optimal operating point is located at a delay of 25ms and a transmission cost of 0.25 yuan / Mbps / hour, corresponding to a bandwidth utilization of 92.3%. For a comparison of the end-to-end delay-bandwidth utilization Pareto front between this embodiment and the traditional NSGA-II algorithm, please refer to [reference needed]. Figure 7 Experimental results show that, under the same end-to-end latency constraints, this embodiment can achieve higher link bandwidth utilization. At a typical operating point of 25ms, the bandwidth utilization of this embodiment can reach 92.3%, while that of the traditional algorithm is only 84.2%, which verifies the optimization performance advantage of this embodiment.

[0041] The calculation logic for bandwidth utilization is as follows: single-link bandwidth utilization equals the link's real-time traffic bandwidth divided by the total available bandwidth of the link. The network-wide average bandwidth utilization is the arithmetic mean of the bandwidth utilization of all service access links. The core calculation process for a typical operating point is as follows: Under a 25ms end-to-end latency constraint, after dynamic pruning of the solution space, the link combination within the narrow feasible domain is 2 dedicated lines + 1 5G public network link. The total available bandwidth of the three links is 50Mbps, 30Mbps, and 20Mbps, respectively, for a total available bandwidth of 100Mbps. After optimization, the actual bandwidth used by the three links is 42.5Mbps, 25.5Mbps, and 24.3Mbps, respectively, for a total actual bandwidth of 92.3Mbps. The calculated network-wide average bandwidth utilization is 92.3%, perfectly matching the experimental data.

[0042] Specifically, such as Figure 8 As shown, after generating the bandwidth aggregation strategy, the bandwidth aggregation strategy is pre-tested and verified based on the lightweight digital twin image corresponding to the edge domain. After the verification is passed, it is sent out for execution, and a baseline strategy and an automatic fallback path are bound to each strategy.

[0043] The lightweight digital twin mirror is a lightweight simulation model of the edge network built into the edge controller. It retains only the core elements necessary for bandwidth aggregation strategy pre-simulation and does not require full network protocol stack modeling. The lightweight digital twin mirror adopts a layered construction, including a static topology layer, a dynamic link feature layer, and a service SLA rule layer. The synchronization verification between the mirror and the real network is carried out every 5 minutes to ensure the consistency between the mirror data and the real network status. If the core features of the mirror and the real network deviate by more than 5%, full synchronization is automatically triggered to ensure the accuracy of the pre-simulation.

[0044] The pre-run verification of bandwidth aggregation strategies adopts a three-level progressive process. Only after all verifications pass will the bandwidth aggregation strategy be deployed to the corresponding edge nodes for execution: The first level is hard constraint compliance verification, which verifies whether the pre-run results meet all service SLA hard constraints, packet out-of-order constraints, and link load constraints. If not, it is directly returned to the optimization module to adjust the feasible domain range and optimization target priority, and then resolved. The second level is link oscillation risk quantification verification, which calculates the change in link bandwidth utilization after strategy execution. If the change in bandwidth utilization of a single link exceeds 50%, or if the change in bandwidth utilization of more than 20% of the links in the edge domain exceeds 30%, it is determined that there is a risk of link oscillation, and it is returned to the optimization module to smooth traffic allocation weights and reduce the change in strategy. The third level is global optimization effect verification, which verifies whether the optimization targets after the pre-run are better than the currently running baseline strategy. It requires that at least 3 optimization targets are better than the baseline, and no optimization target deteriorates by more than 5%. If these conditions are met, it is determined to be a compliant strategy and can be deployed.

[0045] Each issued bandwidth aggregation policy is bound to a corresponding baseline policy and an automatic fallback path. The baseline policy is divided into three levels: primary baseline policy, backup baseline policy, and emergency backup policy. The primary baseline policy is a compliant policy that has been running stably for more than 30 minutes in the previous cycle. The backup baseline policy is the historical best policy that has been running stably for more than 1 hour in the past 24 hours. The emergency backup policy is the single-link optimal path policy. Based on the deviation between the actual execution effect of the policy and the pre-simulation result, the corresponding level of automatic fallback operation is triggered: if the deviation is between 5% and 10%, a level 1 fallback is executed, and the traffic allocation weight is adjusted online without switching the baseline policy; if the deviation is between 10% and 20%, a level 2 fallback is executed, automatically switching to the primary baseline policy, while suspending incremental optimization of the service flow and waiting for the link status to stabilize before resuming; if the deviation exceeds 20%, a level 3 fallback is executed, automatically switching to the emergency backup policy, while reporting an alarm to the edge controller, triggering a full solution space global optimization, and generating a new compliant policy.

[0046] Specifically, such as Figure 9 As shown, during the strategy execution process, the marginal impact of each optimization objective on the business SLA compliance rate is quantified based on the interpretability analysis framework, a weight benchmark is generated, the weight of the optimization objective is adaptively adjusted according to the dynamic drift of the business SLA, and the entire process of weight adjustment is recorded to form a traceable audit link.

[0047] The interpretability analysis framework is a feature importance analysis framework built on the SHAP interpretability algorithm, used to quantify the marginal contribution of each optimization objective to the business SLA compliance rate. The weight benchmarks for optimization objectives are generated based on incremental interpretability analysis. During off-peak periods, the global controller, based on the benchmark feasible domain library, pre-calculates the benchmark feature impact values ​​of each optimization objective on the business SLA compliance rate under different business types and SLA levels, and distributes them to each edge controller. During incremental optimization, only the feature impact values ​​of each optimization objective are incrementally updated based on the feature changes in this optimization, without requiring a full recalculation. The weight coefficient of a single optimization objective is equal to the absolute value of its feature impact, divided by the sum of the absolute values ​​of the feature impacts of all optimization objectives; the sum of all weight coefficients is 1.

[0048] Service SLA dynamic drift refers to the phenomenon of dynamic changes in the sensitivity of service flows to different SLA indicators. A 1-minute sliding window is used to monitor the default risk of service flow SLA indicators in real time. The default risk is calculated by dividing the current indicator value by the service SLA default threshold. When the default risk of a certain indicator exceeds 80% for three consecutive windows, it is determined that SLA drift has occurred, and the sensitivity of the corresponding optimization target increases. Based on incremental SHAP analysis results, the weight coefficients of the corresponding optimization targets are automatically adjusted. The magnitude of a single weight adjustment does not exceed ±20%. Simultaneously, upper and lower limits are set for the weights of each optimization target: the lower limit for link reliability is 0.1, and the upper limit is 0.5; the lower limit for transmission cost is 0.05, and the upper limit is 0.4, to avoid policy oscillations caused by excessive weight concentration. The policy generated for the first time after weight adjustment must be verified through a digital twin mirror pre-run. If the pre-run results show a decrease in SLA compliance rate, the adjustment is automatically revoked, and the original weight baseline is restored.

[0049] The entire process of weight adjustment is logged to form a complete traceable audit link. The traceable audit link is a chain-stored full-process operation log that fully records all operations from incremental data reporting, feasible domain pruning, incremental optimization solution, policy pre-playing, weight adjustment, policy distribution, and execution rollback. It is tamper-proof, traceable, and meets the controllable and auditable compliance requirements of telecommunications-grade networks and critical information infrastructure networks.

[0050] Example: This example is implemented in a multi-branch interconnected SD-WAN networking scenario for a chain enterprise. The enterprise headquarters is located in Shanghai, with four branches in Beijing, Guangzhou, Shenzhen, and Chengdu. Each branch deploys one SD-WAN CPE device as an edge node. Each CPE is connected to two dedicated lines from operators and one 5G public network link. The enterprise is divided into four autonomous edge units according to geographical location, with one edge controller for each edge unit. The Shanghai headquarters deploys a global controller with dual-machine hot standby, thus constructing a three-level hierarchical autonomous architecture of end-edge-cloud.

[0051] The company operates three types of core business between its branches and headquarters. The first type is a high-definition video conferencing system, which is an extremely sensitive business. The SLA requires end-to-end latency of no more than 50ms, packet loss rate of no more than 0.1%, and a minimum guaranteed bandwidth of 10Mbps. The second type is the company's SaaS office system, which is a medium-sensitive business. The SLA requires end-to-end latency of no more than 100ms, packet loss rate of no more than 1%, and a minimum guaranteed bandwidth of 2Mbps. The third type is nighttime data backup, which is a low-sensitive business. The SLA requires a minimum guaranteed bandwidth of 50Mbps and minimizes transmission costs.

[0052] Edge nodes collect local link status data and service traffic characteristics in real time. Trigger thresholds are adaptively set based on SLA sensitivity coefficients: ±3% for video conferencing service latency and ±25% for data backup service bandwidth. No data is reported in steady state; incremental feature data is reported only when data changes exceed the thresholds. The edge controller pre-builds a baseline feasible domain library corresponding to three types of services. During service flow access, feasible domains are dynamically pruned based on SLA hard constraints. Video conferencing services are prioritized for pruning based on packet loss rate and latency constraints, while data backup services are prioritized for pruning based on bandwidth and cost constraints. The adaptive robustness buffer band width is calculated using a core formula: ±5% for leased link buffers and ±30% for 5G links.

[0053] Incremental optimization is initiated only when a preset trigger event occurs. Using the Pareto optimal solution set from the previous convergence as the initial population core, an improved NSGA-II algorithm is executed incrementally within a pruned, narrow feasible region. Convergence is determined using the quantitative optimization convergence extreme value termination criterion formula, with an average of 8 iterations and a convergence time consistently below 0.8 milliseconds. The generated bandwidth aggregation strategy first undergoes a three-level progressive pre-verification in a lightweight digital twin image of the edge domain. After successful verification, it is deployed to the CPE device for execution. Each strategy is bound to a 3-level baseline strategy, and tiered rollback is automatically triggered if deviations occur during execution. Incremental SHAP analysis is used to quantify the marginal impact of each optimization objective in real time, automatically generating weight benchmarks. When the packet loss rate of the video conferencing service suddenly increases, the weights of the corresponding optimization objectives for packet loss rate and latency are automatically increased, with a single adjustment not exceeding ±20%. The entire process is recorded in audit logs, generating interpretable reports.

[0054] The actual test results of this embodiment show that the overall network link bandwidth utilization rate is stable at over 91%, the SLA compliance rate of video conferencing services is 99.95%, the effective bandwidth maintenance rate is 90.2% under extreme weak network conditions with a 40% packet loss rate, and no link oscillation events caused by policy scheduling occur. Compared with the centralized full-scale calculation scheme, the controller computing power consumption is reduced by 90%, and compared with the fixed-period full-scale reporting scheme, the control signaling bandwidth consumption is reduced by 92%.

[0055] Comparative Example: This comparative example adopts the existing conventional centralized full-scale computation SD-WAN bandwidth scheduling scheme. The deployment environment, business scenarios, and hardware configuration are completely consistent with the example. It adopts full data collection and full-scale iterative computation with a fixed period of 500 milliseconds. The multi-objective optimization weights are preset by human experience. The strategy execution adopts the mode of first issuing and then feedback. There is no pre-drilling verification step. The buffer band adopts a fixed width of ±15%. The convergence judgment adopts a fixed threshold HV change rate ≤1%.

[0056] The comparative test results show that the average convergence time of multi-objective optimization is 22 milliseconds, the highest link bandwidth utilization is 84%, the SLA compliance rate of video conferencing service is 99.7%, the control signaling bandwidth accounts for 11.2% of the total WAN bandwidth, the link oscillation caused by policy scheduling occurs at a rate of 1.1%, the effective bandwidth maintenance rate is 73.5% under extreme weak network conditions with a 40% packet loss rate, and there are two service SLA default events caused by the loss of the optimal solution in the extreme fluctuation scenario of link.

[0057] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A method for dynamic bandwidth aggregation in AI-SD-WAN based on a multi-objective optimization algorithm, characterized in that, The process includes the following steps: S1, constructing a three-tiered autonomous architecture (edge-cloud), dividing the network into an edge node layer, a domain controller layer, and a global controller layer, each with differentiated data and decision-making permissions; S2, edge nodes perform incremental feature data collection based on business SLA requirements, reporting the changed incremental feature data to their respective domain controllers only when link status or business requirements change beyond preset triggering conditions; S3, the domain controller dynamically prunes the full solution space of multi-objective optimization based on hard constraints of the business SLA, setting an adaptive robust buffer zone at the boundary of the pruned narrow feasible domain; S4, using an event-triggered mechanism to initiate incremental multi-objective optimization, using the optimal solution converged in the previous cycle as the iteration starting point, generating a bandwidth aggregation strategy within the narrow feasible domain; S5, performing lightweight pre-test verification on the generated bandwidth aggregation strategy, distributing it to edge nodes for execution after successful verification, and binding a multi-level automatic fallback mechanism to the strategy; S6, adaptively adjusting the target weights of multi-objective optimization based on an interpretable analysis framework, simultaneously generating a fully tamper-proof and traceable auditable link.

2. The AI-SD-WAN bandwidth dynamic aggregation method based on a multi-objective optimization algorithm according to claim 1, characterized in that, In S2, incremental feature data collection and reporting are achieved through a three-level linkage closed-loop mechanism of edge nodes, domain controllers, and global controllers. Specifically, this includes: edge nodes constructing dual-dimensional reporting trigger thresholds based on service SLA levels. These dual-dimensional reporting trigger thresholds include a normal fluctuation adaptive threshold to adapt to normal link fluctuations and an extreme value fallback trigger threshold to adapt to extreme link jumps. Edge nodes only send a reporting request to their respective domain controllers when the data change exceeds either trigger threshold. If no threshold is triggered, they only update the baseline data locally and do not perform a reporting operation. After receiving the reporting request, the domain controller allocates differentiated reporting time slots to each edge node based on the current reporting node scale and network load status. It performs aggregation and compression processing on incremental feature data of the same feature dimension and performs batch staggered scheduling according to service SLA priority when the reported traffic exceeds a preset threshold. The global controller performs time synchronization on cross-domain incremental feature data and uses the synchronized unified data as the input baseline for dynamic pruning of the solution space and incremental multi-objective optimization.

3. The AI-SD-WAN bandwidth dynamic aggregation method based on a multi-objective optimization algorithm according to claim 1, characterized in that, In S3, a multi-objective optimization baseline feasible domain library matching the service SLA level is pre-constructed. The baseline feasible domain library is a standardized set of feasible domains pre-constructed according to three dimensions: service type, SLA level, and link scenario. Each baseline feasible domain corresponds to a clear service SLA hard constraint boundary and a Pareto optimal solution interval under the corresponding constraint. The edge controller parses the SLA hard constraint requirements of the service flow, matches the corresponding baseline feasible domain from the baseline feasible domain library, and, combined with the real-time link status data of the edge domain under its jurisdiction, eliminates all invalid solution intervals in the baseline feasible domain that do not meet the real-time SLA hard constraints, thus completing the dynamic pruning of the solution space and compressing the high-dimensional full solution space into a low-dimensional narrow feasible domain that matches the current service requirements. An adaptive robustness buffer band is set outside the boundary of the pruned narrow feasible domain. The solutions within the buffer are included in the optimization iteration range. The width of the robustness buffer band is adaptively adjusted based on the coefficient of variation and kurtosis coefficient of the link status index. Under steady-state network conditions, multi-objective optimization is performed only within the narrow feasible domain. In extreme scenarios, it automatically switches to the full solution space to perform global optimization.

4. The AI-SD-WAN bandwidth dynamic aggregation method based on a multi-objective optimization algorithm according to claim 1, characterized in that, In S4, the incremental multi-objective optimization solution adopts the improved NSGA-II algorithm. The Pareto optimal solution set that converged in the previous cycle is used as the core of the initial population. The remaining individuals in the initial population are randomly generated within the narrow feasible region after pruning. The non-dominated sorting process of the algorithm is embedded with SLA hard constraint verification, so that the non-dominated sorting level of feasible individuals is always higher than that of infeasible individuals.

5. The AI-SD-WAN bandwidth dynamic aggregation method based on a multi-objective optimization algorithm according to claim 4, characterized in that, In the improved NSGA-II algorithm, the execution intensity of the crossover operator and the mutation operator is adaptively adjusted according to the relative change rate of the hypervolume between successive iterations, and the convergence judgment condition is dynamically adjusted based on the incremental convergence characteristics.

6. The AI-SD-WAN bandwidth dynamic aggregation method based on a multi-objective optimization algorithm according to claim 1, characterized in that, In S5, the lightweight pre-test verification of the bandwidth aggregation strategy adopts a three-level progressive process, namely, hard constraint compliance verification, link oscillation risk quantification verification, and global optimization effect verification. Only when all verifications pass will the bandwidth aggregation strategy be issued.

7. The AI-SD-WAN bandwidth dynamic aggregation method based on a multi-objective optimization algorithm according to claim 6, characterized in that, In the three-level progressive process, if any level of verification fails, the bandwidth aggregation strategy is returned to the multi-objective optimization solution stage, and the constraints that failed the verification are simultaneously updated as optimization boundaries to the solution space pruning stage.

8. The AI-SD-WAN bandwidth dynamic aggregation method based on a multi-objective optimization algorithm according to claim 6, characterized in that, The lightweight pre-simulation verification is performed based on the lightweight digital twin image corresponding to the edge domain. The lightweight digital twin image is constructed in a layered manner, including a static topology layer, a dynamic link feature layer, and a service SLA rule layer. Incremental updates are performed on the corresponding link features in the image only when the edge node reports incremental feature data. When the core features of the lightweight digital twin image and the real network deviate from a preset threshold, a full synchronization operation is automatically triggered. After synchronization is completed, the pre-simulation verification is re-executed.

9. The AI-SD-WAN bandwidth dynamic aggregation method based on a multi-objective optimization algorithm according to claim 1, characterized in that, In S5, a multi-level automatic fallback mechanism is bound to the bandwidth aggregation strategy, and a multi-level baseline strategy is set accordingly. The multi-level baseline strategy includes a primary baseline strategy, a backup baseline strategy, and an emergency backup strategy. The primary baseline strategy is a compliant strategy that has been running stably in the previous cycle, the backup baseline strategy is the historically best strategy that has been running stably within a preset cycle, and the emergency backup strategy is the single-link optimal path strategy. Based on the deviation between the actual execution effect of the strategy and the pre-simulation results, the corresponding level of automatic rollback operation is triggered, and the deviation data is simultaneously fed back to the multi-objective optimization model.

10. The AI-SD-WAN bandwidth dynamic aggregation method based on a multi-objective optimization algorithm according to claim 1, characterized in that, In step S6, the adaptive adjustment of optimization target weights and the generation of a traceable audit link specifically include: based on an interpretability analysis framework and combined with incremental feature data after time-series synchronization, incrementally calculating the marginal contribution of each optimization target to the business SLA compliance rate, and generating an optimization target weight benchmark; based on the real-time default risk of business SLA indicators, identifying business SLA sensitivity drift events, and correspondingly increasing the weight ratio of the optimization target to which the drift indicator belongs; setting upper and lower limit constraints for the weight of a single optimization target, and requiring the bandwidth aggregation strategy generated after weight adjustment to undergo pre-rehearsal verification, with the weight adjustment being revoked if the verification fails; and storing the operation data of the entire weight adjustment process and all stages of strategy execution in a chain, generating an immutable traceable audit link.

Images