Machine learning based low latency private line operation intelligent agent management and control system
Patent Information
- Application Number
- CN202611229114.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-13
- Publication Date
- 2026-10-13
AI Technical Summary
[0003]本发明的目的在于提供基于机器学习的低时延专线运维智能体管控系统,其解决了现有的专线运维智能体管控系统,面向故障发生后的判断和定位,难以在服务等级协议违约前识别低时延专线的亚健康退化趋势,也缺乏对运维处置动作执行前的影子仿真、风险预算门控和分阶段回滚控制,容易造成处置动作扩大时延波动或引发业务震荡,无法满足使用需求
1、本发明将实际时延变化拆分为链路固有时延、业务扰动时延、控制事件时延和异常退化残差时延,能够将业务突发、路由切换、服务质量策略调整、保护倒换和维护操作引起的可解释时延因素从实际时延变化中剥离出来,从而更准确地识别具有退化意义的异常残差。
Smart Images

Figure CN122909143A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of communication network operation and maintenance, and more specifically, to a machine learning-based intelligent agent management and control system for low-latency leased line operation and maintenance. Background Technology
[0002] Low-latency leased lines are commonly used in financial transactions, industrial control, cloud-network collaboration, telemedicine, and critical business transmission scenarios for government and enterprises. Their operational quality is affected by end-to-end latency, single-hop latency, jitter, packet loss rate, port queue length, link utilization, routing changes, and service quality policy adjustments. Existing patent CN115242621B discloses a network leased line monitoring method, device, equipment, and computer-readable storage medium. It determines whether a network leased line is faulty by analyzing the error value between network leased line link information, current service call success information, and historical service call success information, and identifies the faulty end based on service operation information. However, this type of solution mainly focuses on judgment and location after a fault occurs. It struggles to identify the sub-health degradation trend of low-latency leased lines before service level agreement (SLA) breaches, and lacks shadow simulation, risk budget gating, and phased rollback control before execution of maintenance actions. This can easily lead to increased latency fluctuations or service disruptions caused by maintenance actions. Therefore, there is an urgent need for an intelligent agent system capable of multi-source perception, residual identification, simulation verification, and closed-loop management of low-latency leased lines. Summary of the Invention
[0003] The purpose of this invention is to provide a machine learning-based intelligent agent management and control system for low-latency leased line operation and maintenance. This system addresses the shortcomings of existing intelligent agent management and control systems for leased line operation and maintenance, which are geared towards judgment and location after a fault occurs. These systems struggle to identify the sub-health degradation trend of low-latency leased lines before service level agreement (SLA) breaches and lack shadow simulation, risk budget gating, and phased rollback control before the execution of operation and maintenance actions. This can easily lead to increased latency fluctuations or business disruptions caused by the actions taken, and thus fail to meet user requirements.
[0004] This invention achieves the above objectives through the following technical solution: a low-latency leased line operation and maintenance intelligent agent management and control system based on machine learning, the system comprising: The multi-source acquisition module is used to collect end-to-end latency, single-hop latency, jitter, packet loss rate, port queue length, link utilization, routing change records, service quality policy change records, protection switching records, maintenance operation logs, and business traffic data of low-latency leased lines. The path time alignment module is used to determine the source node, destination node, intermediate forwarding node, primary path and backup path based on the leased line service identifier, and to map multi-source operation and maintenance data with different sampling periods to the same leased line path and the same time axis to form a leased line operation status matrix. The delay disturbance decomposition module is used to decompose actual delay changes into link-inherent delay, service disturbance delay, control event delay, and abnormal degradation residual delay through a machine learning model. The residual fingerprint generation module is used to generate time-delay causal residual fingerprints based on the time delay of abnormally degraded residuals. The intelligent agent decision-making module is used to identify the sub-healthy state before the default of the service level agreement based on the time delay causal residual fingerprint recognition, and generate candidate operation and maintenance actions. The shadow simulation gating module is used to predict the risk consumption of candidate operation and maintenance actions in the shadow leased line simulation model, and to control whether the candidate operation and maintenance actions enter the phased issuance process based on the latency risk budget. The phased execution module is used to progressively issue, verify, and roll back candidate operation and maintenance actions that have passed the gating.
[0005] Furthermore, the path time alignment module includes: Business path parsing unit, time granularity unification unit, and state matrix construction unit; The service path parsing unit queries the service activation record, routing table and protection relationship based on the leased line service identifier to determine the source node, destination node, intermediate forwarding node, primary path and backup path; The unified time granularity unit uses the monitoring indicator with the highest service level agreement requirement as the benchmark sampling period, sets event markers for event-type records, and performs window aggregation for continuous indicators; The state matrix construction unit organizes and aligns the multi-source operation and maintenance data according to the node dimension, link dimension, business flow dimension and time dimension to generate a leased line operation state matrix; the event marker includes event type, occurrence node, start time, end time and affected path number.
[0006] Furthermore, the time delay perturbation decomposition module includes: Model input unit, component output unit, and residual determination unit; The model input unit is used to input the leased line operation status matrix, link length, node forwarding capability, link bandwidth, peak service traffic, port queue length, link utilization, routing switching records, quality of service policy adjustment records, protection switching records, and maintenance operation logs into the machine learning model; The component output unit is used to output the inherent delay of the link, the delay caused by service disturbances, and the delay caused by control events; The residual determination unit is used to determine the difference between the actual delay change and the link inherent delay, service disturbance delay and control event delay as the abnormal degradation residual delay.
[0007] Furthermore, the time delay perturbation decomposition module also includes a training sample construction unit; The training sample construction unit is used to use low-load stable operation samples as link inherent latency training samples, historical samples corresponding to service traffic peaks and port queue backlogs as service disturbance latency training samples, and latency change samples after routing switching, service quality policy adjustment, protection switching and maintenance operations as control event latency training samples. The machine learning model establishes the contribution relationship of different factors to the actual time delay variation based on the above training samples.
[0008] Furthermore, the residual fingerprint generation module includes: Residual feature extraction unit and fingerprint encoding unit; The residual feature extraction unit is used to extract the residual duration, residual growth slope, residual fluctuation density, residual cross-node propagation direction, residual lag correlation coefficient with packet loss rate, and residual lag correlation coefficient with port queue length of abnormally degraded residual delay within a sliding time window. The fingerprint encoding unit is used to combine the above residual features into a time-delay causal residual fingerprint; The direction of residual propagation across nodes is encoded according to the order of nodes where the abnormally degraded residual first exceeds the residual threshold.
[0009] Furthermore, the residual fingerprint generation module also includes a threshold determination unit; The threshold determination unit is used to determine the residual threshold based on the residual distribution of the historical healthy operation phase, and to determine the residual growth slope threshold based on the residual slope distribution of the historical healthy operation phase. Wherein, the residual threshold is taken as the 95th or 99th quartile of the absolute value of the historical health residuals, and the residual growth slope threshold is taken as the 95th quartile of the absolute value of the slope of the historical health residuals. When the business priority is higher than the preset priority level, the residual threshold is taken as the 95th percentile of the absolute value of the historical healthy residual.
[0010] Furthermore, the intelligent agent decision-making module includes: Sub-health identification unit, degradation localization unit, and candidate action generation unit; The sub-health identification unit is used to compare the current latency causal residual fingerprint with the historical health baseline, the residual template of the same type of leased line, and the residual sample before the historical failure through dynamic time normalization distance comparison. When the similarity is higher than the similarity threshold, the abnormal degradation residual latency is higher than the residual threshold, and the residual growth slope is higher than the residual growth slope threshold, the low latency leased line is determined to be in a sub-healthy state before the service level agreement default. The degradation location unit is used to output the degradation location, degradation type, and risk level based on the node where the residual first appears, the direction of residual propagation across nodes, and the coupling index. The candidate action generation unit is used to generate candidate operation and maintenance actions based on the degradation location, degradation type, and risk level.
[0011] Furthermore, the candidate action generation unit is configured with a candidate action library; The candidate action library includes path adjustment actions, backup link switching actions, service quality queue rearrangement actions, rate limiting policy adjustment actions, abnormal node isolation actions, protection failover actions, and observation mode actions. When the degradation location is concentrated in a single intermediate forwarding node and the hysteresis correlation coefficient between the residual and the port queue length is higher than the coupling threshold, the candidate action generation unit prioritizes generating a service quality queue rearrangement action or an abnormal node isolation action. When degradation propagates continuously along the primary path, the candidate action generation unit prioritizes generating backup link switching actions or path adjustment actions.
[0012] Furthermore, the shadow simulation gating module includes: Shadow leased line construction unit, motion simulation unit, risk budget calculation unit, and gating decision unit; The shadow leased line construction unit constructs a shadow leased line simulation model that is synchronously updated with the real leased line, based on the real leased line topology, primary path, backup path, link bandwidth, node forwarding capability, port queue configuration, service quality policy, historical latency curve, and current business traffic distribution. The action simulation unit is used to simulate each candidate operation and maintenance action without changing the actual leased line configuration, and to predict the end-to-end latency, jitter, packet loss rate, routing convergence time and the impact on service continuity after the action is executed. The risk budget calculation unit is used to calculate the latency risk budget based on the service level agreement latency limit, the current end-to-end actual latency, business priority, historical latency fluctuation range, historical jitter percentile, fault recovery time limit and customer business continuity requirements. The gating decision unit is used to allow candidate operation and maintenance actions to enter the phased issuance process when the predicted risk consumption does not exceed the latency risk budget and the simulated end-to-end latency is lower than the service level agreement latency limit.
[0013] Furthermore, the phased execution module includes: Action decomposition unit, feedback verification unit, and rollback control unit; The action splitting unit is used to split the candidate operation and maintenance handling actions through gating into multiple execution units corresponding to single node adjustment, single queue adjustment, partial traffic switching or backup path detection; The feedback verification unit is used to trigger the multi-source acquisition module to re-acquire multi-source operation and maintenance data, trigger the path time alignment module to regenerate the dedicated line operation status matrix, trigger the delay disturbance decomposition module to recalculate the abnormal degradation residual delay, and trigger the residual fingerprint generation module to regenerate the delay causal residual fingerprint after each execution unit is completed. When the new latency causal residual fingerprint converges and the end-to-end latency, jitter, or packet loss rate improves, the feedback verification unit controls the subsequent execution unit to continue sending. When the new latency causal residual fingerprint fails to converge, or when any of the end-to-end latency, jitter, or packet loss rate exceeds the rollback trigger threshold, the rollback control unit suspends subsequent execution units and restores the low-latency leased line to the previous stable configuration state.
[0014] The beneficial effects of this invention are as follows: 1. This invention breaks down actual latency changes into link-inherent latency, service disturbance latency, control event latency, and abnormal degradation residual latency. It can separate the explainable latency factors caused by service bursts, routing switching, service quality policy adjustments, protection switching, and maintenance operations from actual latency changes, thereby more accurately identifying abnormal residuals with degradation significance.
[0015] 2. This invention generates a latency causal residual fingerprint by considering the residual duration, residual growth slope, residual fluctuation density, residual cross-node propagation direction, and the coupling relationship between residual and packet loss rate and port queue length. This fingerprint can identify sub-health trends before service level agreement (SLP) defaults occur on low-latency leased lines and output the degradation location, degradation type, and risk level, thereby improving the ability to identify faults in advance.
[0016] 3. After the intelligent agent generates candidate operation and maintenance actions, the present invention simulates the execution effects of path adjustment, backup link switching, service quality queue rearrangement, rate limiting policy adjustment, abnormal node isolation, protection switching, and observation mode actions through a shadow leased line simulation model. This allows for the pre-assessment of the impact of different actions on end-to-end latency, jitter, packet loss rate, routing convergence time, and service continuity, thereby reducing the risk of blind adjustments.
[0017] 4. This invention introduces a latency risk budget, which compares the predicted risk consumption of candidate maintenance actions with the latency risk budget that the current leased line can bear. Only when the predicted risk consumption does not exceed the budget and the simulated end-to-end latency is lower than the service level agreement latency limit is the phased deployment process allowed. When the above conditions are not met, the system performs observation, narrows down the action scope, regenerates candidate actions, delays execution, or performs manual review, thereby avoiding the handling action itself from causing new network instability.
[0018] 5. This invention breaks down the gated operation and maintenance actions into multiple verifiable small steps, which are executed step by step in the form of single node adjustment, single queue adjustment, partial traffic switching or backup path detection. After each stage is executed, a latency causal residual fingerprint is regenerated. The execution is continued, paused or rolled back to the previous stable configuration state based on whether the residual converges and whether the latency, jitter and packet loss rate have improved. This achieves closed-loop verification and security control of low-latency leased line operation and maintenance actions. Attached Figure Description
[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating the overall system workflow of the present invention. Figure 2 This is a flowchart of the time delay disturbance decomposition of the present invention; Figure 3 This is a flowchart of the intelligent decision-making and gating process of the present invention; Figure 4 This is a flowchart illustrating the phased execution and feedback process of the present invention. Detailed Implementation
[0020] The present application will now be described in further detail with reference to the accompanying drawings. It should be noted that the following specific embodiments are only used to further illustrate the present application and should not be construed as limiting the scope of protection of the present application. Those skilled in the art can make some non-essential improvements and adjustments to the present application based on the above application content.
[0021] Example 1: Please see Figure 1-4 This invention provides a machine learning-based intelligent agent management and control system for low-latency leased lines. The system continuously monitors the operational status of low-latency leased lines, identifies anomalies and degradation, generates candidate maintenance actions, performs shadow simulation gating, phased execution, and rollback control. Low-latency leased lines can be financial transaction lines, industrial control lines, cloud-network collaboration lines, telemedicine lines, or government and enterprise private network lines.
[0022] The system comprises a multi-source acquisition module, a path-time alignment module, a delay disturbance decomposition module, a residual fingerprint generation module, an agent decision-making module, a shadow simulation gating module, and a phased execution module. These modules communicate with each other via data interfaces, message queues, or shared databases to form a closed-loop management process from operational status acquisition to feedback verification of maintenance and handling actions.
[0023] The multi-source acquisition module is used to collect multi-source operation and maintenance data for low-latency leased lines. This multi-source data includes end-to-end latency, single-hop latency, jitter, packet loss rate, port queue length, link utilization, routing change records, quality of service policy change records, protection switching records, maintenance operation logs, and service traffic data. Specifically, end-to-end latency characterizes the overall transmission time of a service flow from the source node to the destination node; single-hop latency characterizes the forwarding time between adjacent nodes; jitter characterizes the stability of latency changes; packet loss rate characterizes packet transmission loss; port queue length characterizes the degree of data queuing and backlog at the port; link utilization characterizes the occupancy of link bandwidth resources; routing change records, quality of service policy change records, protection switching records, and maintenance operation logs characterize control events that may cause latency changes; and service traffic data characterizes service bursts, traffic peaks, and changes in service traffic distribution.
[0024] Before multi-source operation and maintenance data enters subsequent calculations, the system first performs dimension unification processing. End-to-end latency, single-hop latency, abnormal degradation residual latency, jitter, route convergence time, and fault recovery time are uniformly converted to milliseconds; packet loss rate, link utilization, service quality queue occupancy rate, and business continuity impact are uniformly converted to dimensionless proportional values between 0 and 1; business traffic data is in megabits per second as the raw unit and is divided by the corresponding link bandwidth before entering the model to convert it into business traffic occupancy rate; if the port queue length is collected in terms of the number of data packets or bytes, it is divided by the corresponding port queue capacity to convert it into queue occupancy rate; route change records, service quality policy change records, protection switching records, and maintenance operation logs are converted into event markers, with event marker values of 0 or 1, and accompanied by event type, occurrence node, start time, end time, and affected path number. Through the above processing, data from different sources, with different sampling granularities, and in different units can be compared within the same computing framework.
[0025] The path time alignment module includes a service path resolution unit, a time granularity unification unit, and a state matrix construction unit. The service path resolution unit queries service activation records, routing tables, and protection relationships based on the leased line service identifier to determine the source node, destination node, intermediate forwarding nodes, primary path, and backup path. The leased line service identifier can be a leased line number, customer service number, virtual leased line instance number, or network slice service number. Through service path resolution, the system can clearly identify the actual transmission path and backup protection path corresponding to each low-latency leased line.
[0026] The unified time-granularity unit uses the monitoring indicators with the highest service level agreement requirements as the baseline sampling period. Event-type records are marked with event tags, and continuous indicators are aggregated within a window. Event-type records include routing change records, service quality policy change records, protection switching records, and maintenance operation logs; continuous indicators include end-to-end latency, single-hop latency, jitter, packet loss rate, port queue length, link utilization, and service traffic data. For data collected at different periods, the unified time-granularity unit resamples, aggregates, or marks data according to a unified time window, enabling different data sources to be mapped to the same time axis.
[0027] Time window length The priority level is determined based on the dedicated line service level and equipment data acquisition capabilities. For high-priority services such as financial transactions, industrial control, and remote control, [further details are needed]. Take 1 to 5 seconds; for general cloud-network collaboration and government / enterprise office applications, The sampling period is 10 to 60 seconds. If the service level agreement (SLA) explicitly specifies the monitoring period, the SLA-specified monitoring period will be used as the priority. If different devices have different sampling periods, the shortest sampling period among latency, packet loss rate, and jitter, which also meets the reliability requirements, will be used as the baseline sampling period. Other indicators will be mapped to the same time window through window aggregation, nearest neighbor alignment, or event marking. Window aggregation includes the average, maximum, 95th percentile, and median values. End-to-end latency and jitter will preferentially use the 95th percentile value, while link utilization and queue occupancy will preferentially use the maximum or average value. Event-type records will be marked with whether an event has occurred.
[0028] The state matrix construction unit organizes and aligns the multi-source operation and maintenance data according to the dimensions of node, link, service flow, and time to generate a dedicated line operation state matrix. The dedicated line operation state matrix is represented as follows: in, Indicates the first A matrix of dedicated line operation status corresponding to each time window; This indicates the number of nodes, links, or service flows included in the analysis under the current dedicated line path; Indicates the first The node, link, or service flow in the first The runtime status vector is defined within a time window. The runtime status vector includes end-to-end latency, single-hop latency, jitter, packet loss rate, port queue length, link utilization, event markers, and service traffic characteristics.
[0029] The delay perturbation decomposition module includes a model input unit, a component output unit, a residual determination unit, and a training sample construction unit. The model input unit is used to input the leased line operation status matrix, link length, node forwarding capacity, link bandwidth, peak service traffic, port queue length, link utilization, routing switching records, quality of service policy adjustment records, protection switching records, and maintenance operation logs into the machine learning model. The machine learning model can employ a gradient boosting tree model, a random forest model, a temporal neural network model, or an attention mechanism model.
[0030] The training sample construction unit is used to build different types of training samples. Specifically, low-load stable operation samples are used as training samples for inherent link latency; historical samples corresponding to peak service traffic and port queue backlogs are used as training samples for service disturbance latency; and samples of latency changes after routing switching, quality of service policy adjustments, protection switching, and maintenance operations are used as training samples for control event latency. Through these training samples, the machine learning model establishes the contribution relationship between different factors to actual latency changes.
[0031] The component output unit outputs the inherent link delay, service disturbance delay, and control event delay. The residual determination unit determines the difference between the actual delay variation and the resulting values of the inherent link delay, service disturbance delay, and control event delay as the abnormal degradation residual delay, specifically expressed as: in, Indicates the first The abnormal degradation residual delay within a time window is measured in milliseconds. This represents the actual time delay change of the current time window relative to the reference time window, measured in milliseconds. This represents the inherent latency of the link, measured in milliseconds. This represents the latency caused by service disturbances, measured in milliseconds. The control event latency is expressed in milliseconds. The inherent link latency is determined by low-load stable operation samples; the service disturbance latency is determined by training with samples of service traffic peaks and port queue backlogs; and the control event latency is determined by training with samples of latency changes after historical control events. Through these methods, the system can isolate interpretable factors from actual latency changes and retain abnormal degradation residual latency that indicates degradation.
[0032] The residual fingerprint generation module includes a residual feature extraction unit, a fingerprint encoding unit, and a threshold determination unit. The threshold determination unit is used to determine the residual threshold based on the residual distribution during historical healthy operation phases. The residual growth slope threshold is determined based on the residual slope distribution during historical healthy operation phases. Among them, the residual threshold The unit of measurement is milliseconds, and the residual threshold is taken as the 95th or 99th quartile of the absolute value of the historical health residual. For high-priority services, to improve early identification capability, the residual threshold is taken as the 95th quartile of the absolute value of the historical health residual. For ordinary services, to reduce the false alarm rate, the residual threshold is taken as the 99th quartile of the absolute value of the historical health residual. If there is a measurement error in the equipment delay acquisition, the residual threshold should also be no less than twice the upper limit of the measurement error to avoid misjudgment caused by acquisition jitter.
[0033] Residual growth slope threshold The unit of measurement is milliseconds per time window, or milliseconds per second after conversion. The residual growth slope threshold is taken as the 95th quartile of the absolute value of the historical healthy residual slope. When the business priority is high, a value between the 90th and 95th quartiles can be used to improve the ability to identify early continuous degradation; when the business priority is low, a value between the 95th and 99th quartiles can be used to reduce false alarms caused by short-term fluctuations.
[0034] The residual feature extraction unit is used to extract the residual duration, residual growth slope, residual fluctuation density, residual cross-node propagation direction, residual lag correlation coefficient with packet loss rate, and residual lag correlation coefficient with port queue length within a sliding time window. The residual duration characterizes the continuous duration of the abnormally degraded residual; the residual growth slope characterizes the rate at which the abnormally degraded residual increases over time; the residual fluctuation density characterizes the frequency with which the abnormally degraded residual crosses the residual threshold; and the residual cross-node propagation direction is encoded according to the node order in which the abnormally degraded residual first exceeds the residual threshold.
[0035] Sliding window length The sliding window coverage duration is determined based on the service failure propagation speed and network convergence time. For high-priority services, the preferred sliding window coverage duration is 30 to 60 seconds; for ordinary services, the preferred sliding window coverage duration is 2 to 5 minutes. The system follows... Determine the number of time windows contained in the sliding window, where This represents the duration of the sliding window coverage. If the calculated result is not an integer, it is rounded up. To avoid false positives due to an excessively short window, Not less than 6; to avoid response delays caused by excessively long windows, for high-priority services, No more than 12, for general business. No more than 30.
[0036] In one alternative implementation, the residual growth slope and residual fluctuation density are determined as follows: in, Indicates the first The growth slope of abnormally degraded residuals within a time window; Indicates the number of time windows contained in the sliding window; Indicates the first Abnormal degradation residual delay within a time window; This represents the average of the time sequence numbers within the sliding window; This represents the mean residual delay within the sliding window; This represents the residual fluctuation density, which is a dimensionless proportional value. This indicates the number of times the residual changes from below the residual threshold to above the residual threshold, or from above the residual threshold to below the residual threshold, within the sliding window. Residual fluctuation density threshold. The value is determined based on the distribution of the number of times the residual crosses above and below the residual threshold during the historical healthy operation phase, with the 95th quantile of the historical healthy residual fluctuation density being the preferred choice.
[0037] In one alternative implementation, the hysteresis correlation coefficient between the residual and the packet loss rate and the port queue length is determined as follows: in, Representing the abnormally degraded residual sequence and index sequence The maximum lag correlation coefficient between them is a dimensionless value between 0 and 1; index series This can be a sequence of packet loss rates or a sequence of port queue lengths. Indicates the number of lag time windows; Indicates the maximum number of hysteresis windows allowed for comparison; Indicator sequence The mean within the sliding window. When When the packet loss rate sequence is used, Used to characterize the degree of coupling between anomalous degraded residuals and packet loss variations; when When the port queue length sequence is used, Used to characterize the degree of coupling between anomalous degenerate residuals and queue backlog.
[0038] Coupling threshold The lag correlation coefficients between residuals and packet loss rate, and between residuals and port queue length, are determined based on the distribution of these coefficients in historical healthy samples, with the 95th quantile of the historical healthy lag correlation coefficients being the preferred choice. If the current lag correlation coefficient between residuals and port queue length is higher than... If the degradation type is determined to be queue backlog type degradation; if the lag correlation coefficient between the current residual and the packet loss rate is higher than... If so, the degradation type will be classified as either link quality degradation or forwarding anomaly degradation. Maximum hysteresis window number. Based on the sampling period and network convergence time, the optimal method is to... This covers a typical route convergence or protection switching process. For example, with a sampling period of 5 seconds and a typical convergence time of 30 seconds, Take 6.
[0039] The fingerprint encoding unit combines the residual duration, residual growth slope, residual fluctuation density, residual cross-node propagation direction, lag correlation coefficient between residual and packet loss rate, and lag correlation coefficient between residual and port queue length into a time-delay causal residual fingerprint. The time-delay causal residual fingerprint is represented as: in, Indicates the first The time-delay causal residual fingerprint corresponding to each time window; Indicates the duration of abnormal degradation residuals, measured in milliseconds or the number of time windows; This represents the slope of the residual growth, measured in milliseconds per time window. This represents the residual fluctuation density, which is a dimensionless proportional value. Indicates the direction of residual propagation across nodes; This represents the lag correlation coefficient between residuals and packet loss rate; This represents the lag correlation coefficient between the residual and the port queue length.
[0040] The intelligent agent decision-making module includes a sub-health identification unit, a degradation localization unit, and a candidate action generation unit. The sub-health identification unit performs dynamic time-normalized distance comparison between the current latency causal residual fingerprint and historical health baselines, residual templates of similar leased lines, and historical pre-failure residual samples. When the similarity exceeds a similarity threshold, the abnormal degradation residual latency exceeds a residual threshold, and the residual growth slope exceeds a residual growth slope threshold, the low-latency leased line is determined to be in a sub-healthy state prior to a service level agreement (SLA) default. (Similarity threshold) This is a dimensionless value between 0 and 1, determined based on the recall requirements of samples prior to historical failures, with the lowest similarity value preferred to cover more than 90% of the samples prior to historical failures. If the number of samples prior to historical failures is insufficient, a residual template of the same type of dedicated line will be used to supplement it, and it will be continuously updated in subsequent operations based on actual alarms, handling results, and rollback records.
[0041] Before performing dynamic time-normalized distance comparison, all fingerprint features are dimensionless. Among them, time-delay features are divided by the corresponding historical health 95th percentile value, proportional features are kept in the range of 0 to 1, and propagation direction features are converted into node sequence encoding or path direction encoding to avoid bias in distance calculation caused by different dimensional features.
[0042] The degradation localization unit outputs the degradation location, degradation type, and risk level based on the node where the residual first appears, the direction of residual propagation across nodes, and coupling indicators. For example, when an abnormal degradation residual first appears at an intermediate forwarding node, and the hysteresis correlation coefficient between the residual and the port queue length is higher than the coupling threshold, the degradation type is identified as node queue backlog degradation; when an abnormal degradation residual propagates continuously along the primary path, and the hysteresis correlation coefficient between the residual and the packet loss rate is higher than the coupling threshold, the degradation type is identified as link quality degradation or forwarding anomaly degradation.
[0043] The candidate action generation unit is configured with a candidate action library. This library includes path adjustment actions, backup link switching actions, quality of service queue rearrangement actions, rate limiting policy adjustment actions, abnormal node isolation actions, protection failover actions, and observation mode actions. When the degradation location is concentrated on a single intermediate forwarding node and the hysteresis correlation coefficient between the residual and the port queue length is higher than the coupling threshold, the candidate action generation unit prioritizes generating quality of service queue rearrangement actions or abnormal node isolation actions. When degradation propagates continuously along the primary path, the candidate action generation unit prioritizes generating backup link switching actions or path adjustment actions. When the abnormal degradation residual does not continuously exceed the residual threshold, the candidate action generation unit prioritizes generating observation mode actions.
[0044] If the candidate action generation unit uses candidate action ranking weights, these weights are not fixed manually, but are determined based on historical success rates, average recovery times, number of rollbacks, and the degree of impact on customer business. Specifically, a higher historical success rate corresponds to a higher ranking weight for the action; a shorter average recovery time corresponds to a higher ranking weight for the action; more historical rollbacks correspond to a lower ranking weight for the action; and a higher degree of impact on customer business corresponds to a lower ranking weight for the action. Each time the candidate action library is updated or the machine learning model is retrained, the system re-evaluates the weights and limits any single weight to no more than 50% of the sum of all weights to avoid any single metric excessively dominating the action ranking.
[0045] The shadow leased line simulation gating module includes a shadow leased line construction unit, an action simulation unit, a risk budget calculation unit, and a gating decision unit. The shadow leased line construction unit builds a shadow leased line simulation model that is synchronized with the real leased line, based on the real leased line topology, primary path, backup path, link bandwidth, node forwarding capacity, port queue configuration, quality of service policy, historical latency curves, and current business traffic distribution. The shadow leased line simulation model is used to simulate the execution effects of candidate operation and maintenance actions without changing the real leased line configuration.
[0046] The action simulation unit simulates each candidate operational action, predicting end-to-end latency, jitter, packet loss rate, routing convergence time, and service continuity impact after the action is executed. It also predicts risk consumption. Used to indicate candidate operation and maintenance actions The potential risk occupancy caused by the current leased line operation status is a dimensionless proportion between 0 and 1. The predicted risk consumption is determined by converting the end-to-end latency increment, jitter increment, packet loss rate increment, routing convergence time, and service continuity impact into their respective occupancy proportions relative to the allowable upper limit, and then taking the maximum occupancy proportion. By using the maximum occupancy proportion instead of a simple average, the risk of a certain key indicator can be prevented from being diluted by other low-risk indicators. For example, if the backup link switching action has an occupancy proportion of 0.35 for end-to-end latency, 0.42 for jitter, 0.18 for packet loss rate, and 0.60 for routing convergence time, then the predicted risk consumption for this action is 0.60.
[0047] The risk budget calculation unit is used to calculate the latency risk budget based on the service level agreement latency cap, current end-to-end actual latency, service priority, historical latency fluctuation range, historical jitter percentile, fault recovery time, and customer business continuity requirements. Latency Risk Budget This is used to limit the risk space that can be consumed before the execution of candidate operation and maintenance actions. It is consistent with the predicted risk consumption and is preferably expressed as a dimensionless ratio value between 0 and 1. The current latency margin is the difference between the service level agreement latency limit and the current end-to-end actual latency, measured in milliseconds. For high-priority services, the latency risk budget is 30% to 50% of the current latency margin; for ordinary services, the latency risk budget is 50% to 70% of the current latency margin. If the historical latency fluctuation range is large, the 95th percentile of historical jitter is high, or the customer's business continuity requirements are high, then a lower value within the above range should be used; if the historical operation is stable and the fault recovery time limit is lenient, then a higher value within the above range should be used.
[0048] The gating decision unit is used to allow candidate maintenance actions to enter the phased deployment process when the predicted risk consumption does not exceed the latency risk budget and the simulated end-to-end latency is lower than the service level agreement latency limit. in, Indicates the candidate operation and maintenance handling action; This indicates the gating result. A value of 1 means that the candidate operation and maintenance action is allowed to enter the phased issuance process, while a value of 0 means that direct issuance is prohibited. This indicates the predicted risk consumption of candidate operation and maintenance actions; This indicates the current latency risk budget for the dedicated line; This represents the end-to-end latency after the predicted action is executed in the shadow leased line simulation model, measured in milliseconds. This represents the maximum latency specified in the service level agreement, measured in milliseconds. When the gating result is 0, the system performs observation, narrows down the action range, regenerates candidate actions, delays execution, or requires manual review.
[0049] The phased execution module includes an action splitting unit, a feedback verification unit, and a rollback control unit. The action splitting unit is used to break down gated candidate maintenance actions into multiple execution units corresponding to single-node adjustments, single-queue adjustments, partial traffic switching, or backup path detection. For example, for a backup link switching action, backup path detection can be performed first, followed by a small proportion of service traffic switching, and the switching proportion can be expanded based on feedback results; for a service quality queue reordering action, parameters of a single queue can be adjusted first, and then expanded to related service flows based on feedback results.
[0050] The feedback verification unit is used to trigger the multi-source acquisition module to re-acquire multi-source operation and maintenance data, the path time alignment module to regenerate the dedicated line operation status matrix, the delay disturbance decomposition module to recalculate the abnormal degradation residual delay, and the residual fingerprint generation module to regenerate the delay causal residual fingerprint after each execution unit is completed. When the new delay causal residual fingerprint converges and the end-to-end delay, jitter, or packet loss rate improves, the feedback verification unit controls subsequent execution units to continue sending data.
[0051] The rollback control unit is used to pause subsequent execution units and restore the low-latency leased line to its previous stable configuration state when a new latency causal residual fingerprint fails to converge, or when any of the end-to-end latency, jitter, or packet loss rate exceeds the rollback trigger threshold. jitter deterioration threshold Packet loss rate deterioration threshold and residual fingerprint convergence threshold Among them, the end-to-end latency degradation threshold The unit is milliseconds, preferably 20% to 30% of the current latency margin; jitter deterioration threshold The unit is milliseconds, and the 95th quartile of historical healthy jitter is preferred; packet loss rate deterioration threshold. The value is a dimensionless proportionality between 0 and 1, preferably 50% to 80% of the upper limit of packet loss rate allowed by the service level agreement; residual fingerprint convergence threshold. This is a dimensionless proportional value between 0 and 1, used to determine whether the latency-cause residual fingerprint converges after execution relative to before execution. It is preferably determined based on the tenth percentile of the percentage decrease in residual fingerprints among historical successfully processed samples. For high-priority services, the aforementioned rollback trigger threshold is set to a lower value; for ordinary services, the aforementioned rollback trigger threshold is set to a higher value.
[0052] The previous stable configuration state refers to the network configuration state after the previous execution unit has completed, where end-to-end latency, jitter, packet loss rate, and latency causal residual fingerprint all meet the stability conditions. The stability conditions include end-to-end latency being lower than the service level agreement latency limit, jitter being lower than the 95th percentile of historical healthy jitter, packet loss rate being lower than the service level agreement's allowed packet loss rate limit, and the latency causal residual fingerprint showing a convergence trend relative to before execution.
[0053] In a preferred embodiment, the system establishes a parameter update cycle. For the sampling period, sliding window length, residual threshold, growth slope threshold, coupling threshold, and similarity threshold, the system updates them weekly or monthly based on historical healthy samples and pre-failure samples. For latency risk budget, predicted risk consumption, and rollback trigger threshold, the system recalculates them after each service level agreement change, leased line topology change, primary / backup path adjustment, or customer business priority change. For candidate action ranking weights, the system updates them synchronously when the machine learning model is retrained or the candidate action library is updated. Through this parameter update mechanism, the system can adapt to changes in low-latency leased line topology, traffic distribution, and business level, avoiding long-term use of fixed thresholds that could lead to recognition lag or increased false alarms.
[0054] Through the above implementation methods, this application can identify sub-health degradation trends when low-latency leased lines have not yet defaulted on service level agreements, and verify the risks of candidate operation and maintenance actions through a shadow leased line simulation model. Then, it can make gating judgments through latency risk budgeting, and finally reduce the risk of business shock caused by operation and maintenance actions through phased execution and rollback control, thereby improving the security, interpretability and closed-loop management capabilities of low-latency leased line operation and maintenance.
[0055] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A machine learning-based low-latency leased line operation and maintenance intelligent agent management and control system, characterized in that, The system includes: The multi-source acquisition module is used to collect end-to-end latency, single-hop latency, jitter, packet loss rate, port queue length, link utilization, routing change records, service quality policy change records, protection switching records, maintenance operation logs, and business traffic data of low-latency leased lines. The path time alignment module is used to determine the source node, destination node, intermediate forwarding node, primary path and backup path based on the leased line service identifier, and to map multi-source operation and maintenance data with different sampling periods to the same leased line path and the same time axis to form a leased line operation status matrix. The delay disturbance decomposition module is used to decompose actual delay changes into link-inherent delay, service disturbance delay, control event delay, and abnormal degradation residual delay through a machine learning model. The residual fingerprint generation module is used to generate time-delay causal residual fingerprints based on the time delay of abnormally degraded residuals. The intelligent agent decision-making module is used to identify the sub-healthy state before the default of the service level agreement based on the time delay causal residual fingerprint recognition, and generate candidate operation and maintenance actions. The shadow simulation gating module is used to predict the risk consumption of candidate operation and maintenance actions in the shadow leased line simulation model, and to control whether the candidate operation and maintenance actions enter the phased issuance process based on the latency risk budget. The phased execution module is used to progressively issue, verify, and roll back candidate operation and maintenance actions that have passed the gating.
2. The low-latency leased line operation and maintenance intelligent agent management and control system based on machine learning according to claim 1, characterized in that, The path time alignment module includes: Business path parsing unit, time granularity unification unit, and state matrix construction unit; The service path parsing unit queries the service activation record, routing table and protection relationship based on the leased line service identifier to determine the source node, destination node, intermediate forwarding node, primary path and backup path; The unified time granularity unit uses the monitoring indicator with the highest service level agreement requirement as the benchmark sampling period, sets event markers for event-type records, and performs window aggregation for continuous indicators; The state matrix construction unit organizes and aligns the multi-source operation and maintenance data according to the node dimension, link dimension, business flow dimension and time dimension to generate a dedicated line operation state matrix; the event marker includes event type, occurrence node, start time, end time and affected path number.
3. The low-latency leased line operation and maintenance intelligent agent management and control system based on machine learning according to claim 1, characterized in that, The time delay disturbance decomposition module includes: Model input unit, component output unit, and residual determination unit; The model input unit is used to input the leased line operation status matrix, link length, node forwarding capability, link bandwidth, peak service traffic, port queue length, link utilization, routing switching records, quality of service policy adjustment records, protection switching records, and maintenance operation logs into the machine learning model; The component output unit is used to output the inherent delay of the link, the delay caused by service disturbances, and the delay caused by control events; The residual determination unit is used to determine the difference between the actual delay change and the link inherent delay, service disturbance delay and control event delay as the abnormal degradation residual delay.
4. The low-latency leased line operation and maintenance intelligent agent management and control system based on machine learning according to claim 3, characterized in that: The time delay perturbation decomposition module also includes a training sample construction unit; The training sample construction unit is used to use low-load stable operation samples as link inherent latency training samples, historical samples corresponding to service traffic peaks and port queue backlogs as service disturbance latency training samples, and latency change samples after routing switching, service quality policy adjustment, protection switching and maintenance operations as control event latency training samples. The machine learning model establishes the contribution relationship of different factors to the actual time delay variation based on the above training samples.
5. The low-latency leased line operation and maintenance intelligent agent management and control system based on machine learning according to claim 1, characterized in that, The residual fingerprint generation module includes: Residual feature extraction unit and fingerprint encoding unit; The residual feature extraction unit is used to extract the residual duration, residual growth slope, residual fluctuation density, residual cross-node propagation direction, residual lag correlation coefficient with packet loss rate, and residual lag correlation coefficient with port queue length of abnormally degraded residual delay within a sliding time window. The fingerprint encoding unit is used to combine the above residual features into a time-delay causal residual fingerprint; The direction of residual propagation across nodes is encoded according to the order of nodes where the abnormally degraded residual first exceeds the residual threshold.
6. The low-latency leased line operation and maintenance intelligent agent management and control system based on machine learning according to claim 5, characterized in that: The residual fingerprint generation module also includes a threshold determination unit; The threshold determination unit is used to determine the residual threshold based on the residual distribution of the historical healthy operation phase, and to determine the residual growth slope threshold based on the residual slope distribution of the historical healthy operation phase. Wherein, the residual threshold is taken as the 95th or 99th quartile of the absolute value of the historical health residuals, and the residual growth slope threshold is taken as the 95th quartile of the absolute value of the slope of the historical health residuals. When the business priority is higher than the preset priority level, the residual threshold is taken as the 95th percentile of the absolute value of the historical healthy residual.
7. The low-latency leased line operation and maintenance intelligent agent management and control system based on machine learning according to claim 1, characterized in that, The intelligent agent decision-making module includes: Sub-health identification unit, degradation localization unit, and candidate action generation unit; The sub-health identification unit is used to compare the current latency causal residual fingerprint with the historical health baseline, the residual template of the same type of leased line, and the residual sample before the historical failure through dynamic time normalization distance comparison. When the similarity is higher than the similarity threshold, the abnormal degradation residual latency is higher than the residual threshold, and the residual growth slope is higher than the residual growth slope threshold, the low latency leased line is determined to be in a sub-healthy state before the service level agreement default. The degradation location unit is used to output the degradation location, degradation type, and risk level based on the node where the residual first appears, the direction of residual propagation across nodes, and the coupling index. The candidate action generation unit is used to generate candidate operation and maintenance actions based on the degradation location, degradation type, and risk level.
8. The low-latency leased line operation and maintenance intelligent agent management and control system based on machine learning according to claim 7, characterized in that: The candidate action generation unit is configured with a candidate action library; The candidate action library includes path adjustment actions, backup link switching actions, service quality queue rearrangement actions, rate limiting policy adjustment actions, abnormal node isolation actions, protection failover actions, and observation mode actions. When the degradation location is concentrated in a single intermediate forwarding node and the hysteresis correlation coefficient between the residual and the port queue length is higher than the coupling threshold, the candidate action generation unit prioritizes generating a service quality queue rearrangement action or an abnormal node isolation action. When degradation propagates continuously along the primary path, the candidate action generation unit prioritizes generating backup link switching actions or path adjustment actions.
9. The low-latency leased line operation and maintenance intelligent agent management and control system based on machine learning according to claim 1, characterized in that, The shadow simulation gating module includes: Shadow leased line construction unit, motion simulation unit, risk budget calculation unit, and gating decision unit; The shadow leased line construction unit constructs a shadow leased line simulation model that is synchronously updated with the real leased line, based on the real leased line topology, primary path, backup path, link bandwidth, node forwarding capability, port queue configuration, service quality policy, historical latency curve, and current business traffic distribution. The action simulation unit is used to simulate each candidate operation and maintenance action without changing the actual leased line configuration, and to predict the end-to-end latency, jitter, packet loss rate, routing convergence time and the impact on service continuity after the action is executed. The risk budget calculation unit is used to calculate the latency risk budget based on the service level agreement latency limit, the current end-to-end actual latency, business priority, historical latency fluctuation range, historical jitter percentile, fault recovery time limit and customer business continuity requirements. The gating decision unit is used to allow candidate operation and maintenance actions to enter the phased issuance process when the predicted risk consumption does not exceed the latency risk budget and the simulated end-to-end latency is lower than the service level agreement latency limit.
10. The low-latency leased line operation and maintenance intelligent agent management and control system based on machine learning according to claim 9, characterized in that, The phased execution module includes: Action decomposition unit, feedback verification unit, and rollback control unit; The action splitting unit is used to split the candidate operation and maintenance handling actions through gating into multiple execution units corresponding to single node adjustment, single queue adjustment, partial traffic switching or backup path detection; The feedback verification unit is used to trigger the multi-source acquisition module to re-acquire multi-source operation and maintenance data, trigger the path time alignment module to regenerate the dedicated line operation status matrix, trigger the delay disturbance decomposition module to recalculate the abnormal degradation residual delay, and trigger the residual fingerprint generation module to regenerate the delay causal residual fingerprint after each execution unit is completed. When the new latency causal residual fingerprint converges and the end-to-end latency, jitter, or packet loss rate improves, the feedback verification unit controls the subsequent execution unit to continue sending. When the new latency causal residual fingerprint fails to converge, or when any of the end-to-end latency, jitter, or packet loss rate exceeds the rollback trigger threshold, the rollback control unit suspends subsequent execution units and restores the low-latency leased line to the previous stable configuration state.
