Network configuration optimization method and system based on dynamic multi-level rollback
By using a dynamic multi-level rollback network configuration optimization system, risks are monitored and assessed in real time. Combined with reinforcement learning optimization strategies, the system solves the problems of fixed thresholds and insufficient risk prediction in the configuration change rollback mechanism of 5G networks, and achieves efficient and controllable network configuration management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GANNAN UNIV OF SCI & TECH
- Filing Date
- 2026-01-16
- Publication Date
- 2026-06-02
AI Technical Summary
In existing 5G networks and intelligent operation and maintenance systems, configuration change rollback mechanisms suffer from problems such as fixed thresholds, inability to predict change risks, coarse rollback granularity, high recovery costs, and insufficient learning and optimization capabilities, leading to erroneous rollbacks and service interruptions.
A network configuration optimization system based on dynamic multi-level rollback is adopted, including a configuration change acquisition module, an indicator monitoring module, an anomaly detection module, a risk assessment module, a rollback strategy module, and an optimization learning module. The system optimizes network configuration by combining reinforcement learning algorithms with real-time monitoring, anomaly detection, risk assessment, and multi-level rollback execution modules.
It enables automatic anomaly detection, risk assessment, and multi-level rollback execution, dynamically adjusts rollback strategies, reduces erroneous rollbacks, improves network stability and resource utilization, supports online fine-tuning and anomaly handling, and adapts to different network environments.
Smart Images
Figure CN122137742A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of network configuration optimization technology, and in particular to a network configuration optimization method and system based on dynamic multi-level rollback. Background Technology
[0002] Currently, in 5G networks and intelligent operation and maintenance systems, configuration change rollback mechanisms mainly employ fixed threshold triggering or static rollback strategies. While these solutions can restore the system state in the event of configuration failure, they suffer from several drawbacks, including fixed thresholds, inability to predict change risks, coarse rollback granularity, high recovery costs, and insufficient learning and optimization capabilities. Specifically:
[0003] 1. Fixed Threshold: Existing rollback mechanisms typically set fixed performance thresholds (e.g., packet loss rate > 5%, latency > 100ms triggers rollback), which cannot be dynamically adjusted based on different service loads, time periods, or network environment changes. When the network is under high load but still fluctuating normally, this threshold is easily triggered falsely, leading to false rollbacks or rollback delays.
[0004] 2. Inability to predict change risks: Current systems mostly perform rollbacks only after a change fails, which is a "passive response." There is a lack of risk prediction before or in the early stages of a change, based on abnormal evolution trends and historical experience, to achieve "preventive rollback decisions" based on risk evolution. If a rollback is delayed, it may cause a surge in latency, business interruptions, and other losses.
[0005] 3. Coarse rollback granularity and high recovery cost: Most current systems perform "global rollback," which undoes all configurations. If only some modules malfunction, a global rollback will affect other normal modules, leading to wasted resources and a wider range of business interruptions.
[0006] 4. Insufficient learning and optimization capabilities: Rollback thresholds and strategies often rely on manual experience to set, and cannot learn and optimize themselves based on historical data, leading to the recurrence of similar problems.
[0007] Therefore, there is an urgent need for an intelligent network configuration management system and method that can automatically detect anomalies, assess risks, perform multi-level rollbacks, and automatically optimize. Summary of the Invention
[0008] To address the aforementioned shortcomings, the present invention aims to provide an intelligent network configuration management system and method capable of automatically detecting anomalies, assessing risks, performing multi-level rollbacks, and automatically optimizing.
[0009] To achieve the above objectives, the technical solution provided by this invention is as follows:
[0010] A network configuration optimization system based on dynamic multi-level rollback includes a configuration change acquisition module, an indicator monitoring module, an anomaly detection module, a risk assessment module, a rollback strategy module, a rollback execution module, and an optimization learning module. The configuration change acquisition module records network configuration change information; the indicator monitoring module monitors key performance indicators of network operation in real time; the anomaly detection module analyzes the key performance indicators based on an anomaly detection model, identifies network operation anomalies, and calculates indicator offsets; the risk assessment module performs risk grading assessment based on the offsets to determine the risk level of network operation; the rollback strategy module selects the appropriate rollback level based on the determined risk level; the rollback execution module restores the network configuration at the specified level; and the optimization learning module compares and analyzes the network performance before and after the rollback, and updates the network optimization strategy based on a reinforcement learning algorithm to form a closed-loop optimization.
[0011] A network configuration optimization method based on dynamic multi-level rollback includes the following steps:
[0012] S1: Collect network configuration change information, record the change type, change content, change time and the set of affected network nodes, and store it in the change log database;
[0013] S2: Real-time monitoring of key network performance indicators and normalization of monitoring data;
[0014] S3: Construct an anomaly detection model to detect anomalies in the key performance indicators of the network, and calculate the offset and anomaly confidence of each indicator;
[0015] S4: Based on the comparison between the offset and the preset dynamic risk threshold and the risk classification assessment, determine the risk level of network operation;
[0016] S5: Dynamically determine the rollback level based on the risk level and the impact radius of the configuration change;
[0017] S6: Call the rollback execution module to restore the network configuration at the specified level, and monitor key performance indicators in real time during the rollback process;
[0018] S7: After the rollback is completed, collect network performance data before and after the rollback to analyze the differences in network performance, and update the network optimization strategy based on reinforcement learning algorithm to form a closed-loop optimization.
[0019] As a preferred embodiment of the present invention, in step S1, a configuration change acquisition module continuously collects network configuration change information from the network management system, SDN controller, and automated operation and maintenance scripts, and records the change type, change content, change time, and the set of affected network nodes for each change; the change types include: routing policy adjustment, bandwidth resource reallocation, port operation status modification, and link priority change; the network nodes include switches, routers, base stations, core network equipment, and other terminal nodes.
[0020] As a preferred embodiment of the present invention, in step S2, the indicator monitoring module monitors key network performance indicators in real time through network probes, traffic sampling, network protocol statistics, or log analysis. Normalization is performed uniformly before entering the anomaly detection module to ensure that indicators of different dimensions can be used within a unified model system. The key performance indicators include packet loss rate, round-trip time (RTT), one-way latency, throughput, RRC establishment success rate, UE connection retention rate, cell availability, handover success rate, PRB utilization, uplink and downlink rates, base station CPU utilization, base station memory utilization, uplink BLER, downlink BLER, and backhaul link latency.
[0021] In a preferred embodiment of the present invention, in step S3, an anomaly detection module constructs an anomaly detection model based on the key performance indicators collected in step S2. This anomaly detection model is used to evaluate and calculate the degree of deviation of the indicators, and to identify potential anomalies or performance degradation trends. The anomaly detection module adopts an unsupervised anomaly detection method based on an autoencoder, and establishes a baseline for each network key performance indicator using an exponentially weighted moving average. The standardized observation vector within the sliding window is used as the input of the autoencoder to calculate the reconstruction error. The reconstruction error is normalized to obtain an anomaly score. When the anomaly score exceeds a preset threshold, it is determined to be an anomaly, and the offset and anomaly confidence are output.
[0022] As a preferred embodiment of the present invention, in step S4, a risk classification assessment is performed by the risk assessment module based on the comparison result of the indicator offset and the preset threshold; the risk level includes at least level 1 risk, level 2 risk and level 3 risk, and a dynamic threshold adjustment mechanism is adopted.
[0023] As a preferred embodiment of the present invention, in step S5, the rollback strategy module determines the corresponding rollback level according to the risk level. The rollback level includes local rollback, module rollback and global rollback. The optimal rollback level is dynamically determined based on the impact radius of the configuration change on the network node and the correlation between the change and each module.
[0024] As a preferred embodiment of the present invention, in step S6, the SDN controller uniformly issues configuration rollback instructions, and after the instructions reach the corresponding network node, the network configuration is restored to the state before the change through the rollback execution module. During the rollback execution process, key performance indicators are monitored in real time.
[0025] As a preferred embodiment of the present invention, in step S7, the reinforcement learning algorithm uses the Actor-Critic algorithm for policy gradient training, modeling the network operation and maintenance scenario as a closed-loop RL environment, wherein: the network state includes indicator offset vector, current threshold configuration, current rollback level policy, network topology changes, and historical rollback success rate; adjustable policies include adjusting risk classification threshold, adjusting rollback trigger conditions, adjusting indicator weights, adjusting influence radius threshold, selecting rollback level policy preference, and adjusting anomaly detection model trigger sensitivity; the reward function is automatically generated based on the rollback effect, including the degree of performance recovery after rollback, rollback damage, rollback speed, erroneous rollback, abnormal state after rollback, and overall network stability improvement.
[0026] The beneficial effects of this invention are as follows: This invention can automatically detect anomalies, assess risks, and execute multi-level rollbacks. It not only achieves dynamic risk perception of configuration changes but also, through multi-level rollback strategies and a reinforcement learning-driven optimization module, enables adaptive control and rapid recovery of network configuration risks, ensuring network stability and high reliability. Specifically, this invention has the following advantages:
[0027] (1) Strong dynamic risk perception capability and efficient preventive rollback: identify potential risks in advance before configuration changes or in the early stage of changes, reducing the probability of business interruption and network performance degradation;
[0028] (2) Multi-level rollback strategy with controllable rollback granularity: avoids unnecessary rollback operations, reduces the impact on normal business modules, improves network resource utilization, and shortens recovery time;
[0029] (3) Closed-loop optimization mechanism to achieve adaptive learning: Training samples are generated through each rollback operation to accumulate experience, so that the network configuration management strategy can be continuously improved over time.
[0030] (4) Supports online fine-tuning and anomaly handling to enhance system robustness: automatically adapts to different network environments, service loads and topology changes, reducing the risk of misjudgment and accidental rollback;
[0031] (5) Performance quantification evaluation system to ensure controllable rollback effect: Provide clear reward signals for optimization learning module to achieve a balance between performance recovery, cost control and business impact;
[0032] (6) Wide range of applications: Applicable to 5G / 6G networks, SDN networks, cloud networks, enterprise networks and operator network operation and maintenance scenarios, taking into account both flexibility and intelligence.
[0033] The present invention will be further described below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0034] Figure 1 A flowchart of the system of this invention. Detailed Implementation
[0035] Example: This example provides a network configuration optimization method and system based on dynamic multi-level rollback. The system includes a configuration change acquisition module, an indicator monitoring module, an anomaly detection module, a risk assessment module, a rollback strategy module, a rollback execution module, and an optimization learning module. The configuration change acquisition module records network configuration change information; the indicator monitoring module monitors key performance indicators of network operation in real time; the anomaly detection module analyzes the key performance indicators based on an anomaly detection model, identifies network operation anomalies, and calculates indicator offsets; the risk assessment module performs risk grading assessment based on the offsets to determine the risk level of network operation; the rollback strategy module selects the appropriate rollback level based on the determined risk level; the rollback execution module restores the network configuration at the specified level; and the optimization learning module compares and analyzes the network performance before and after the rollback, and updates the network optimization strategy based on a reinforcement learning algorithm to form a closed-loop optimization.
[0036] The aforementioned configuration change acquisition module, indicator monitoring module, anomaly detection module, risk assessment module, rollback strategy module, rollback execution module, and optimization learning module interact through data channels and control commands to form a complete closed-loop control system, realizing a continuous self-evolving network management mode from "change → detection → risk → rollback → learning → optimization strategy".
[0037] The network configuration optimization system based on dynamic multi-level rollback of this invention can be deployed in network management platforms (NMS), SDN controllers, cloud architecture control planes, or edge orchestration systems.
[0038] When working, see Figure 1 The optimization method of the network configuration optimization system based on dynamic multi-level rollback in this invention is as follows:
[0039] S1: Collect network configuration change information, record the change type, change content, change time and the set of affected network nodes, and store it in the change log database;
[0040] Specifically, the configuration change collection module continuously collects network configuration change information from the Network Management System (NMS), SDN controller, and automated operation and maintenance scripts, recording the change type, change content, change time, and the set of network nodes affected for each change.
[0041] The types of changes include: routing policy adjustment, bandwidth resource reallocation, port operation status modification, link priority change, etc.; the network nodes include switches, routers, base stations, core network equipment and other terminal nodes.
[0042] To ensure the integrity of the changed data, this embodiment stores the above information in a unified change log database, providing basic data for subsequent risk assessment, impact radius calculation, and rollback strategy reasoning.
[0043] S2: Real-time monitoring of key network performance indicators and normalization of monitoring data;
[0044] Specifically, the indicator monitoring module monitors key network performance indicators in real time, including packet loss rate, round-trip time (RTT), one-way latency, throughput, RRC establishment success rate, UE connection retention rate, cell availability, handover success rate, PRB utilization, uplink and downlink speeds, base station CPU utilization, base station memory utilization, uplink BLER, downlink BLER, and backhaul link latency.
[0045] In application, monitoring data can be obtained through network probes, traffic sampling, network protocol statistics, or log analysis. Before entering the anomaly detection module, all data undergoes normalization to ensure that different metrics can be used within a unified model system. The monitoring frequency can be dynamically adjusted according to network scale and business requirements, for example, collecting data every 1 to 5 seconds to ensure timely reflection of network status.
[0046] S3: Construct an anomaly detection model to detect anomalies in the key performance indicators of the network, and calculate the offset and anomaly confidence of each indicator;
[0047] Specifically, an anomaly detection model is constructed based on the key performance indicators collected in step S2 through the anomaly detection module. This anomaly detection model is used to evaluate and calculate the degree of deviation of the indicators and identify potential anomalies or performance degradation trends.
[0048] In this embodiment, the anomaly detection module in the system of the present invention preferably adopts an unsupervised anomaly detection method based on an autoencoder. For each network metric, an exponentially weighted moving average (EWMA) is first used to establish a baseline. And the standardized observation vector within the sliding window As input to the autoencoder. The autoencoder is composed of an encoder. With decoder The training objective is to minimize the regularized mean square reconstruction error. .
[0049] Calculate reconstruction error for online samples Normalized by the mean and standard deviation of the training period error. .when Exceeding the threshold (For example If a value is detected as abnormal, the abnormality confidence level will be output. Offset of each indicator and normalized offset magnitude The above output serves as input to the subsequent risk assessment module, used to determine the rollback strategy and rollback level. This embodiment also preferably supports online fine-tuning (using an incremental learning strategy to fine-tune model parameters through small batches of real-time samples), missing value interpolation (using a time-series-based linear interpolation method to complete missing data), and continuous window determination (using a fixed-length sliding window to confirm the consistency of abnormal results across multiple consecutive windows) to improve robustness.
[0050] S4: Based on the comparison between the offset and the preset dynamic risk threshold and the risk classification assessment, determine the risk level of network operation.
[0051] Specifically, the risk assessment module performs a risk grading assessment based on the comparison results of the indicator offset and the preset threshold; the risk levels include at least Level 1 risk, Level 2 risk, and Level 3 risk. The output offset is based on step S3. The system determines the risk level of each KPI, and its offset is defined as follows:
[0052] in:
[0053] This is the current observation value;
[0054] This is a benchmark or expected value.
[0055] The system presets three risk thresholds for each indicator: .
[0056] The corresponding risk level is determined based on the comparison between the offset and the threshold:
[0057] Level 1 risk: When the offset satisfies: This indicates that the indicator deviates only slightly from the benchmark, and the impact is limited to local nodes or fluctuations in a single indicator.
[0058] Level 2 risk: When the offset satisfies: This indicates that the indicator is moderately abnormal, which may affect some network modules or service areas.
[0059] Level 3 risk: When the offset satisfies: This indicates that the indicator deviates significantly from the baseline, posing a potential risk of overall network performance degradation or widespread service interruption.
[0060] To improve adaptability, this embodiment employs a dynamic threshold adjustment mechanism, updating the threshold based on network type, service characteristics, and historical monitoring data. The dynamic threshold update formula can be defined as:
[0061]
[0062] in:
[0063] The threshold of the previous period ( =1,2,3);
[0064] This represents the historical standard deviation of the indicator.
[0065] This is a smoothing coefficient used to control the update speed.
[0066] Historical fluctuation standard deviation and Under the same metric system, it can be directly used for dynamic threshold updates.
[0067] In applications, preset thresholds can be dynamically adjusted based on different network types, service types, and historical performance data.
[0068] S5: Dynamically determine the rollback level based on the risk level and the impact radius of the configuration change;
[0069] Specifically, the rollback strategy module determines the corresponding rollback level based on the risk level. Rollback levels include partial rollback, module rollback, and global rollback.
[0070] Partial rollback: Rollback the configuration of a portion of nodes or interfaces;
[0071] Module rollback: Rollback is performed on a network subsystem or module;
[0072] Global rollback: Restores the entire network configuration to its stable state before the change.
[0073] In application, the system dynamically determines the optimal rollback level based on the impact radius of the configuration change on network nodes and the correlation between the change and each module.
[0074] Assuming any configuration change The system calculates its radius of influence. and for each network module Relevance ,in:
[0075] In the formula:
[0076] For configuration changes For modules Weighting coefficients;
[0077] For module Propagation distance or dependency depth in the topology;
[0078] This represents the number of related modules.
[0079] Further define the rollback level set as These correspond to partial rollback, module-level rollback, and global rollback, respectively. The system is based on the radius of influence. With preset level threshold , (in Compare and determine the optimal rollback level:
[0080] Partial rollback: ;
[0081] Module-level rollback: ;
[0082] Global rollback: .
[0083] To achieve self-adaptation, this embodiment employs a dynamic adjustment strategy for the threshold, calculating the average influence radius based on historical change data. with standard deviation The threshold update formula is:
[0084]
[0085] in This represents the strategy coefficient.
[0086] S6: Call the rollback execution module to restore the network configuration at the specified level, and monitor key performance indicators in real time during the rollback process;
[0087] In the application, the rollback execution module receives configuration rollback commands uniformly from the SDN controller and automatically executes the rollback operation after the commands reach the corresponding network nodes. Rollback triggering conditions include the risk level reaching a preset threshold or the detection of an abnormal event.
[0088] S7: After the rollback is completed, collect network performance data before and after the rollback to analyze the differences in network performance, and update the network optimization strategy based on reinforcement learning algorithm to form a closed-loop optimization to continuously improve network performance and configuration.
[0089] The optimization learning module writes the learning results back to the network configuration control system, influencing the subsequent network configuration rollback control process by dynamically adjusting risk thresholds, rollback level judgment parameters, and indicator weights. Its output serves as parameter input to the anomaly detection module and the rollback strategy module, guiding network configuration rollback decisions and execution.
[0090] In application, the optimization learning module is used to analyze the network performance differences before and after the rollback operation, and dynamically adjust system parameters based on an adaptive algorithm of reinforcement learning, thereby achieving true closed-loop optimization. Its closed-loop optimization process can be divided into the following core steps:
[0091] (1) Rollback result collection and performance comparison analysis
[0092] The optimization learning module first obtains:
[0093] Performance metrics before rollback: packet loss rate, round-trip time (RTT), one-way end-to-end latency, throughput, RRC establishment success rate, UE connection retention rate, cell availability, handover success rate, PRB utilization, uplink rate, downlink rate, base station CPU utilization, base station memory utilization, uplink BLER, downlink BLER, and backhaul link latency;
[0094] Performance metrics after rollback: The recovery status of the same set of metrics after rollback;
[0095] Rollback execution process data: execution time of a single rollback operation, number of rollback execution failures, number of abnormal events triggered by rollback, network performance recovery time after rollback, recovery extent of key performance indicators (KPIs) after rollback, performance difference measurement before and after rollback, rollback strategy hit rate (i.e., whether the selected rollback level matches the actual risk level), and the impact of rollback operation on business interruption duration.
[0096] Module calculation:
[0097] Performance recovery rate: ;
[0098] Overall Performance Recovery Ratio (Multi-metric PRR): ;
[0099] in The weights of the indicators are set by the risk assessment module or manually. .
[0100] Rollback proceeds: ;
[0101] Rollback cost: ;
[0102] Rollback effect score: ;
[0103] The result serves as the model's reward signal, which is then input into the subsequent learning process.
[0104] (2) Construct a State, Action, and Reward model
[0105] The optimization learning module models network operation and maintenance scenarios as a closed-loop RL environment:
[0106] ① Network State
[0107] Includes: metric offset vector, current threshold configuration, current rollback level policy, network topology changes, and historical rollback success rate.
[0108] Example:
[0109]
[0110] ② Adjustable strategies (Actions)
[0111] This includes: adjusting risk classification thresholds (raising / lowering level 1, 2, and 3 thresholds), adjusting rollback trigger conditions (delayed triggering / early triggering), adjusting the weights of different indicators, adjusting the influence radius threshold of the decision tree, selecting different rollback level strategy preferences (prioritizing local / prioritizing modules), and adjusting the trigger sensitivity of the anomaly detection model.
[0112] ③ Reward function (Reward)
[0113] Automatically generated based on the rollback effect, including: the degree of performance recovery after rollback (positive reward), rollback damage (e.g., unnecessary rollback → negative reward), fast rollback speed (positive reward), erroneous rollback (strong negative reward), still abnormal after rollback (negative reward), and improved overall network stability (positive reward).
[0114] (3) Strengthen the learning and training process
[0115] The optimization learning module gradually develops the optimal strategy through continuous training:
[0116] Use Actor-Critic to implement policy gradient training;
[0117] Historical data and real-time data are used together as training samples;
[0118] Each rollback provides a new source of training samples;
[0119] Enable the system to learn automatically:
[0120] A certain abnormal offset is more suitable for triggering a module rollback;
[0121] The threshold should be automatically increased during certain time periods (such as peak hours);
[0122] Some configuration changes were dangerous in the past network environment and should be rolled back preventively.
[0123] Some nodes have a particularly large radius of influence and should be given higher priority.
[0124] Ultimately, the strategy achieves self-optimization.
[0125] (4) Strategy update and closed-loop feedback
[0126] After training is complete, the optimization learning module will update the results and write them back to the system:
[0127] Update the trigger threshold of the anomaly detection module;
[0128] Update the risk classification thresholds;
[0129] Optimize the rules of the rollback strategy module;
[0130] Adjust the execution granularity of the rollback execution module (whether it is more inclined towards local or module-wide execution).
[0131] Update the decision tree parameters (influence radius, weight).
[0132] The update logic is as follows:
[0133] Learning outcome → New threshold / new strategy → Take effect → New monitoring data → New round of training, achieving a true "closed loop";
[0134] Data → Detection → Risk → Rollback → Learning → Parameter Adjustment → New Data → Re-detection... This continuous cycle makes the system more accurate and stable with continued use.
[0135] (5) Long-term effects
[0136] Through the aforementioned closed-loop mechanism, the system possesses the following capabilities:
[0137] Automatically identify risk patterns for different types of changes;
[0138] Automatically adapts to network load, topology, and service fluctuations;
[0139] Automatically reduces erroneous rollbacks;
[0140] Automatically anticipate risks and roll back in advance;
[0141] Automatic parameter adjustment reduces manual adjustments and improves stability.
[0142] The system has been upgraded from "human experience-driven" to data-driven + adaptive learning-driven.
[0143] By adopting the technology of this invention, the intelligence, adaptability, and reliability of network configuration management can be significantly improved, with the following specific advantages:
[0144] (1) Strong dynamic risk perception capability and efficient preventive rollback: identify potential risks in advance before configuration changes or in the early stage of changes, realize preventive rollback, and reduce the probability of business interruption and network performance degradation.
[0145] (2) Multi-level rollback strategy with controllable rollback granularity: It can effectively avoid unnecessary rollback operations, reduce the impact on normal business modules, improve network resource utilization, and shorten recovery time.
[0146] (3) Closed-loop optimization mechanism to achieve adaptive learning: The system can generate training samples through each rollback operation, gradually accumulate experience, realize "data-driven + self-learning" closed-loop optimization, and make the network configuration management strategy continuously improve with the use time, thereby improving network stability and operating efficiency.
[0147] (4) Supports online fine-tuning and anomaly handling to enhance system robustness: The system can automatically adapt to different network environments, service loads and topology changes, reducing the risk of misjudgment and rollback.
[0148] (5) Performance quantification evaluation system to ensure controllable rollback effect: The system performs quantitative evaluation of rollback operations. This evaluation system can provide clear reward signals for the optimization learning module, ensuring that the rollback strategy achieves the best balance between performance recovery, cost control and business impact.
[0149] (6) Wide range of applications: This invention is applicable to 5G / 6G networks, SDN networks, cloud networks, enterprise networks and operator network operation and maintenance scenarios. It can maintain stable operation in high-frequency configuration changes and complex network environments, taking into account both flexibility and intelligence, and realizing automated upgrades of network operation and maintenance.
[0150] Based on the disclosure and teachings of the above specification, those skilled in the art can make changes and modifications to the above embodiments. Therefore, the present invention is not limited to the specific embodiments disclosed and described above, and some modifications and changes to the present invention should also fall within the protection scope of the claims of the present invention. Furthermore, although some specific terms are used in this specification, these terms are only for convenience of explanation and do not constitute any limitation on the present invention. Other methods and systems obtained by using the same or similar methods are all within the protection scope of the present invention.
Claims
1. A network configuration optimization system based on dynamic multi-level rollback, characterized in that... It includes: The configuration change collection module is used to record network configuration change information; The performance indicator monitoring module is used to monitor key performance indicators of network operation in real time. An anomaly detection module is used to analyze the key performance indicators based on the anomaly detection model, identify network operation anomalies, and calculate indicator offsets. The risk assessment module is used to perform risk classification assessment based on the offset to determine the risk level of network operation; The rollback strategy module is used to select the appropriate rollback level based on the determined risk level. The rollback execution module is used to restore the network configuration at a specified level. The optimization learning module is used to compare and analyze the network performance before and after the rollback, and update the network optimization strategy based on the reinforcement learning algorithm to form a closed-loop optimization.
2. A network configuration optimization method based on dynamic multi-level rollback, characterized in that, It includes the following steps: S1: Collect network configuration change information, record the change type, change content, change time and the set of affected network nodes, and store it in the change log database; S2: Real-time monitoring of key network performance indicators and normalization of monitoring data; S3: Construct an anomaly detection model to detect anomalies in the key performance indicators of the network, and calculate the offset and anomaly confidence of each indicator; S4: Based on the comparison between the offset and the preset dynamic risk threshold and the risk classification assessment, determine the risk level of network operation; S5: Dynamically determine the rollback level based on the risk level and the impact radius of the configuration change; S6: Call the rollback execution module to restore the network configuration at the specified level, and monitor key performance indicators in real time during the rollback process; S7: After the rollback is completed, collect network performance data before and after the rollback to analyze the differences in network performance, and update the network optimization strategy based on reinforcement learning algorithm to form a closed-loop optimization.
3. The network configuration optimization method based on dynamic multi-level rollback according to claim 2, characterized in that: In step S1, the configuration change collection module continuously collects network configuration change information from the network management system, SDN controller, and automated operation and maintenance scripts, and records the change type, change content, change time, and set of affected network nodes for each change. The types of changes include: routing policy adjustment, bandwidth resource reallocation, port operation status modification, and link priority change; The network nodes include switches, routers, base stations, core network equipment, and other terminal nodes.
4. The network configuration optimization method based on dynamic multi-level rollback according to claim 3, characterized in that: In step S2, the indicator monitoring module monitors key network performance indicators in real time through network probes, traffic sampling, network protocol statistics, or log analysis. Before entering the anomaly detection module, the indicators are uniformly normalized to ensure that indicators of different dimensions can be used within a unified model system. The key performance indicators include packet loss rate, round-trip time (RTT), one-way latency, throughput, RRC establishment success rate, UE connection retention rate, cell availability, handover success rate, PRB utilization, uplink and downlink rates, base station CPU utilization, base station memory utilization, uplink BLER, downlink BLER, and backhaul link latency.
5. The network configuration optimization method based on dynamic multi-level rollback according to claim 4, characterized in that: In step S3, an anomaly detection model is constructed based on the key performance indicators collected in step S2 by the anomaly detection module. The anomaly detection model is used to evaluate and calculate the degree of deviation of the indicators and identify potential anomalies or performance degradation trends. The anomaly detection module adopts an unsupervised anomaly detection method based on an autoencoder, and uses an exponentially weighted moving average to establish a baseline for each key network performance indicator. The standardized observation vector within the sliding window is used as the input of the autoencoder to calculate the reconstruction error. The reconstruction error is normalized to obtain the anomaly score. When the anomaly score exceeds a preset threshold, it is determined to be an anomaly, and the offset and anomaly confidence are output.
6. The network configuration optimization method based on dynamic multi-level rollback according to claim 5, characterized in that: In step S4, the risk assessment module performs a risk grading assessment based on the comparison results between the indicator offset and the preset threshold. The risk levels include at least Level 1 risk, Level 2 risk, and Level 3 risk, and a dynamic threshold adjustment mechanism is adopted.
7. The network configuration optimization method based on dynamic multi-level rollback according to claim 6, characterized in that: In step S5, the rollback strategy module determines the corresponding rollback level based on the risk level. Rollback levels include partial rollback, module rollback, and global rollback. The optimal rollback level is dynamically determined based on the impact radius of the configuration change on network nodes and the correlation between the change and each module.
8. The network configuration optimization method based on dynamic multi-level rollback according to claim 7, characterized in that: In step S6, the SDN controller uniformly issues configuration rollback instructions, and after the instructions reach the corresponding network nodes, the rollback execution module restores the network configuration to the state before the change. During the rollback execution process, key performance indicators are monitored in real time.
9. The network configuration optimization method based on dynamic multi-level rollback according to claim 8, characterized in that: In step S7, the reinforcement learning algorithm uses the Actor-Critic algorithm for policy gradient training, modeling the network operation and maintenance scenario as a closed-loop RL environment, wherein: Network status includes metric offset vector, current threshold configuration, current rollback level policy, network topology changes, and historical rollback success rate; Adjustable strategies include adjusting risk classification thresholds, adjusting rollback trigger conditions, adjusting indicator weights, adjusting impact radius thresholds, selecting rollback level strategy preferences, and adjusting anomaly detection model trigger sensitivity. The reward function is automatically generated based on the rollback effect, including the degree of performance recovery after rollback, rollback damage, rollback speed, erroneous rollback, abnormal states after rollback, and overall network stability improvement.