Network topology adaptive optimization method and system

Abstract

This disclosure provides a network topology adaptive optimization method and system. The method includes: collecting network node status and link load data, and performing data standardization, noise filtering, and time-series alignment; constructing a dynamic network topology model based on a Stancu-type neural network operator to capture the dynamic relationships and topological features between nodes; actively detecting network weaknesses and potential fault points using a perturbation sampling strategy, and conducting a comprehensive risk assessment based on service importance; generating candidate topology adjustment schemes using a multi-objective optimization algorithm, predicting the performance of each scheme under different load conditions using a network simulation engine, and obtaining the optimal scheme after evaluation; using a two-phase commit protocol and a distributed coordination mechanism to achieve synchronous execution of configuration changes among multiple nodes, completing topology reconstruction and resource scheduling; deploying real-time monitoring points to verify the optimization effect, and continuously optimizing the decision model based on actual results. This embodiment can safely and reliably execute network changes.

Description

Technical Field

[0001] This disclosure relates to the field of computer network technology, specifically to a network topology adaptive optimization method and system. Background Technology

[0002] Network topology optimization is a key technology in computer networks and communication systems, focusing on how to effectively organize network nodes and links to improve network performance, reliability, and resource utilization. As network size and complexity increase, network topology optimization techniques become particularly important and are crucial for enhancing overall network performance.

[0003] Traditional network topology optimization methods mainly fall into two categories: static planning and simple dynamic adjustment. Static planning methods typically design network topologies using linear programming or heuristic algorithms based on predefined network traffic models and constraints. Simple dynamic adjustment methods, on the other hand, rely on predefined rules or thresholds to trigger topology changes, such as routing adjustment mechanisms based on load balancing algorithms and failover techniques based on redundant links. These methods are effective when the network environment is relatively stable and changes are predictable. In recent years, machine learning-based network topology optimization techniques have gradually emerged. These techniques predict traffic changes and identify potential faults by building network behavior models, thereby guiding topology adjustments. These methods use standard neural network structures for traffic prediction and topology generation, employ reinforcement learning methods to optimize routing strategies, and trigger corresponding topology adjustment operations when network conditions change.

[0004] However, existing machine learning optimization methods suffer from two significant technical drawbacks: first, standard neural network architectures are insufficiently adaptable to multivariate and multi-objective optimization problems in network topology, resulting in unsatisfactory prediction accuracy and convergence speed in complex network environments; second, the lack of effective sampling mechanisms to capture abrupt changes and abnormal patterns in network states causes system response lags when facing network fluctuations and failures, making it difficult to achieve true millisecond-level topology reconstruction and resource scheduling. These problems severely restrict the application effectiveness of network topology optimization technology in large-scale, highly dynamic network environments. Summary of the Invention

[0005] To address the problems existing in the prior art, this disclosure provides a network topology adaptive optimization method and system. This disclosure introduces a Stancu type neural network operator to construct a dynamic network topology model, combines a perturbation sampling strategy to actively detect network weaknesses, and adopts a distributed collaborative mechanism to achieve millisecond-level topology reconstruction, thereby improving the accuracy, speed, and reliability of network topology optimization.

[0006] To achieve the above objectives, the present disclosure provides the following technical solutions: In a first aspect, embodiments of this disclosure provide a network topology adaptive optimization method, including: Based on distributed network probes, network node status data and link load data are acquired through a multi-level acquisition strategy. The network node status data and link load data are then processed by data standardization, noise filtering and time-series alignment to obtain a standardized network status data stream. The standardized network state data stream is received, and a dynamic network topology model is constructed using a Stancu-type neural network operator. The Stancu-type neural network operator includes adjustable higher-order terms and nonlinear transformations. The dynamic network topology model is used to capture the dynamic relationships and topological features between network nodes, and to obtain network topology feature vectors and relationship matrices. Based on the network topology feature vector and relation matrix, network weaknesses and potential fault points are actively detected through a perturbation sampling strategy, network bottlenecks and abnormal patterns are identified, a comprehensive risk assessment is conducted, and a network risk assessment report and optimization target list are obtained. Based on the network risk assessment report and the optimization target list, a set of candidate topology adjustment schemes is generated using a multi-objective optimization algorithm. The optimal topology adjustment scheme is obtained by comprehensively evaluating each candidate topology adjustment scheme. The system receives the optimal topology adjustment scheme, converts it into a network topology change instruction sequence, and uses a distributed coordination mechanism to achieve synchronous execution of configuration changes among multiple nodes. This completes the distribution and execution of topology reconstruction instructions, while simultaneously performing dynamic resource scheduling and route updates. The system obtains topology adjustment results and performance monitoring data, and continuously optimizes the decision model based on these results and data.

[0007] As one implementation method, the method based on distributed network probes acquires network node status data and link load data through a multi-level acquisition strategy. The network node status data and link load data are then processed for data standardization, noise filtering, and time-series alignment to obtain a standardized network status data stream, including: Based on the network topology and node distribution, intelligent algorithms are used to determine the probe deployment locations and deploy monitoring agents on network nodes at different levels to obtain a probe deployment scheme that covers the entire network. Based on the probe deployment scheme, differentiated sampling frequencies and monitoring granularities are set, and the sampling frequency of each node is determined according to the node load characteristics to obtain multi-level network raw monitoring data. The system receives the original monitoring data of the multi-level network and processes it through unit conversion, dimension normalization and feature scaling to unify the network data from different sources and of different types into a standard format, thereby obtaining a standardized network indicator dataset. Based on the standardized network index dataset, noise and outliers are removed by applying sliding window midpoint filtering and wavelet transform methods to obtain cleaned network state data. Based on the cleaned network status data, time alignment is achieved through timestamp correction and interpolation algorithms. Multi-source data fusion technology is used to correlate and analyze the node status and link load data in the cleaned network status data to obtain the standardized network status data stream.

[0008] As one implementation, the standardized network state data stream is received, and a dynamic network topology model is constructed using a Stancu-type neural network operator. This Stancu-type neural network operator includes adjustable higher-order terms and nonlinear transformations. The dynamic network topology model captures the dynamic relationships and topological features between network nodes, resulting in a network topology feature vector and a relationship matrix, including: Based on the requirements of network topology optimization, a mathematical model of a Stancu-type neural network operator is constructed. The adjustable higher-order terms and nonlinear transformations are introduced to obtain the framework of a Stancu-type neural network operator. Based on the Stancu type neural network operator framework and the standardized network state data stream, a graph structure representation of the network topology is constructed. The embedding vectors of nodes and links in the graph structure representation are learned by the neural network to obtain a low-dimensional feature representation of the network elements. Based on the low-dimensional feature representation of the network elements, a hybrid model of temporal convolutional network and long short-term memory network is applied to capture the temporal change pattern of the network state in the low-dimensional feature representation of the network elements, thereby obtaining the network dynamic temporal features. Based on the dynamic temporal characteristics of the network, the dynamic association strength and influence propagation mode between nodes in the dynamic temporal characteristics of the network are learned through graph attention mechanism and relation reasoning network to obtain the network node relation matrix. Based on the network node relationship matrix, the low-dimensional feature representation of the network elements, and the network dynamic temporal features, a multi-level fusion network is used to integrate the static topology structure and dynamic state features. The computational efficiency is optimized by applying model distillation technology to obtain the network topology feature vector and relationship matrix.

[0009] As one implementation method, based on the network topology feature vector and relationship matrix, a perturbation sampling strategy is used to actively detect network weaknesses and potential fault points, identify network bottlenecks and abnormal patterns, conduct a comprehensive risk assessment, and obtain a network risk assessment report and an optimization target list, including: Based on the network topology feature vector and relation matrix, a multi-level progressive perturbation sampling strategy is designed, which includes micro-perturbation, moderate perturbation and strong perturbation, to obtain the network perturbation sampling execution plan; Based on the network disturbance sampling execution plan, disturbance injection is executed collaboratively by a distributed controller. The network response data to the disturbance is collected in real time during the execution of the network disturbance sampling execution plan, and the state changes of each node and link are recorded to obtain the disturbance response dataset. Based on the disturbance response dataset and the network topology feature vector and relation matrix, gradient analysis and influence propagation model are used to identify the network nodes and links in the disturbance response dataset that are sensitive to disturbances, assess the vulnerability of the network nodes and links that are sensitive to disturbances, and obtain a network sensitivity map. Based on the network sensitive point map and the standardized network status data stream, clustering algorithms and anomaly detection models are applied to detect and classify network bottlenecks in the network sensitive point map, distinguishing between capacity bottlenecks, configuration bottlenecks, and topology bottlenecks, and obtaining a network bottleneck analysis report. Based on the network bottleneck analysis report and the network sensitive point map, a comprehensive risk assessment is conducted in conjunction with business importance and service level agreement requirements. Priorities are assigned to the risk points identified in the network bottleneck analysis report and the network sensitive point map, and corresponding optimization targets are generated, resulting in the network risk assessment report and the optimization target list.

[0010] As one implementation method, the step of generating a candidate topology adjustment scheme set using a multi-objective optimization algorithm based on the network risk assessment report and the optimization target list, comprehensively evaluating each candidate topology adjustment scheme, and obtaining the optimal topology adjustment scheme includes: Based on the network risk assessment report and the list of optimization objectives, the optimization objectives in the list of optimization objectives are transformed into mathematical objective functions and weight coefficients are set for each objective to obtain a formalized definition of a multi-objective optimization problem. Based on the formalized definition of the multi-objective optimization problem and the network topology feature vector and relation matrix, network physical constraints, service constraints and resource constraints are identified, and the network physical constraints, service constraints and resource constraints are transformed into mathematical constraints to obtain the optimization problem constraint model; Based on the optimization problem constraint model and the formalized multi-objective optimization problem definition, a multi-objective evolutionary algorithm and Pareto optimal front technique are applied to generate topology adjustment schemes for different combinations of objectives in the formalized multi-objective optimization problem definition, thereby obtaining the candidate topology adjustment scheme set. Based on the candidate topology adjustment scheme set, each scheme in the candidate topology adjustment scheme set is simulated and tested using a network simulation engine. The performance indicators of each scheme under different load conditions are predicted and potential risks and benefits are evaluated to obtain a scheme performance prediction report. Based on the performance prediction report of the proposed scheme and the weight coefficients in the formalized multi-objective optimization problem definition, a multi-criteria decision analysis method is applied to comprehensively evaluate each candidate topology adjustment scheme in the candidate topology adjustment scheme set and select the best balance point, generating a detailed implementation plan to obtain the optimal topology adjustment scheme.

[0011] As one implementation method, the process involves receiving the optimal topology adjustment scheme, converting it into a network topology change instruction sequence, and using a distributed coordination mechanism to achieve synchronous execution of configuration changes among multiple nodes. This completes the distribution and execution of topology reconstruction instructions, while simultaneously performing dynamic resource scheduling and route updates, resulting in topology adjustment results and performance monitoring data. Based on the optimal topology adjustment scheme, the optimization decisions in the optimal topology adjustment scheme are converted into a network configuration instruction set to ensure the atomicity and consistency of the instructions in the network configuration instruction set, thereby obtaining the network topology change instruction sequence; Based on the network topology change instruction sequence, a distributed transaction management framework is constructed. A two-phase commit protocol is used to ensure the synchronization of configuration changes among multiple nodes, and a rollback mechanism is designed to deal with the situation of node execution failure, thus obtaining a collaborative execution plan. According to the collaborative execution plan, the instructions in the network topology change instruction sequence are rapidly distributed and executed through optimized message queues and parallel processing mechanisms, and the change risk is controlled by time window technology to obtain execution status feedback data. Based on the execution status feedback data and the network topology change instruction sequence, dynamic scheduling of computing, storage and network resources is performed, and a load balancing algorithm is applied to optimize traffic allocation to obtain a resource scheduling result report. Based on the resource scheduling result report and the execution status feedback data, real-time monitoring points are deployed to collect network performance data after the topology change. The data is then compared and analyzed with the performance indicators predicted in the scheme performance prediction report to verify the optimization effect, thereby obtaining the topology adjustment results and performance monitoring data.

[0012] As one implementation method, determining the sampling frequency of each node based on node load characteristics includes: Based on the node types in the probe deployment scheme, nodes are divided into a first type of node and a second type of node according to their hierarchical position in the network topology and their traffic carrying capacity. The first type of node is a high-load node carrying backbone traffic, and the second type of node is a low-load node carrying access traffic. The sampling frequency for the first type of node is set to the millisecond level, and the sampling frequency for the second type of node is dynamically adjusted to the second level to the minute level according to the changes in node load, thereby obtaining the original monitoring data of the multi-layer network.

[0013] As one implementation method, the step of learning the dynamic association strength and influence propagation pattern between nodes in the dynamic temporal features of the network through graph attention mechanism and relational reasoning network to obtain the network node relation matrix includes: Based on the dynamic temporal characteristics of the network, the attention weights between nodes are calculated using a graph attention mechanism to obtain the node attention weight matrix. Based on the node attention weight matrix, the influence propagation path between nodes is modeled through a relational reasoning network to obtain the network node relation matrix containing dynamic association strength and influence propagation mode.

[0014] As one implementation method, the design rollback mechanism addresses node execution failures and generates a collaborative execution plan, including: Based on the two-phase commit protocol, a node execution status monitoring mechanism is set up. When a node execution failure is detected, a distributed rollback instruction is triggered to restore the node that has undergone changes to its state before the change, thereby obtaining a system state consistency guarantee result. The system state consistency guarantee result is then integrated into the collaborative execution plan.

[0015] Secondly, embodiments of this disclosure also provide a network topology adaptive optimization system, comprising: The network status acquisition module is used to acquire network node status data and link load data based on distributed network probes and through a multi-level acquisition strategy. The network node status data and link load data are then processed for data standardization, noise filtering and time-series alignment to obtain a standardized network status data stream. The dynamic topology modeling module is used to receive the standardized network state data stream, apply a Stancu type neural network operator to construct a dynamic network topology model, the Stancu type neural network operator includes adjustable higher-order terms and nonlinear transformations, and capture the dynamic relationships and topological features between network nodes through the dynamic network topology model to obtain network topology feature vectors and relationship matrices. An anomaly identification module is used to actively detect network weaknesses and potential fault points based on the network topology feature vector and relation matrix, identify network bottlenecks and anomaly patterns, conduct a comprehensive risk assessment, and obtain a network risk assessment report and an optimization target list. The optimization decision module is used to generate a set of candidate topology adjustment schemes based on the network risk assessment report and the optimization target list, and to comprehensively evaluate each candidate topology adjustment scheme to obtain the optimal topology adjustment scheme. The topology reconfiguration module is used to receive the optimal topology adjustment scheme, convert the optimal topology adjustment scheme into a network topology change instruction sequence, realize the synchronous execution of configuration changes among multiple nodes through a distributed coordination mechanism, complete the distribution and execution of topology reconfiguration instructions, and simultaneously perform dynamic resource scheduling and route updates to obtain topology adjustment results and performance monitoring data, and continuously optimize the decision model based on the topology adjustment results and performance monitoring data.

[0016] Compared with the prior art, the embodiments of this disclosure have at least the following beneficial effects: This disclosure introduces a Stancu-type neural network operator to model the dynamic characteristics of network topology. Compared to traditional neural network structures, it includes adjustable higher-order terms and nonlinear transformations, enabling it to more accurately capture the complex relationships between network nodes and improve prediction accuracy and optimization efficiency. Simultaneously, a multi-level progressive perturbation sampling strategy is designed to actively detect network weaknesses and potential fault points, solving the problem that traditional passive monitoring methods struggle to effectively detect potential risks and identifying network bottlenecks and abnormal patterns in advance. Specifically, by employing a distributed collaborative mechanism and a two-phase commit protocol, network topology reconstruction and resource scheduling are ensured to be completed within milliseconds, while guaranteeing the synchronicity and consistency of multi-node configuration changes, significantly improving the response speed and reliability of network topology adjustments. Furthermore, by combining multi-objective optimization and multi-criteria decision analysis methods, it can automatically balance multiple optimization objectives such as throughput, latency, and reliability under different network conditions, generating the optimal topology adjustment scheme and improving network resource utilization efficiency. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the network topology adaptive optimization method provided in this embodiment; Figure 2 This is a block diagram of the network topology adaptive optimization system provided in this embodiment. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the implementation methods of the embodiments of this disclosure will be further described in detail below with reference to the accompanying drawings.

[0020] Example 1: Figure 1 A flowchart of the network topology adaptive optimization method of this embodiment is shown, as follows: Figure 1 As shown, this embodiment provides a network topology adaptive optimization method, which can be executed by a computer system. Specifically, the method includes the following steps: Step S1: Based on distributed network probes, network node status data and link load data are acquired through a multi-level acquisition strategy. The network node status data and link load data are then processed for data standardization, noise filtering, and time-series alignment to obtain a standardized network status data stream.

[0021] This step begins by deploying a distributed probe system at key network locations to build a comprehensive network monitoring system. Specifically, based on the characteristics of the network topology, monitoring agents are deployed on different types of network devices, such as core switching nodes, edge routing nodes, and critical business server nodes. These monitoring points form a probe network covering the entire network, capable of simultaneously collecting network operational status information from multiple perspectives.

[0022] Typically, the selection of deployment locations is based on graph theory-based probe coverage optimization algorithms to ensure effective monitoring of the entire network using the minimum number of probes.

[0023] The probe system employs a multi-level acquisition strategy, applying differentiated sampling frequencies and monitoring granularities to different types of nodes. For example, high-load nodes such as core switches and critical routing nodes use high-frequency sampling at the millisecond level (typically 5-50 milliseconds) to accurately capture transient network fluctuations; while low-load nodes such as edge nodes and terminal access devices use an adaptive sampling strategy, with sampling intervals dynamically adjustable from the second level to the minute level.

[0024] Typically, adaptive sampling adjusts the sampling frequency based on changes in node load. When a sudden increase in traffic or an abnormal pattern is detected, the sampling frequency is automatically increased; when the network status is stable, the sampling frequency is decreased, thereby balancing monitoring accuracy and system resource consumption.

[0025] In step S1, the collected raw data mainly includes network node status data and link load data. Specifically, node status data includes indicators such as CPU utilization, memory usage, queue depth, and forwarding table size, reflecting the device's processing capacity and resource status; link load data includes indicators such as traffic throughput, bandwidth utilization, latency, packet loss rate, and error packet statistics, reflecting connection quality and transmission efficiency.

[0026] Those skilled in the art will recognize that these raw data come from equipment from different manufacturers, have different formats, and inconsistent units of measurement, and therefore require standardization.

[0027] The data standardization process begins with unit conversion and dimension normalization, unifying data from different sources and units into a standard format, such as converting all traffic data to bits per second and latency data to milliseconds. Then, the min-max normalization method maps various indicators to the [0,1] interval, eliminating dimensional differences. For unevenly distributed data, feature scaling techniques such as logarithmic transformation or Z-score normalization are applied to make the data distribution more uniform.

[0028] Typically, standardized data still contains various noises and occasional outliers. This embodiment uses sliding window midpoint filtering and wavelet transform for data cleaning. Sliding window midpoint filtering, as a nonlinear filtering technique, effectively removes impulse noise. By setting an appropriate window size (usually 5-15 sampling points), the median value within the window is used to replace the center point value, effectively removing short-term interference. For complex noise patterns, discrete wavelet transform is used for multi-scale analysis to identify and filter high-frequency noise while preserving the main features of the data.

[0029] Because the data comes from distributed probes, there may be slight differences in the internal clocks of different devices, resulting in incomplete synchronization of the timestamps of the collected data. This embodiment achieves timing alignment through timestamp correction and interpolation algorithms. First, the system synchronizes the clocks of all probes with a standard network time server, calculates the time offset, and adjusts the data timestamps. For data with different sampling frequencies, the system applies a cubic spline interpolation algorithm to generate data values ​​at a unified time point.

[0030] Finally, multi-source data fusion technology is applied to correlate node status data with link load data, establishing a network topology graph model. This maps various data types to corresponding nodes and links, and graph analysis algorithms are used to uncover the causal relationships and mutual influences between node status and link load. This process yields a standardized network status data stream, providing a high-quality data foundation for subsequent analysis.

[0031] Step S2: Receive the standardized network state data stream, and construct a dynamic network topology model using a Stancu type neural network operator. The Stancu type neural network operator includes adjustable higher-order terms and nonlinear transformations. The dynamic network topology model captures the dynamic relationships and topological features between network nodes, and obtains the network topology feature vector and relationship matrix.

[0032] The core of step S2 is to apply Stancu-type neural network operators to construct a dynamic network topology model, which is used to capture the dynamic relationships in complex networks. Specifically, Stancu-type neural network operators are an extension of traditional neural networks. By introducing adjustable higher-order terms and nonlinear transformations, the expressive power of the model is enhanced, enabling it to more accurately model the complex nonlinear relationships in network topology.

[0033] In this step, a mathematical model framework for a Stancu-type neural network operator is first constructed based on the network topology optimization requirements. This framework, based on Stancu polynomial interpolation theory, extends the activation function of traditional neural networks by introducing higher-order polynomial terms and parameterized nonlinear transformations. Unlike traditional neural networks that use fixed activation functions, each activation unit of a Stancu-type neural network operator contains multiple adjustable parameters, enabling adaptive adjustment of its nonlinearity and higher-order characteristics, thus more accurately modeling complex network topology relationships. These parameters are optimized along with other weights during training, allowing the model to automatically adapt to the characteristics of different network environments.

[0034] Next, based on the Stancu-type neural network operator framework, a graph representation of the network topology is constructed, and the embedding vectors of nodes and links are learned. First, the physical topology of the network is represented as a graph structure, with nodes representing network devices and edges representing physical links, and standardized network state data is mapped onto this graph structure. Then, graph neural network techniques are applied, particularly a combined architecture of Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT), to learn low-dimensional embedding representations of network elements.

[0035] In this process, considering the characteristics of the node itself, the characteristics of neighboring nodes, and the attributes of the links, a low-dimensional vector representation that can represent the topological location and functional characteristics of network elements is extracted through multi-layer graph convolution operations and attention mechanisms. It is usually 64-256 dimensions.

[0036] Next, a hybrid model combining Temporal Convolutional Network (TCN) and Long Short-Term Memory (LSTM) is applied to capture the temporal change patterns of the network state. Low-dimensional feature representations are organized into a temporal sequence, and a hybrid temporal model is used for modeling. This model comprises two main components: a Temporal Convolutional Network and a LSTM network. The Temporal Convolutional Network captures changing features at different time scales through multiple layers of dilated convolutions; its receptive field grows exponentially with the number of layers, effectively capturing long-term dependencies. The LSTM network focuses on modeling long-term dependencies and sequential memory, retaining important information and filtering out irrelevant information through gating mechanisms. The outputs of the two models are fused through an attention mechanism to form a comprehensive temporal feature representation.

[0037] Next, based on the dynamic temporal characteristics of the network, the system learns the dynamic association strength and influence propagation patterns between nodes through graph attention mechanisms and relational reasoning networks. The system employs a multi-head graph attention mechanism to calculate the attention weights between nodes, with each attention head learning the association patterns between nodes from different perspectives.

[0038] Specifically, attention scores between node pairs are calculated, representing the degree of influence of a node's state change on another node, and then normalized using a softmax function to form an attention weight matrix. Next, a relational reasoning network is applied, and the influence propagation path between nodes is modeled based on the attention weight matrix. A message-passing mechanism is used to simulate the dynamic process of information flow and influence propagation in the network. Through multiple rounds of iterative updates, complex influence links are captured, ultimately generating a network node relation matrix.

[0039] Finally, a multi-layered fusion network is used to integrate the static topology and dynamic state features, and model distillation is applied to optimize computational efficiency. This embodiment designs a multi-layered fusion network architecture, initially integrating static topology features and dynamic state features through feature-level fusion, and then using an attention mechanism to weightedly combine different features.

[0040] To optimize computational efficiency, this embodiment employs model distillation technology to transfer knowledge from a complex teacher model to a simplified student model, significantly reducing the computational complexity of online inference. Based on this, this embodiment ultimately outputs network topology feature vectors and relation matrices, providing a foundation for subsequent perturbation sampling and anomaly identification.

[0041] Step S3: Based on the network topology feature vector and relationship matrix, actively detect network weaknesses and potential fault points through a perturbation sampling strategy, identify network bottlenecks and abnormal patterns, conduct a comprehensive risk assessment, and obtain a network risk assessment report and an optimization target list.

[0042] In this step, the focus shifts from passive monitoring to active detection, using perturbation sampling strategies to detect potential weaknesses and failure risks in the network.

[0043] Specifically, based on network topology feature vectors and relationship matrices, a multi-level progressive perturbation sampling strategy is designed, including three levels: micro-perturbation, moderate perturbation, and strong perturbation. Micro-perturbation is the lightest probing method, testing the network's sensitivity to minor load changes by sending specially crafted probe packets or briefly increasing low-intensity traffic (typically not exceeding 5% of link capacity) on a specific path. Moderate perturbation increases the perturbation intensity, generating moderate traffic pressure (approximately 10%-30% of link capacity) or altering traffic distribution patterns, used to detect potential performance bottlenecks under certain loads. Strong perturbation is the most invasive test, including simulating node failures, link interruptions, or sudden large traffic spikes, used to test the network's response and recovery capabilities under extreme conditions.

[0044] Based on the above perturbation strategy, this embodiment uses a distributed controller to collaboratively execute perturbation injection and collect network response data in real time.

[0045] In this step, a distributed controller network is deployed, with each controller responsible for managing the execution of disturbances in a specific network area. Before executing a disturbance, the system performs a pre-check to confirm that the current network state meets the conditions for safe execution. Then, the disturbance is injected gradually according to the plan, while response data is collected in real time through distributed monitoring probes, including the dynamic changes in indicators such as CPU utilization, memory usage, queue length changes of each node, and bandwidth utilization, latency, and packet loss rate of each link.

[0046] Typically, it is necessary to focus on the comparison of states before and after the disturbance and the fluctuation process, record the magnitude, duration and recovery pattern of the indicator changes, and form a structured disturbance response dataset.

[0047] Next, based on the disturbance response dataset and network topology feature vectors and relation matrices, gradient analysis and influence propagation models are used to identify network nodes and links sensitive to disturbances. First, gradient analysis is performed to calculate the rate of change of performance indicators for each network element before and after the disturbance, quantifying its sensitivity to the disturbance and constructing a comprehensive sensitivity score.

[0048] The high sensitivity score indicates that the element reacts strongly to disturbances and may be a potential vulnerability. Next, the system applies an impact propagation model to analyze how disturbances propagate and spread within the network, identifying key propagation paths and amplification nodes.

[0049] This embodiment generates a network sensitivity map by integrating the results of gradient analysis and propagation model. The map displays the vulnerability of each element in the network in the form of a heat map and marks key propagation paths and risk cluster areas.

[0050] Specifically, based on a network sensitivity map and standardized network status data stream, clustering algorithms and anomaly detection models are applied to detect and classify network bottlenecks. By combining the sensitivity map with historical network status data, unsupervised clustering algorithms are applied to group and identify patterns of sensitivity points, considering multi-dimensional features such as response characteristics, location characteristics, and historical performance indicators.

[0051] Based on the clustering results, a classification model is further applied to categorize bottlenecks into several major types: capacity bottlenecks (caused by hardware resource limitations), configuration bottlenecks (caused by improper parameter settings or strategies), and topology bottlenecks (caused by network structure defects). This embodiment also applies a timing anomaly detection algorithm to identify the time patterns of bottleneck occurrence, distinguish between persistent and periodic bottlenecks, and generate a detailed network bottleneck analysis report.

[0052] Finally, a comprehensive risk assessment is conducted based on the network bottleneck analysis report and network sensitivity map, combined with business importance and service level agreement (SLA) requirements. Typically, this involves importing business criticality mappings and SLA requirements, overlaying the business mappings with network bottleneck and sensitivity information, and calculating a risk score. This score incorporates factors such as bottleneck severity, impact scope, business importance, and SLA requirements.

[0053] Based on risk scoring, all risk points are prioritized, and specific optimization goals are set for high-priority risk points, including performance indicator improvement goals, reliability enhancement goals, and resilience enhancement goals. This embodiment also considers optimization difficulty and resource requirements, dividing the goals into short-term, medium-term, and long-term categories to form a phased optimization plan, generating a comprehensive network risk assessment report and optimization goal list.

[0054] Step S4: Based on the network risk assessment report and the optimization target list, a set of candidate topology adjustment schemes is generated using a multi-objective optimization algorithm. Each candidate topology adjustment scheme is comprehensively evaluated to obtain the optimal topology adjustment scheme.

[0055] Understandably, step S4 transforms the optimization objective into a concrete and feasible solution. First, based on the network risk assessment report and the list of optimization objectives, the objectives are converted into mathematical objective functions with assigned weighting coefficients. Then, each objective in the list is analyzed and transformed into a quantifiable mathematical expression; for example, the bandwidth utilization balance objective is defined as minimizing the standard deviation of link utilization, and the latency optimization objective is defined as minimizing the average end-to-end latency of the network.

[0056] In this approach, a weight coefficient is assigned to each objective based on service importance and SLA requirements. For example, a network supporting real-time services might assign a higher weight to the latency objective and a medium weight to the bandwidth utilization objective. In this way, this embodiment can transform the list of optimization objectives into a formalized multi-objective optimization problem definition.

[0057] Next, this embodiment identifies network physical constraints, service constraints, and resource constraints, transforming these constraints into mathematical constraints. Typically, physical constraints include physical connection limits, link capacity limits, and node processing capabilities; service constraints include quality of service requirements, path limitations, and security policies; and resource constraints include equipment budget limits, energy consumption limits, and management complexity thresholds. The system converts these constraints into mathematical expressions, forming a feasible solution space for the optimization problem.

[0058] This embodiment generates a set of candidate topology adjustment schemes based on the constraint model of the optimization problem and the formal definition of the multi-objective optimization problem, using a multi-objective evolutionary algorithm and Pareto optimal front technique. This embodiment employs an improved non-dominated sorting genetic algorithm (NSGA-III) to encode the topology adjustment schemes as chromosomes, optimizing the quality of the solution generation by generation through selection, crossover, and mutation operations.

[0059] Unlike single-objective optimization, multi-objective optimization seeks a Pareto optimal solution set, meaning it cannot improve solutions for other objectives without compromising at least one objective. This embodiment identifies Pareto fronts of different levels through non-dominated sorting and uses a reference point assistance mechanism to ensure population diversity. In summary, after multiple generations of evolution, the algorithm converges to a set of candidate solutions distributed along the Pareto optimal fronts, forming a candidate topology adjustment scheme set.

[0060] Next, based on the candidate topology adjustment scheme set, each scheme is simulated and tested using a network simulation engine. Specifically, a high-precision network simulation engine is used, capable of simulating various characteristics of a real network environment. First, a corresponding network topology model is built in the simulation environment according to the configuration parameters of the candidate schemes. Then, various load test scenarios are prepared, including normal working load, peak load, abnormal traffic patterns, and fault conditions. By performing simulation tests on each scheme and scenario combination, key performance indicator data is collected, potential risks and expected benefits are assessed, and a scheme performance prediction report is generated.

[0061] Finally, based on the performance prediction report and target weight coefficients, a multi-criteria decision analysis method is applied to comprehensively evaluate candidate solutions and select the optimal balance point. Specifically, an evaluation matrix is ​​constructed, and an evaluation is performed using a combination of the analytic hierarchy process (AHP) and the ideal solution method. The AHP is used to determine the relative importance weights of the evaluation criteria, while the ideal solution method is used to determine the relative closeness between the candidate solutions and the ideal solution.

[0062] In summary, this embodiment can select the optimal topology adjustment scheme based on the score ranking and generate a detailed implementation plan, including the order of change steps, change window suggestions, preparatory conditions, verification methods, and rollback strategies.

[0063] Step S5: Receive the optimal topology adjustment scheme, convert the optimal topology adjustment scheme into a network topology change instruction sequence, realize the synchronous execution of configuration changes among multiple nodes through a distributed coordination mechanism, complete the distribution and execution of topology reconstruction instructions, and simultaneously perform dynamic resource scheduling and route updates to obtain topology adjustment results and performance monitoring data. Based on the topology adjustment results and performance monitoring data, continuously optimize the decision model.

[0064] Step S5 implements the entire process from optimization scheme to actual network adjustment. First, the optimization decisions in the optimal topology adjustment scheme are converted into a network configuration instruction set. A network device configuration template library is built, each decision in the optimization scheme is parsed, and it is mapped to the corresponding configuration change operation. Typically, during the conversion process, the atomicity and consistency of the instructions are ensured, the dependencies between instructions are analyzed, the correct execution order is determined, and a structured sequence of network topology change instructions is generated.

[0065] In this embodiment, a distributed transaction management framework is constructed based on the network topology change instruction sequence, and a two-phase commit protocol is used to ensure the synchronization of configuration changes among multiple nodes. It can be understood that the two-phase commit protocol, as a distributed algorithm, can be used to coordinate whether all participants agree to execute a transaction.

[0066] Typically, in the first phase (preparation phase), the central coordinator sends change plans to all participating nodes, and each node verifies the feasibility of the plans and returns a response. Only when all nodes report readiness does the coordinator proceed to the second phase (commit phase) to execute the actual configuration changes. This embodiment also includes a comprehensive rollback mechanism. When execution fails or unexpected results are detected, rollback commands can be executed in reverse order to restore the network to its state before the change, ensuring that complex network topology changes can be executed safely and reliably.

[0067] Next, according to the collaborative execution plan, instructions are rapidly distributed and executed through optimized message queues and parallel processing mechanisms. Specifically, this embodiment uses a publish-subscribe model message queue architecture to distribute instructions and implements an intelligent parallel processing mechanism. By analyzing the dependencies between instructions, independent task groups are executed in parallel. This embodiment also employs time window technology to control change risks. Large changes are broken down into multiple smaller batches, executed within a predetermined time window, and verification checkpoints are set. By collecting execution status feedback from each node in real time, a detailed execution status database is formed.

[0068] Next, based on the execution status feedback data, dynamic scheduling of computing, storage, and network resources is performed. Specifically, the system adjusts the deployment locations of virtual machines, containers, or microservices according to the location and connection characteristics of nodes in the new topology; it optimizes data sharding strategies and cache locations based on access pattern analysis and new topology characteristics; this embodiment applies a load balancing algorithm to optimize traffic allocation, dynamically allocating traffic based on weighted round-robin and minimum connection count strategies. Furthermore, this embodiment also applies a path optimization algorithm based on traffic engineering, guiding traffic along the optimal path and avoiding congestion points by adjusting routing weights.

[0069] Finally, network performance data after the topology change is collected by deploying real-time monitoring points, and compared with the performance indicators predicted in the solution performance prediction report. Specifically, the system deploys real-time monitoring points on key nodes and links to collect fine-grained performance data, and deploys end-to-end performance probes to measure end-to-end service quality indicators. By performing a three-way comparison analysis between the actual monitored performance data and the baseline data and expected values ​​before optimization, the degree of improvement of each indicator is calculated, the deviation between the actual effect and the predicted value is analyzed, and the accuracy of the model prediction is evaluated. This embodiment also performs root cause analysis to identify potential problems and propose further optimization suggestions, generating topology adjustment results and performance monitoring data reports.

[0070] Based on the aforementioned topology adjustment results and performance monitoring data, the decision-making model is continuously optimized, forming a closed-loop optimization system in this embodiment. Specifically, by analyzing the difference between actual and predicted results, parts of the model that need improvement are identified, such as inaccurate constraints, unreasonable weight settings, or missing influencing factors. Simultaneously, model parameters are updated using incremental learning methods, adjusting the weights and hyperparameters of the prediction model based on new observation data. This embodiment also applies reinforcement learning techniques, treating network optimization as a continuous decision-making process. By trying different optimization strategies and observing their actual effects, the decision-making strategy is gradually improved.

[0071] Understandably, this continuous optimization mechanism enables the system to learn and adapt to changes in the network environment, thereby improving the accuracy and effectiveness of subsequent optimization decisions.

[0072] In summary, based on steps S1-S5 above, this embodiment achieves adaptive optimization of network topology, forming a complete closed-loop system from data acquisition to scheme execution and effect verification. The optimization method in this embodiment combines distributed monitoring, deep learning, active detection, and collaborative execution, enabling it to accurately identify network bottlenecks, generate optimal topology adjustment schemes, and execute network changes safely and reliably, significantly improving network performance and reliability, and allowing the network to better adapt to business needs and environmental changes.

[0073] The following will provide further explanation of steps S1-S5.

[0074] In this embodiment, step S1, which involves acquiring network node status data and link load data based on a distributed network probe and a multi-level acquisition strategy, and then performing data standardization, noise filtering, and time-series alignment on the network node status data and link load data to obtain a standardized network status data stream, includes: Step S11: Based on the network topology and node distribution, determine the probe deployment location through intelligent algorithms, deploy monitoring agents on network nodes at different levels, and obtain a probe deployment scheme covering the entire network.

[0075] In step S11, the system analyzes the current network topology and identifies different types of network nodes, such as core switching nodes, edge routing nodes, and critical service server nodes. Then, a graph-based probe coverage optimization algorithm is used to determine the minimum number of probe deployment locations to ensure that the status of any link in the network can be monitored by at least one probe.

[0076] For large-scale networks, this embodiment employs a hierarchical deployment strategy, deploying different types of probes at the backbone, aggregation, and access layers to form a multi-layered probe network. After determining the probe deployment locations, the system automatically generates probe deployment scripts and deploys lightweight monitoring agents to the corresponding network devices through the network configuration management system.

[0077] Understandably, these monitoring agents are able to collect key performance indicators such as device CPU utilization, memory usage, interface traffic, and packet loss rate, and transmit the data to the central management platform through a secure channel.

[0078] Step S12: Based on the probe deployment scheme, set differentiated sampling frequencies and monitoring granularities, determine the sampling frequency of each node according to the node load characteristics, and obtain multi-level network raw monitoring data.

[0079] In this step, the system adopts different sampling strategies for different nodes based on their importance and load characteristics in the network.

[0080] For example, a high-frequency sampling strategy is used for high-load nodes such as core switches and critical routing nodes, with a sampling interval typically set to 5-50 milliseconds to capture transient network fluctuations. For low-load nodes such as edge nodes and terminal access devices, an adaptive sampling strategy is used, with an initial sampling interval set to 1-5 seconds, dynamically adjusted according to load changes. Specifically, when a sudden increase or anomaly in traffic is detected, the sampling frequency is automatically increased to the millisecond level, and reduced to the minute level when the network status is stable, in order to balance monitoring accuracy and system overhead.

[0081] At the same time, the system will also set different monitoring granularities according to the changing characteristics of different indicators. For example, a higher sampling frequency will be used for indicators with large fluctuations (such as instantaneous traffic), and a lower sampling frequency will be used for indicators with slow changes (such as link utilization trends), thereby forming a multi-dimensional and multi-granular raw monitoring dataset.

[0082] Step S13: Receive the original monitoring data of the multi-level network, and unify the network data from different sources and of different types into a standard format through unit conversion, dimension normalization and feature scaling to obtain a standardized network indicator dataset; based on the standardized network indicator dataset, apply sliding window midpoint filtering and wavelet transform methods to remove noise and outliers to obtain cleaned network status data.

[0083] Those skilled in the art will recognize that network devices come from different manufacturers and may report performance metrics using different units and metrics. For example, some devices may report bandwidth in bits per second, while others may use bytes per second.

[0084] Based on this, the system first performs a unit conversion, converting all traffic metrics to bits per second and latency metrics to milliseconds, etc. Then, for metrics with different dimensions (such as CPU utilization of 0-100% and traffic of 0-10Gbps), the system uses a min-max normalization method to map various metrics to the [0,1] interval, eliminating dimensional differences. For unevenly distributed metric data, the system also applies feature scaling techniques such as logarithmic transformation or Z-score normalization to make the data distribution more uniform, facilitating subsequent analysis and processing.

[0085] In summary, after the above processing, the raw data from different devices and different indicators are transformed into a standardized network indicator dataset, laying the foundation for subsequent analysis.

[0086] Network monitoring data typically contains various noises and occasional anomalies, which may arise from measurement errors, network jitter, or temporary faults. This embodiment uses sliding window mid-range filtering (a non-linear filtering technique) to remove impulse noise.

[0087] Typically, by setting an appropriate window size (usually 5-15 sampling points) and performing sliding processing on the data sequence, the median value of the data within the window is used to replace the center point value, which effectively removes short-term interference and isolated outliers while preserving data edge and trend information.

[0088] Specifically, for complex noise patterns, this embodiment employs the discrete wavelet transform method for multi-scale analysis, decomposing the signal into different frequency components, identifying and filtering high-frequency noise, and preserving the main features of the signal. In practical applications, by selecting wavelet basis functions suitable for the characteristics of the network data (such as Daubechies wavelet or Haar wavelet), the data is decomposed into multiple layers, a threshold is set to filter noise coefficients, and then the signal is reconstructed to obtain denoised and smoothed data.

[0089] In summary, this embodiment, through the aforementioned signal processing techniques, can effectively remove noise and outliers from network monitoring data and extract the true characteristics of network status changes.

[0090] Step S14: Based on the cleaned network status data, time sequence alignment is achieved through timestamp correction and interpolation algorithms. Multi-source data fusion technology is used to perform correlation analysis on the node status and link load data in the cleaned network status data to obtain the standardized network status data stream.

[0091] It is understandable that the data collected by distributed network probes may have time synchronization issues, as the internal clocks of different devices may have slight differences, resulting in inconsistencies in timestamps.

[0092] Based on this, this embodiment first performs timestamp correction. By synchronizing with a standard network time server, the time offset of each probe is calculated, and the timestamps of the collected data are adjusted accordingly. For data with different sampling frequencies, a cubic spline interpolation algorithm is used to generate data values ​​at a unified time point to achieve time sequence alignment. Then, multi-source data fusion technology is applied to perform correlation analysis between node status data (such as CPU utilization and memory usage) and link load data (such as traffic, latency, and packet loss rate).

[0093] Specifically, a network topology graph model is established, mapping various types of data to corresponding nodes and links. Graph analysis algorithms are used to uncover the causal relationships and mutual influences between node states and link loads, such as identifying patterns where increased CPU load on a node leads to increased processing latency on connected links. Finally, a standardized network state data stream containing three dimensions—time, space, and metrics—is generated, providing a complete and consistent data foundation for subsequent dynamic topology modeling.

[0094] In summary, through steps S11-S14, this embodiment realizes the transformation process from raw monitoring data to a standardized network status data stream, providing high-quality data input for adaptive network topology optimization. This multi-level acquisition and processing mechanism can adapt to network environments of different scales and complexities, ensuring that subsequent analysis models obtain accurate and complete network status information.

[0095] In this embodiment, in step S2, the standardized network state data stream is received, and a dynamic network topology model is constructed using a Stancu-type neural network operator. The Stancu-type neural network operator includes adjustable higher-order terms and nonlinear transformations. The dynamic network topology model captures the dynamic relationships and topological features between network nodes, resulting in a network topology feature vector and a relationship matrix, including: Step S21: Based on the network topology optimization requirements, construct a mathematical model of a Stancu type neural network operator, introduce the adjustable higher-order terms and nonlinear transformations, and obtain the Stancu type neural network operator framework.

[0096] Understandably, the Stancu-type neural network operator is an extension of the traditional neural network. Its core feature is the introduction of adjustable higher-order terms and nonlinear transformations, giving it stronger function approximation and expressive power. In this embodiment, the Stancu-type neural network operator, based on Stancu polynomial interpolation theory, extends the activation function of the traditional feedforward neural network, enabling it to capture complex nonlinear network topological relationships.

[0097] Specifically, the theoretical foundation of Stancu-type neural network operators originates from the Stancu polynomial in numerical analysis. It is a generalized form of the Bernstein polynomial, offering more flexible parameter configuration and stronger approximation capabilities. Unlike traditional polynomials, Stancu polynomials include two adjustable parameters, α and β, giving them greater flexibility and accuracy in approximating nonlinear functions. The core advantage of this polynomial lies in its ability to alter the shape and approximation characteristics by adjusting these two parameters, thereby better adapting to different types of function shapes and nonlinear relationships.

[0098] In this embodiment, the idea of ​​Stancu polynomials is incorporated into the neural network architecture, forming a Stancu-type neural network operator. Specifically, the activation function of each neuron in the neural network is modified. Traditional neural networks typically use fixed activation functions (such as ReLU, Sigmoid, or Tanh), while the Stancu-type neural network operator replaces these fixed activation functions with parameterized activation functions based on Stancu polynomials. This activation function can be expressed as: ; Where, φ( x ) is the basic activation function (such as ReLU), Stancu( x ; α, β, i ) is an i-th order Stancu polynomial term, λ i These are learnable weight coefficients, α and β are parameters that control the shape of the Stancu polynomial, and n is the highest order of the polynomial.

[0099] Understandably, this design makes the activation function adaptive, enabling it to automatically adjust its nonlinearity and complexity based on the characteristics of the data.

[0100] Typically, the training process for a Stancu type neural network operator includes the following steps: First, in the initialization phase, appropriate initial parameter values ​​are set for each neuron based on the complexity of the network topology and the characteristics of the data. Generally, higher-order polynomial representations are assigned to complex network regions (such as core switching nodes or high-traffic convergence points), while lower-order representations are used for relatively simple edge nodes, achieving a reasonable allocation of computational resources. Second, in the forward propagation phase, the Stancu activation value of each neuron is calculated to capture the complex nonlinear characteristics of the input signal. Finally, in the backpropagation phase, not only are the traditional weights and bias parameters updated, but the coefficients and shape parameters of the Stancu polynomial are also optimized simultaneously, allowing the activation function to adaptively adjust to best fit the data distribution characteristics.

[0101] Furthermore, the Stancu-type neural network operator introduces a dynamic order adjustment mechanism, which automatically determines the optimal polynomial order required for each neuron during training. Specifically, starting from a lower order, the order is gradually increased according to changes in the fitting error until the error improvement is no longer significant or a preset maximum order limit is reached. This progressive complexity growth strategy ensures the model's expressive power while effectively preventing overfitting, making it suitable for data with complex and dynamically changing network topologies.

[0102] In specific applications of network topology optimization, firstly, the Stancu-type neural network operator can more accurately capture the nonlinear variation patterns of network traffic, especially in scenarios such as traffic bursts, congestion propagation, and cascading failures, where traditional linear or simple nonlinear models often struggle to accurately model these complex phenomena. Secondly, it can effectively handle multi-scale features in the network, simultaneously focusing on micro-level node state changes and macro-level global topology evolution, providing a unified framework for a comprehensive understanding of network behavior. Finally, it possesses stronger generalization ability and robustness, adapting to dynamic changes in network topology and unseen network states, maintaining good performance even with limited training data.

[0103] In this step, the basic neural network architecture is first defined, and then the activation function of each neuron is modified to a Stancu type, introducing higher-order polynomial terms and parameterized nonlinear transformations. Unlike traditional neural networks that use fixed activation functions (such as ReLU, sigmoid, etc.), each activation unit of a Stancu type neural network operator contains multiple adjustable parameters, which can adaptively adjust its nonlinearity and higher-order characteristics, thereby more accurately modeling complex network topologies.

[0104] Typically, these parameters are optimized along with other weights during network training, enabling the model to automatically adapt to the characteristics of different network environments. Through this improvement, the Stancu-type neural network operator can more effectively handle multivariate, multi-objective optimization problems in network topology, improving modeling accuracy and convergence speed.

[0105] Step S22: Based on the Stancu type neural network operator framework and the standardized network state data stream, construct a graph structure representation of the network topology, and learn the embedding vectors of nodes and links in the graph structure representation through a neural network to obtain a low-dimensional feature representation of network elements.

[0106] In step S22, the network physical topology is first represented as a graph structure, where nodes represent network devices (such as routers and switches) and edges represent physical links connecting devices. Then, standardized network state data is mapped onto this graph structure, associating each node and link with corresponding state features (such as node load, link bandwidth utilization, etc.). Next, graph neural network (GNN) techniques, particularly a combined architecture of graph convolutional networks (GCN) and graph attention networks (GAT), are applied to learn low-dimensional embedding representations of network elements.

[0107] This process also considers the characteristics of the nodes themselves, the characteristics of neighboring nodes, and the attributes of the links. Through multi-layer graph convolution operations and attention mechanisms, low-dimensional vector representations that can represent the topological location and functional characteristics of network elements are gradually extracted. These embedded vectors are typically 64-256 dimensions, which not only preserves the key features of network elements but also greatly reduces the complexity of subsequent processing.

[0108] It is important to understand that applying the Stancu type neural network operator as the core processing unit in the graph neural network at this stage enables the model to more accurately capture the nonlinear relationships and high-order interaction patterns between network elements.

[0109] Step S23: Based on the low-dimensional feature representation of the network elements, apply a hybrid model of temporal convolutional network and long short-term memory network to capture the temporal change pattern of the network state in the low-dimensional feature representation of the network elements, and obtain the dynamic temporal features of the network.

[0110] Understandably, the dynamic characteristics of network topology are reflected in the changing patterns of node states and link loads over time, and capturing these temporal features is crucial for predicting network behavior and optimizing topology.

[0111] In this step, the low-dimensional feature representations obtained in the previous step are first organized into a temporal sequence in chronological order, and then a hybrid temporal model is applied for modeling. This hybrid model consists of two main components: a temporal convolutional network (TCN) and a long short-term memory network (LSTM). The TCN is responsible for extracting local temporal patterns, capturing changing features at different time scales through multiple layers of dilated convolutions. Its receptive field grows exponentially with the number of layers, effectively capturing long-term dependencies. The LSTM focuses on modeling long-term dependencies and sequence memory, retaining important information and filtering out irrelevant information through gating mechanisms.

[0112] The outputs of the two models are fused through an attention mechanism, adaptively adjusting the weights of each component based on the characteristics of different network elements and time points to form a comprehensive temporal feature representation. This hybrid architecture can capture both the periodic changes in network traffic and the impact of sudden events, as well as model long-term evolution trends, providing rich dynamic temporal information for subsequent relational reasoning.

[0113] Step S24: Based on the network dynamic temporal features, the dynamic association strength and influence propagation mode between nodes in the network dynamic temporal features are learned through graph attention mechanism and relational reasoning network to obtain the network node relation matrix.

[0114] It is understandable that there are complex interrelationships between network nodes. A change in the state of one node may propagate through links and affect other nodes, creating a cascading effect.

[0115] This embodiment uses a multi-head graph attention mechanism to calculate the attention weights between nodes. Each attention head learns the association patterns between nodes from different perspectives and captures different types of dependencies.

[0116] Specifically, attention scores are calculated between node pairs, representing the degree to which a change in the state of one node affects the other. These attention scores are calculated based on the node's feature vectors, historical interaction patterns, and network topology, and are normalized using the softmax function to form an attention weight matrix.

[0117] Next, a relational reasoning network is applied to model the influence propagation path between nodes based on an attention weight matrix. This reasoning network employs a message passing mechanism to simulate the dynamic process of information flow and influence propagation in the network, identifying direct and indirect relationships.

[0118] In summary, through multiple rounds of iterative updates, this embodiment captures multi-hop dependencies and complex influence links, and finally generates a network node relationship matrix, which describes the dynamic association strength and influence propagation pattern between nodes.

[0119] Step S25: Based on the network node relationship matrix, the low-dimensional feature representation of the network elements, and the network dynamic temporal features, the static topology structure and dynamic state features are integrated through a multi-level fusion network. The computational efficiency is optimized by applying model distillation technology to obtain the network topology feature vector and relationship matrix.

[0120] Understandably, in order to construct a complete dynamic network topology model, it is necessary to effectively integrate the various features obtained in the previous steps. Based on this, step S25 applies a multi-layered fusion network architecture, which contains multiple fusion modules, each responsible for feature integration at a specific level.

[0121] First, the system initially integrates static topology features (such as node type and link bandwidth capacity) with dynamic state features (such as real-time load and traffic patterns) through feature-level fusion. Then, an attention mechanism is used to weight and combine features from different sources, adaptively adjusting the importance of various features according to the current network state.

[0122] For complex networks, directly computing the complete model can be computationally expensive. Therefore, the system employs model distillation to optimize computational efficiency. Specifically, a complex teacher model, containing complete Stancu-type neural network operators and various features, is first trained offline. Then, its knowledge is transferred to a simplified student model. By matching the output distribution and intermediate representations, the student model approximates the performance of the teacher model.

[0123] Understandably, this distillation method significantly reduces the computational complexity of online inference, enabling the system to operate efficiently in resource-constrained environments. After the above processing, the final output is a network topology feature vector (representing the key features of each node and link) and a relationship matrix (representing the dynamic interaction relationships between nodes), providing a foundation for subsequent perturbation sampling and anomaly identification.

[0124] In this embodiment, in step S3, based on the network topology feature vector and relationship matrix, the network weaknesses and potential fault points are actively detected through a perturbation sampling strategy, network bottlenecks and abnormal patterns are identified, a comprehensive risk assessment is performed, and a network risk assessment report and optimization target list are obtained, including: Step S31: Based on the network topology feature vector and relation matrix, design a multi-level progressive perturbation sampling strategy, which includes micro-perturbation, medium-level perturbation and strong perturbation, to obtain a network perturbation sampling execution plan.

[0125] Understandably, perturbation sampling is a technique for actively probing network vulnerabilities. It involves injecting artificial perturbations of varying degrees into the network and observing the network response to identify potential weaknesses and fault points. In this embodiment, the system designs a three-level progressive perturbation sampling strategy based on the analysis results of the network topology feature vector and relationship matrix, including micro-perturbation, moderate perturbation, and strong perturbation sampling strategies.

[0126] Micro-perturbation is the lightest probing method, primarily testing a network's sensitivity to minor load changes by sending specially crafted probe packets (such as packets of specific size and priority) or briefly increasing low-intensity traffic on a specific path (typically not exceeding 5% of link capacity). This perturbation has almost no impact on normal network operation and is suitable for continuous execution in production environments.

[0127] Moderate disturbances increase the intensity of the disturbance, including generating moderate traffic pressure on selected links (approximately 10%-30% of link capacity) or altering traffic distribution patterns, such as temporarily changing some traffic paths by modifying route weights. These types of disturbances can reveal potential performance bottlenecks under certain loads and are typically executed when network load is low.

[0128] Strong disturbances are the most invasive testing methods, including simulating node failures (such as temporarily shutting down a specific interface), link interruptions, or sudden surges in traffic (potentially approaching the link capacity limit), to test the network's response and recovery capabilities under extreme conditions.

[0129] In summary, based on these three levels of perturbation strategies, this embodiment can formulate detailed sampling execution plans for different network areas and time periods according to network status and service importance, including parameters such as perturbation type, intensity, duration, and target location.

[0130] Step S32: Based on the network disturbance sampling execution plan, disturbance injection is executed collaboratively by a distributed controller. The network response data to the disturbance is collected in real time during the execution of the network disturbance sampling execution plan, and the state changes of each node and link are recorded to obtain the disturbance response dataset.

[0131] Typically, perturbation injection requires precise control and coordination to ensure the accuracy and security of the test. This embodiment deploys a distributed controller network, with each controller responsible for managing perturbation execution in a specific network area. These controllers maintain real-time communication through secure channels to ensure precise synchronization of perturbation operations.

[0132] When a disturbance is initiated, a pre-check is first performed to verify whether the current network status meets the conditions for safe execution. For example, it is confirmed that the current network load does not exceed a preset threshold and that critical service links have sufficient redundancy. Then, the disturbance is injected gradually according to the plan, while network response data is collected in real time through distributed monitoring probes. This response data includes the CPU utilization, memory usage, and queue length changes of each node, as well as the dynamic changes in bandwidth utilization, latency, and packet loss rate of each link.

[0133] In this embodiment, special attention is paid to the comparison of states before and after the disturbance and the fluctuation process, recording the magnitude, duration, and recovery mode of the index changes. For strong disturbance tests, the fault recovery process and convergence time are also monitored. All collected data are accompanied by precise timestamps and disturbance-related markers, forming a structured disturbance response dataset, providing rich experimental evidence for subsequent analysis.

[0134] Step S33: Based on the disturbance response dataset and the network topology feature vector and relationship matrix, gradient analysis and influence propagation model are used to identify the network nodes and links in the disturbance response dataset that are sensitive to disturbances, assess the vulnerability of the network nodes and links that are sensitive to disturbances, and obtain a network sensitive point map.

[0135] The in-depth analysis of the disturbance response data in step S33 is a key step in identifying network weaknesses. The system first performs gradient analysis to calculate the rate of change of performance indicators of each network element (node ​​or link) before and after the disturbance, quantifying its sensitivity to the disturbance.

[0136] Specifically, a comprehensive sensitivity score is constructed by calculating the normalized change magnitude of various indicators (such as bandwidth utilization, latency, and packet loss rate) and combining the duration of the change with the recovery speed.

[0137] Typically, a high sensitivity score indicates that an element reacts strongly to disturbances and may be a potential vulnerability. Next, an impact propagation model is applied, based on the network topology matrix, to analyze how disturbances propagate and spread within the network. This model simulates the process of a fault or congestion spreading from its origin to surrounding nodes, identifying key propagation paths and amplifying nodes that play important roles in cascading failures.

[0138] This embodiment also analyzes the impact range and propagation speed of different types of disturbances. For example, some nodes may be insensitive to traffic surges but highly sensitive to link interruptions. By integrating the results of gradient analysis and propagation models, a network sensitivity map is generated. This map displays the vulnerability level of each element in the network in the form of a heatmap, and marks key propagation paths and risk cluster areas.

[0139] Step S34: Based on the network sensitive point map and the standardized network status data stream, apply clustering algorithms and anomaly detection models to detect and classify network bottlenecks in the network sensitive point map, distinguish between capacity bottlenecks, configuration bottlenecks, and topology bottlenecks, and obtain a network bottleneck analysis report.

[0140] Those skilled in the art will recognize that network bottlenecks are a key factor limiting overall performance, and accurately identifying bottleneck types is crucial for developing effective optimization strategies.

[0141] In step S34, the network sensitive point map is combined with historical network state data for analysis. Unsupervised clustering algorithms (such as K-means, DBSCAN, and hierarchical clustering) are applied to group and identify patterns of sensitive points. The clustering process considers multidimensional features, including response features (such as sensitivity and recovery time), location features (such as topological centrality and connectivity), and historical performance indicators (such as average utilization and volatility), thereby discovering groups of sensitive points with similar characteristics.

[0142] Based on the clustering results, the system further applies a specialized classification model to divide the bottlenecks into several major categories, including: capacity bottlenecks (caused by hardware resource limitations, such as insufficient link bandwidth or limited node processing capacity), configuration bottlenecks (caused by improper parameter settings or strategies, such as unreasonable queue management strategies or overly conservative flow control mechanisms), and topology bottlenecks (caused by network structure defects, such as single-point dependence or insufficient path redundancy).

[0143] This embodiment also applies a timing anomaly detection algorithm to identify the timing patterns of bottleneck occurrences and distinguish between persistent and periodic bottlenecks. Finally, a detailed network bottleneck analysis report is generated, which includes the type, severity, scope of impact, conditions of occurrence, and potential causes of each bottleneck, providing precise guidance for subsequent optimization.

[0144] Step S35: Based on the network bottleneck analysis report and the network sensitive point map, and in combination with the importance of the business and the service level agreement requirements, a comprehensive risk assessment is conducted. Priorities are assigned to the risk points identified in the network bottleneck analysis report and the network sensitive point map, and corresponding optimization targets are generated to obtain the network risk assessment report and the optimization target list.

[0145] Understandably, network optimization resources are usually limited and need to be allocated effectively to solve the most critical problems. Therefore, risk assessment and prioritization are important steps in developing an optimization plan.

[0146] In step S35, a service criticality map is first imported, which defines the service functions supported by each part of the network and their importance. Service Level Agreement (SLA) requirements for each service are also considered, such as availability targets and response time commitments. Then, the service map is overlaid with network bottleneck and sensitivity information to assess the potential impact of each risk point on the service.

[0147] Specifically, a risk score is calculated, which combines factors such as bottleneck severity, impact scope, business importance, and SLA requirements, and considers the likelihood of the risk materializing (based on historical data and disturbance test results). Based on the risk score, all identified risk points are prioritized, and specific optimization objectives are set for high-priority risk points.

[0148] Typically, these optimization goals include performance improvement goals (such as reducing the utilization of a specific link to below 70%), reliability improvement goals (such as increasing the redundancy of the critical path), and resilience enhancement goals (such as improving fault recovery time).

[0149] This embodiment also considers optimization difficulty and resource requirements, dividing the objectives into short-term, medium-term, and long-term categories to form a phased optimization plan. Ultimately, a comprehensive network risk assessment report and a list of optimization objectives are generated, providing clear goals and directions for the next step of topology optimization decisions.

[0150] In summary, through steps S31-S35, this embodiment achieves a shift from passive monitoring to active detection. By employing a perturbation sampling strategy, it reveals potential weaknesses and fault risks in the network, providing a more comprehensive and in-depth basis for network topology optimization. This proactive risk identification method can discover problems that are difficult to detect using traditional passive monitoring, greatly improving the foresight and accuracy of network optimization.

[0151] In this embodiment, step S4, which involves generating a set of candidate topology adjustment schemes based on the network risk assessment report and the optimization target list using a multi-objective optimization algorithm, comprehensively evaluating each candidate topology adjustment scheme, and obtaining the optimal topology adjustment scheme, includes: Step S41: Based on the network risk assessment report and the optimization target list, the optimization targets in the optimization target list are transformed into mathematical objective functions and weight coefficients are set for each target to obtain a formalized definition of the multi-objective optimization problem.

[0152] Network topology optimization typically involves multiple competing objectives, such as maximizing bandwidth utilization, minimizing latency, and improving reliability.

[0153] Based on this, step S41 first analyzes each objective in the optimization target list and transforms it into a quantifiable mathematical expression. For example, for the bandwidth utilization balancing objective, the objective function is defined as minimizing the standard deviation of link utilization; for the latency optimization objective, it is defined as minimizing the average end-to-end latency of the network; for the reliability improvement objective, it is defined as maximizing the redundancy of the critical path, etc. For each objective function, corresponding weight coefficients are assigned based on business importance and SLA requirements. The weight values ​​are usually in the range [0,1], and the sum of all weights is 1.

[0154] For example, a network supporting real-time services might assign a high weight (e.g., 0.5) to latency targets, a medium weight (e.g., 0.3) to bandwidth utilization targets, and a low weight (e.g., 0.2) to cost targets. These weights can be dynamically adjusted according to service requirements to adapt to changes in the network environment.

[0155] In summary, in this way, step S41 transforms the list of optimization objectives into a formalized definition of a multi-objective optimization problem, providing a clear mathematical foundation for subsequent optimization algorithms.

[0156] Step S42: Based on the formalized multi-objective optimization problem definition and the network topology feature vector and relation matrix, identify network physical constraints, service constraints and resource constraints, and transform the network physical constraints, service constraints and resource constraints into mathematical constraints to obtain the optimization problem constraint model.

[0157] Typically, network topology optimization cannot arbitrarily adjust the network structure; various practical constraints must be considered. In step S42, physical constraints are first identified. These constraints include physical connection limitations (such as geographical distance between devices and available interface types), link capacity limits (such as maximum bandwidth), and node processing capabilities (such as maximum forwarding rate). For example, the bandwidth of a physical link cannot exceed the upper limit defined by the hardware specifications, which can be expressed as a mathematical inequality constraint.

[0158] Next, identify business constraints, including quality of service requirements (such as maximum allowable latency and minimum availability percentage), path restrictions (such as certain traffic must or must not pass through specific nodes), and security policies (such as isolation requirements). Finally, consider resource constraints, such as equipment budget limits, energy consumption caps, and management complexity thresholds. Transform these constraints into mathematical expressions, in the form of equality constraints, inequality constraints, or logical relationships.

[0159] For example, the end-to-end delay of communication between any two points must be below a certain threshold, which can be represented as an inequality constraint; the network must maintain connectivity, which can be represented as a connectivity condition in graph theory. These constraints together constitute the feasible solution space of the optimization problem, and only topology adjustment schemes that satisfy all constraints are considered valid solutions. In summary, by explicitly defining the constraint model, the system ensures that the generated topology optimization schemes are feasible in real-world environments.

[0160] Step S43: Based on the optimization problem constraint model and the formalized multi-objective optimization problem definition, apply the multi-objective evolutionary algorithm and Pareto optimal front technique to generate topology adjustment schemes for different combinations of objectives in the formalized multi-objective optimization problem definition, and obtain the candidate topology adjustment scheme set.

[0161] Those skilled in the art will know that network topology optimization is a complex combinatorial optimization problem, and traditional exact solution methods are inadequate to handle its high-dimensional and nonlinear characteristics.

[0162] This embodiment employs a multi-objective evolutionary algorithm (MOEA) to effectively explore and discover high-quality candidate solutions. Specifically, the system applies an improved non-dominated sorting genetic algorithm (NSGA-III), which maintains a population of candidate solutions by simulating biological evolution and optimizes the quality of solutions generation by generation through selection, crossover, and mutation operations.

[0163] In this algorithm, topology adjustment schemes are encoded as chromosomes, with each gene representing a configuration change for a specific link or node. The algorithm first randomly generates an initial population, and then evaluates the performance of each candidate scheme on various objective functions.

[0164] Understandably, unlike traditional single-objective optimization, multi-objective optimization does not simply merge multiple objectives, but rather seeks a Pareto optimal solution set. Pareto optimality refers to a solution that cannot improve other objectives without compromising at least one objective.

[0165] This embodiment identifies Pareto fronts of different levels through non-dominated sorting and uses a reference point assistance mechanism to ensure population diversity, thereby exploring different regions of the target space. After multiple generations of evolution (typically 50-200 generations), the algorithm converges to a set of candidate schemes distributed on the Pareto optimal front. These schemes represent the optimization results under different objective trade-offs, forming a set of candidate topology adjustment schemes.

[0166] In summary, compared with traditional single-solution methods, the method in this embodiment provides a wider range of options and can adapt to different decision preferences.

[0167] Step S44: Based on the candidate topology adjustment scheme set, simulate and test each scheme in the candidate topology adjustment scheme set using a network simulation engine, predict the performance indicators of each scheme under different load conditions, evaluate potential risks and benefits, and obtain a scheme performance prediction report.

[0168] To ensure the reliability of the topology adjustment scheme before actual deployment, the system needs to conduct a detailed evaluation of the candidate schemes.

[0169] Specifically, in step S44, a high-precision network simulation engine is used. This engine can simulate various characteristics of a real network environment, including link propagation delay, queue dynamics, protocol behavior, and traffic patterns. First, based on the configuration parameters of the candidate solution, a corresponding network topology model is constructed in the simulation environment. Then, various load test scenarios are prepared, including normal working load, peak load, abnormal traffic patterns, and fault conditions, to comprehensively evaluate the performance stability of the solution under different conditions.

[0170] For each scheme and scenario combination, the system performs simulation tests to collect key performance indicator data, such as end-to-end latency, throughput, packet loss rate, link utilization, route convergence time, and fault recovery speed. In addition, potential risks are assessed, such as single point of failure risk, link overload risk, and route oscillation risk.

[0171] This embodiment also calculates the expected benefits of each solution, such as the percentage increase in bandwidth utilization, the reduction in latency, and the improvement in availability. All test results undergo statistical analysis and visualization to form a solution performance prediction report. This report contains detailed information such as the performance distribution of each solution in various scenarios, risk assessment, and benefit prediction, providing data support for the final decision.

[0172] Step S45: Based on the performance prediction report of the proposed scheme and the weight coefficients in the formalized multi-objective optimization problem definition, apply the multi-criteria decision analysis method to comprehensively evaluate each candidate topology adjustment scheme in the candidate topology adjustment scheme set and select the best balance point, generate a detailed implementation plan, and obtain the optimal topology adjustment scheme.

[0173] Understandably, selecting the final solution from the pool of candidate solutions requires comprehensive consideration of multiple factors and a balance between different objectives.

[0174] Based on this, the system applies the Multi-Criterion Decision Analysis (MCDA) method for comprehensive evaluation and selection. Specifically, an evaluation matrix is ​​first constructed, where each row represents a candidate solution and each column represents an evaluation criterion (such as performance indicators, risk level, implementation difficulty, etc.).

[0175] Then, an evaluation is conducted using a combination of the Analytic Hierarchy Process (AHP) and the Top-Down Solution (TOPSIS) approach. AHP is used to determine the relative importance weights of the evaluation criteria. A judgment matrix is ​​established through expert evaluation or analysis based on business needs, and eigenvectors are calculated to obtain the criterion weights. TOPSIS is used to determine the relative closeness of candidate solutions to the ideal solution. This method is based on Euclidean distance, calculating the distance of each solution to the positive ideal solution (the optimal combination of all indicators) and the negative ideal solution (the worst combination of all indicators), thereby obtaining a comprehensive evaluation score.

[0176] Based on the scoring ranking, the solution with the highest overall score is usually selected as the optimal topology adjustment solution. In addition, a detailed implementation plan is generated, including the sequence of change steps, suggested change windows, preconditions, verification methods, and rollback strategies. These detailed plans consider the principle of minimizing business impact, typically recommending phased implementation of changes during off-peak periods, and setting detailed verification checkpoints to ensure the change process is controllable and verifiable.

[0177] In summary, through steps S41-S45, this embodiment realizes the transformation process from multiple optimization objectives to a specific and feasible topology adjustment scheme. It employs a multi-objective optimization method to generate diverse candidate schemes and selects the optimal implementation scheme through simulation testing and multi-criteria decision analysis. This method can effectively balance the trade-offs between different network performance objectives, generating optimization schemes that meet both technical requirements and business requirements, providing scientific decision support for adaptive network topology optimization.

[0178] In this embodiment, in step S5, receiving the optimal topology adjustment scheme, converting the optimal topology adjustment scheme into a network topology change instruction sequence, and achieving synchronous execution of configuration changes among multiple nodes through a distributed coordination mechanism, completes the distribution and execution of topology reconstruction instructions, while simultaneously performing dynamic resource scheduling and route updates, and obtains topology adjustment results and performance monitoring data, including: Step S51: Based on the optimal topology adjustment scheme, the optimization decisions in the optimal topology adjustment scheme are converted into a network configuration instruction set to ensure the atomicity and consistency of the instructions in the network configuration instruction set, thereby obtaining the network topology change instruction sequence.

[0179] Topology adjustment schemes are usually presented in the form of high-level decision descriptions, such as increasing the link bandwidth between node A and node B or optimizing the routing strategy of area C, which need to be converted into specific network configuration instructions before they can be executed.

[0180] The system first builds a network device configuration template library, containing configuration syntax and command structures for various device types and manufacturers. Then, it parses each decision in the optimization scheme and maps it to the corresponding configuration change operation. For example, a bandwidth increase decision might be converted into an interface configuration command, and a routing optimization decision might be converted into a routing protocol parameter adjustment command.

[0181] Typically, during the conversion process, the atomicity of instructions is a key focus, ensuring that each instruction either executes completely successfully or fails completely, avoiding intermediate states that could lead to network instability. Simultaneously, instruction consistency must be guaranteed, meaning that related configuration changes on different devices must remain logically consistent, such as matching interface configuration parameters between the two ends of the device.

[0182] Specifically, the system also analyzes the dependencies between instructions to determine the correct execution order and avoid temporary service interruptions caused by improper ordering. Ultimately, it generates a structured sequence of network topology change instructions, each containing information such as the target device, command content, expected result, and verification method. These instruction sequences are organized into multiple task groups based on function, with instructions within each group having inherent interrelationships and requiring coordinated execution.

[0183] Step S52: Based on the network topology change instruction sequence, construct a distributed transaction management framework, ensure the synchronization of configuration changes among multiple nodes through a two-phase commit protocol, design a rollback mechanism to deal with node execution failures, and obtain a collaborative execution plan.

[0184] Network topology changes typically involve configuration adjustments to multiple devices, and these adjustments need to be carried out in a coordinated manner to avoid temporary network inconsistencies.

[0185] Step S52 employs a distributed transaction management framework to coordinate this complex process. This framework is based on a two-phase commit protocol (2PC) to ensure the atomicity and consistency of cross-device configuration changes.

[0186] In essence, the two-phase commit protocol, as a distributed algorithm, is used to coordinate whether all participants agree to execute a transaction. In the first phase (preparation phase), the central coordinator sends a proposed change plan to all participating nodes. Each node verifies the feasibility of the plan, locks resources, and returns a ready or rejected response. Only when all nodes report ready will the coordinator enter the second phase (commit phase), sending a commit command to all nodes to execute the actual configuration changes. If any node rejects the change or fails to respond within the timeout period, the coordinator will send an abort command, and all nodes will roll back to their original state.

[0187] In addition, the system employs a comprehensive rollback mechanism, preparing corresponding rollback commands in advance for each configuration change. If execution fails or unexpected results are detected, the system can execute these rollback commands in reverse order, restoring the network to its state before the change.

[0188] In particular, the system also employs transaction logging and state recovery functions to record the execution status and results of each operation, supporting recovery of the execution status after a system crash. In summary, this embodiment, through this distributed transaction management framework, generates detailed collaborative execution plans, defines the execution order, synchronization points, verification steps, and rollback strategies for change operations, ensuring that complex network topology changes can be executed safely and reliably.

[0189] Step S53: According to the collaborative execution plan, the instructions in the network topology change instruction sequence are quickly distributed and executed through an optimized message queue and parallel processing mechanism. The time window technology is used to control the change risk and obtain execution status feedback data.

[0190] Typically, efficient execution of network topology change commands requires advanced distribution and processing mechanisms. Specifically, this embodiment employs an optimized message queue architecture for command distribution. This architecture, based on a publish-subscribe model, features high throughput and low latency. The central controller publishes commands to the queue, and the relevant execution agents subscribe to and consume commands from the queue.

[0191] This embodiment incorporates several optimizations to the message queue, including message compression, batch processing, and priority sorting, to improve transmission efficiency and processing throughput. Simultaneously, it employs an intelligent parallel processing mechanism that analyzes dependencies between instructions, executing independent task groups in parallel while ensuring that dependent tasks are executed sequentially in the correct order. For example, link adjustments in different regions can be performed in parallel, while route updates must be executed only after link adjustments are complete.

[0192] This embodiment also employs time window technology to control change risks, breaking down large changes into multiple smaller batches, executing them within a predetermined time window, and setting verification checkpoints between each batch. If an anomaly is detected, the execution of subsequent batches can be paused for problem diagnosis and correction.

[0193] During execution, real-time feedback on the execution status of each node is collected, including command execution results, configuration verification results, and performance metric changes. This feedback data is structured to form a detailed execution status database for subsequent analysis and decision-making. This efficient distribution and execution mechanism enables large-scale network topology changes to be completed within milliseconds to seconds, significantly improving network adaptability.

[0194] Step S54: Based on the execution status feedback data and the network topology change instruction sequence, perform dynamic scheduling of computing, storage and network resources, apply a load balancing algorithm to optimize traffic allocation, and obtain a resource scheduling result report.

[0195] After the topology change is completed, resource allocation needs to be adjusted to fully leverage the advantages of the new topology. First, the execution status feedback data is analyzed to assess whether the actual effect of the change meets expectations, and resource allocation decisions are made accordingly.

[0196] For computing resources, based on the location and connectivity characteristics of nodes in the new topology, the deployment locations of virtual machines, containers, or microservices are adjusted to distribute the computing load to appropriate processing nodes and reduce cross-network communication. For storage resources, based on access pattern analysis and the characteristics of the new topology, data sharding strategies and cache locations are optimized, placing frequently accessed data on nodes with superior network locations to improve data access efficiency.

[0197] For network resources, the system applies a load balancing algorithm to optimize traffic allocation. The implemented load balancing algorithm comprehensively considers factors such as link capacity, current load, and network latency, and dynamically allocates traffic based on a weighted round-robin and least-connection strategy.

[0198] Furthermore, this embodiment also applies path optimization algorithms based on traffic engineering, such as the improved Open Shortest Path First (OSPF) weighted optimization algorithm, which guides traffic to take the optimal path and avoids congestion points by adjusting route weights. For environments supporting Software-Defined Networking (SDN), the system utilizes centralized control features to implement more granular traffic control, such as application-type-based path selection and service-level resource reservation. It is understood that all resource scheduling decisions and execution results are recorded in a detailed resource scheduling result report, including information such as resource allocation changes, load distribution status, and route adjustment effects.

[0199] Step S55: Based on the resource scheduling result report and the execution status feedback data, deploy real-time monitoring points to collect network performance data after the topology change, compare and analyze the performance indicators predicted in the scheme performance prediction report, and verify the optimization effect to obtain the topology adjustment result and performance monitoring data.

[0200] To verify the actual effects of topology adjustments and resource scheduling, comprehensive performance monitoring and verification analysis are required.

[0201] In step S55, real-time monitoring points are first deployed on key nodes and links. These monitoring points are based on lightweight proxy technology and can collect fine-grained performance data at high frequency, including indicators such as bandwidth utilization, packet forwarding rate, queue depth, latency jitter and error rate.

[0202] In this embodiment, the system also deploys end-to-end performance probes to simulate real user traffic and measure end-to-end service quality indicators. The data collected from these monitoring points is aggregated and processed in real time to form a complete performance dataset. Subsequently, a three-way comparative analysis is performed between the actual monitored performance data, the baseline data before optimization, and the expected values ​​in the solution performance prediction report.

[0203] For example, the degree of improvement for various indicators is calculated, such as the percentage reduction in latency, the increase in bandwidth utilization, and the improvement in availability. This embodiment also analyzes the deviation between the actual results and the predicted values, evaluates the accuracy of the model predictions, and identifies model parameters or assumptions that need to be adjusted.

[0204] For metrics that fail to meet expectations, the system performs root cause analysis to identify potential problems and propose further optimization suggestions. Typically, all analysis results are integrated into a topology adjustment results and performance monitoring data report. This report not only demonstrates the current optimization effects but also includes long-term trend analysis and suggestions for future optimization directions. These data and analysis results are used to continuously optimize the decision-making model, improving model accuracy and decision quality through a feedback learning mechanism, thus forming a closed-loop optimization system.

[0205] In summary, through steps S51-S55, this embodiment realizes the entire process from optimization scheme to actual network adjustment. It employs a distributed collaborative mechanism and a two-phase commit protocol to ensure the secure and reliable execution of complex changes. Combined with dynamic resource scheduling and effect verification analysis, it achieves efficient adaptive optimization of the network topology. This method not only solves the consistency and efficiency problems in traditional network change methods but also supports continuous network optimization through a closed-loop feedback mechanism, enabling the network to better adapt to business needs and environmental changes.

[0206] In this embodiment, step S12, determining the sampling frequency of each node based on the node load characteristics, specifically includes: Based on the node types in the probe deployment scheme, nodes are divided into a first type of node and a second type of node according to their hierarchical position in the network topology and their traffic carrying capacity. The first type of node is a high-load node carrying backbone traffic, and the second type of node is a low-load node carrying access traffic. The sampling frequency for the first type of node is set to the millisecond level, and the sampling frequency for the second type of node is dynamically adjusted to the second level to the minute level according to the changes in node load, thereby obtaining the original monitoring data of the multi-layer network.

[0207] Specifically, in complex network environments, different nodes have significantly different importance and load characteristics, making it neither economical nor reasonable to use a uniform sampling frequency.

[0208] Based on this, this embodiment proposes a differentiated sampling strategy based on node characteristics to balance monitoring accuracy and system overhead. First, network nodes are classified and evaluated based on node type information in the probe deployment scheme. The hierarchical location information of each node is extracted from the network topology database, including the identifiers of the core layer, aggregation layer, and access layer. Simultaneously, historical traffic statistics are analyzed to calculate indicators such as average traffic capacity, peak traffic, and traffic fluctuation characteristics for each node. By comprehensively considering the importance of the node's position in the topology and its traffic processing capacity, network nodes are divided into two main categories.

[0209] Specifically, the first type of node consists of high-load nodes carrying backbone traffic, typically including core routers, core switches, and boundary devices in major data centers. These nodes occupy central positions in the network topology, exhibiting high connectivity (i.e., connected to multiple other nodes) and high betweenness centrality (the frequency with which a node is part of the shortest path between other nodes in the network). In terms of traffic characteristics, the first type of node typically handles 60%-80% of the total network traffic, maintains a consistently high link utilization rate (e.g., an average of 30%-70%), and experiences frequent and drastic traffic fluctuations. Changes in the state of these nodes have a significant impact on overall network performance; any minor performance fluctuation or anomaly can quickly spread and affect a large number of users' experiences.

[0210] Typically, the second type of nodes are low-load nodes that carry access traffic, mainly including access layer switches, edge routers, and terminal access points. These nodes are usually located at the edge of the network topology, with low connectivity, and are mainly responsible for data access and local forwarding in specific areas or departments. In terms of traffic characteristics, the traffic handled by each of these second-type nodes accounts for a small proportion of the total network traffic (usually no more than 5%), link utilization is generally low (e.g., an average of 10%-30%), and traffic patterns are relatively stable, changing slowly and exhibiting a degree of predictability. Changes in the state of these nodes typically only affect the network performance of a local area, with a limited impact.

[0211] Based on the above classification, differentiated sampling frequency strategies are adopted for different types of nodes. Specifically, for the first type of high-load backbone nodes, the system is configured with high-frequency sampling to accurately capture minute performance fluctuations. Specifically, a precise millisecond-level sampling frequency is used, typically set to a sampling interval of 5 to 50 milliseconds.

[0212] This embodiment further subdivides nodes based on their specific importance. For example, critical core routing nodes may use ultra-high-frequency sampling of 5-10 milliseconds, while general backbone nodes use high-frequency sampling of 20-50 milliseconds. This high-frequency sampling can capture transient network events, minor performance fluctuations, and rapidly changing traffic patterns, providing sufficient data support for the early detection and prevention of potential problems.

[0213] For the second type of low-load access nodes, this embodiment adopts an adaptive dynamic sampling strategy, adjusting the sampling frequency in real time based on load changes, rather than using a fixed frequency. This strategy first establishes a load baseline model for each second-type node, describing its normal traffic change patterns and fluctuation range. During runtime, the node's load is continuously monitored and compared with the baseline model, dynamically adjusting the sampling frequency. When a node's load is detected to be in a stable state (the rate of change is below a preset threshold, such as 5% / minute), the sampling frequency is reduced to a lower level, typically once every 30 seconds to 5 minutes, reducing unnecessary monitoring overhead.

[0214] Typically, when a significant change in load is detected (e.g., a change rate exceeding 10% / minute) or a potential abnormal pattern is identified (e.g., sudden traffic spikes or abnormal connection increases), the sampling frequency is automatically increased to the second level (e.g., once every 1-5 seconds) to obtain more detailed status information. When the load changes abnormally drastically or exceeds the warning threshold, the sampling frequency is temporarily increased to the sub-second level (e.g., 200-500 milliseconds) until the status returns to normal.

[0215] Furthermore, this embodiment also employs a temporary sampling strategy adjustment function for specific time windows (such as peak business periods or planned maintenance periods). This allows for the temporary increase of the sampling frequency of certain nodes according to a preset plan or administrator instructions, to address monitoring needs in special circumstances. This embodiment also utilizes an automatic learning mechanism, continuously optimizing the sampling frequency strategy by analyzing historical data patterns. For example, it identifies regular load fluctuations on certain nodes within specific time periods and pre-adjusts the sampling frequency to capture these changes.

[0216] In summary, this intelligent, differentiated sampling strategy significantly reduces overall monitoring overhead while maintaining the accuracy of monitoring key nodes. Compared to a uniform high-frequency sampling strategy, this differentiated strategy can significantly reduce the amount of data collected while maintaining accurate perception of network status. The multi-layered raw network monitoring data generated by this strategy includes both high-precision millisecond-level time-series data from core nodes and dynamic frequency sampling data from edge nodes, providing a comprehensive and resource-efficient data foundation for subsequent data processing and analysis.

[0217] In this embodiment, step S24, which involves learning the dynamic association strength and influence propagation pattern between nodes in the network's dynamic temporal features through a graph attention mechanism and a relational reasoning network to obtain a network node relation matrix, includes: Step S241: Based on the dynamic temporal characteristics of the network, the attention weights between nodes are calculated using the graph attention mechanism to obtain the node attention weight matrix.

[0218] Specifically, nodes in a network are interconnected and influence each other, and understanding these complex relationships is crucial for network optimization. This embodiment applies a graph attention mechanism and a relational reasoning network to capture the dynamic relationships between nodes, making the interaction patterns implicit in time-series data explicit into a relational matrix that can be used for decision-making.

[0219] First, based on the network dynamic temporal features extracted in the previous steps, the Graph Attention Mechanism (GAT) is applied to calculate the attention weights between nodes, resulting in a node attention weight matrix. The Graph Attention Mechanism implements attention calculation between nodes within a graph neural network framework. It allows each node to selectively aggregate information from its neighbors, rather than simply averaging or weighted summing. Compared to traditional graph neural networks, GAT can adaptively learn the importance weights between nodes, more accurately capturing complex relationships.

[0220] Secondly, this embodiment constructs a multi-head graph attention network. This network first converts the dynamic temporal features of each node into an initial node representation vector. Then, it computes the attention coefficients between nodes in parallel using multiple attention heads. Each attention head independently learns a set of attention weights to capture the relationship patterns between nodes from different perspectives.

[0221] Specifically, for each pair of connected nodes i and j in the network, the system calculates the attention coefficient αij between them, which represents the degree of influence of node j on node i. Typically, the calculation process includes the following main steps: First, the node feature vectors are mapped to a shared feature space through a linear transformation; second, an attention mechanism is applied to each pair of nodes to calculate the original attention score; finally, these scores are normalized into weight coefficients using the softmax function.

[0222] To enhance the model's expressive power, this embodiment also employs a multi-head attention mechanism, typically using 8-16 attention heads. Each head independently calculates a set of attention weights, and the outputs of all heads are concatenated or averaged to obtain the final node representation. This multi-head design allows the model to simultaneously focus on different types of node relationships; for example, some heads may focus on capturing traffic correlations, while others may focus on delay propagation patterns.

[0223] Specifically, this embodiment introduces a temporal-aware attention mechanism to consider the temporal dependencies of node states. Currently, traditional GAT only considers spatial attention, ignoring the temporal factor. Specifically, temporal-aware attention integrates temporal encoding into attention computation, enabling the model to distinguish the interaction intensity at different points in time. For example, some nodes may be strongly correlated during peak periods but almost unrelated during trough periods. This mechanism effectively captures this time-varying characteristic through multi-point attention computation within a time window and a time decay function.

[0224] Through the above steps, this embodiment calculates a set of attention weights for each pair of nodes in the network, forming a node attention weight matrix. This matrix is ​​an N×N matrix (N is the number of nodes), and each element αij represents the influence weight of node j on node i. These weights are directional, meaning that αij is not necessarily equal to αji, reflecting the asymmetry of node influence in the network.

[0225] Step S242: Based on the node attention weight matrix, model the influence propagation path between nodes through the relational reasoning network to obtain the network node relation matrix containing dynamic association strength and influence propagation mode.

[0226] In step S242, based on the node attention weight matrix, the influence propagation path between nodes is modeled using a Relation Reasoning Network (RRN) to obtain a network node relationship matrix containing dynamic association strength and influence propagation patterns. The Relation Reasoning Network is a neural network architecture that can be used to infer complex relationships between entities, and it can learn implicit causal relationships and influence propagation rules from observed interaction patterns.

[0227] In this step, the implemented relational reasoning network employs a message passing framework, capturing remote dependencies and complex influence paths through multiple iterative updates. In each iteration, nodes exchange messages containing node state information and relational features. By incorporating an attention weight matrix to guide the message passing process, edges with higher attention weights transmit more information. Unlike traditional graph neural networks, this reasoning network focuses on explicitly modeling the relationships themselves, rather than just node features.

[0228] This embodiment employs a hierarchical influence propagation modeling method to capture both direct and indirect influences. Direct influence refers to the immediate impact between adjacent nodes, calculated directly through a single-hop message passing; indirect influence refers to the impact propagated through intermediate nodes, requiring multiple hop message passing to capture. The system balances computational complexity and model expressiveness by controlling the number of message passing iterations (typically set to 3-5 rounds).

[0229] To enhance the model's understanding of complex network dynamics, the system also introduces a Causal Inference Module. This module identifies potential causal relationships by comparing the temporal correlation and conditional independence of node state changes. For example, if the performance change of node A always precedes that of node B, and this correlation remains significant after controlling for other factors, the system infers that A may have a causal influence on B. The system employs a counterfactual analysis-based approach to assess the strength of causality, quantifying the magnitude of the causal effect by simulating counterfactual scenarios such as how node B would change if node A's state remained unchanged.

[0230] Through the above relationship reasoning process, this embodiment generates the final network node relationship matrix, which not only includes the dynamic association strength between nodes, but also encodes the direction, path and pattern of influence propagation.

[0231] Specifically, the relationship matrix is ​​a multidimensional tensor, where the relationship between each pair of nodes is represented by an eigenvector, typically containing the following key dimensions: Association strength: quantifies the degree of correlation between the state changes of the two nodes, with values ​​typically ranging from [0,1]; Direction of influence: indicates the main flow of influence, distinguishing between the source node and the target node; Time delay characteristics: describes the time characteristics of influence propagation, such as average propagation delay and delay distribution; Type of influence: classifies different types of influence relationships, such as performance correlation, fault propagation, or resource competition; Stability index: assesses the time stability and environmental sensitivity of the relationship.

[0232] Furthermore, this embodiment generates influence and vulnerability scores for each node. The influence score reflects a node's overall ability to influence other nodes, while the vulnerability score indicates how sensitive a node is to the influence of other nodes. These scores are calculated by analyzing the in-degree and out-degree relationships in the relationship matrix, providing crucial reference for subsequent network optimization.

[0233] In summary, this embodiment, by applying graph attention mechanisms and relational reasoning networks, can transform complex interaction patterns implicit in network time-series data into structured relational knowledge, providing a strong analytical foundation for network topology optimization and anomaly detection. Compared with traditional correlation analysis, this method can capture deeper causal relationships and propagation dynamics, significantly improving the interpretability and predictability of network behavior.

[0234] In this embodiment, step S52, which involves designing a rollback mechanism to handle node execution failures and obtaining a collaborative execution plan, specifically includes: Based on the two-phase commit protocol, a node execution status monitoring mechanism is set up. When a node execution failure is detected, a distributed rollback instruction is triggered to restore the node that has undergone changes to its state before the change, thereby obtaining a system state consistency guarantee result. The system state consistency guarantee result is then integrated into the collaborative execution plan.

[0235] Specifically, in complex network environments, topology changes are a high-risk operation; failure of any node can lead to inconsistent network states, resulting in service interruptions or performance degradation. This embodiment uses a distributed rollback mechanism based on a two-phase commit protocol to ensure network consistency and availability even in the event of partial node failure.

[0236] First, this embodiment establishes a node execution status monitoring mechanism based on the Two-Phase Commit Protocol (2PC). The Two-Phase Commit Protocol is a classic protocol for achieving transaction atomicity in distributed systems, comprising a preparation phase and a commit phase. In scenarios involving network topology changes, the system has been specifically optimized for this protocol to adapt to the characteristics and constraints of network devices.

[0237] Specifically, the system employs a three-layer monitoring architecture to track node execution status in real time. The first layer is device-level monitoring, which involves deploying lightweight monitoring agents on each network device to collect command execution status, configuration change results, and device health indicators in real time. These agents maintain a heartbeat connection with the central control system through a secure channel and periodically (typically at 100-500 millisecond intervals) report their local status.

[0238] The second layer is controller-level monitoring, where the distributed controller is responsible for aggregating and analyzing the status reports of devices within the region, and performing preliminary anomaly detection and classification. The controller maintains a status mapping table for all devices within the region, recording the current execution stage, heartbeat status, and health score of each device.

[0239] The third layer is global coordination-level monitoring, where a central coordinator integrates the status reports of all controllers, maintains a global view, and coordinates cross-regional status consistency checks.

[0240] This embodiment employs a sophisticated status determination mechanism, subdividing device execution status into multiple levels, including normal execution, execution warning, execution delay, partial failure, complete failure, and communication interruption. The system determines status through multi-dimensional indicator analysis, including command response time, configuration verification results, system resource indicators (such as CPU utilization and memory usage), and network connection status. The system also sets adaptive timeout thresholds, automatically adjusting determination parameters based on device type, operational complexity, and historical performance to avoid false alarms or missed alarms.

[0241] Furthermore, this embodiment employs a predictive failure detection algorithm, which analyzes time-series patterns in device status to predict potential execution failure risks. For example, if the system observes a continuous increasing trend in device response time or a sharp rise in memory usage, it may indicate an impending execution failure. This predictive detection enables the system to take preventative measures before an actual failure occurs, such as pausing subsequent command execution or preparing backup resources.

[0242] In this embodiment, when a node execution failure is detected, a distributed rollback instruction is triggered to restore the node that has undergone changes to its state before the change. It is understood that the rollback process is a complex collaborative operation, employing multiple technologies to ensure its reliability and efficiency.

[0243] First, this embodiment maintains a detailed change log database, recording all configuration change operations performed by each node, including command content, execution time, original configuration, and result status. This log uses a distributed storage architecture to ensure data integrity even in the event of a partial system crash. For each change record, a corresponding rollback command is automatically generated. These commands are organized into rollback scripts, arranged in the reverse order of their original execution, ensuring a safe restoration to the previous state.

[0244] Specifically, this embodiment employs a layered rollback strategy, prioritizing the recovery of critical equipment and core functions. The rollback process optimizes and prioritizes operations based on topology and service dependencies to ensure network functionality is restored as quickly as possible. For example, routing protocol configurations and core connections are restored first, followed by security policies and QoS configurations, and finally monitoring and management functions. This embodiment utilizes parallel rollback technology, allowing simultaneous rollback operations for independent devices or areas, significantly shortening recovery time.

[0245] To handle potential failures during the rollback process itself, this embodiment employs a two-level rollback mechanism (NestedRollback). When a rollback command failure is detected, a more basic recovery process is initiated, such as restarting the interface, reloading the configuration file, or, in extreme cases, triggering a device reboot. The system also maintains an emergency recovery configuration library, storing verified baseline configurations that can be directly applied to quickly restore basic functionality in the event of a complex rollback failure.

[0246] Meanwhile, this embodiment ensures the consistency of rollback decisions through distributed consensus algorithms (such as Paxos or Raft), achieving globally consistent rollback decisions even in the event of network partitions or controller failures. All rollback operations undergo strict access control and audit logging to ensure security and traceability.

[0247] Specifically, after the rollback operation is completed, a comprehensive consistency verification is performed to ensure that all nodes have successfully recovered to the consistent state before the change, thus obtaining a system state consistency guarantee result. Typically, the verification process includes three levels: configuration consistency check, connectivity test, and functional verification.

[0248] The configuration consistency check compares the current configuration of the device with the baseline configuration before the change to ensure that key parameters are fully matched; the connectivity test verifies whether the connectivity between network nodes has been restored to normal, including physical connection and logical path testing; and the functional verification checks whether key network functions are operating normally, such as route convergence, load balancing, and security policy implementation.

[0249] In summary, this embodiment generates a detailed consistency verification report. Typically, the verification report includes the recovery status of each node, the verification results, and a list of potential issues. For nodes that pass verification, the report confirms successful recovery; for nodes with inconsistencies, the report records the discrepancies in detail and provides resolution suggestions. The system also calculates the overall recovery rate and service availability metrics to assess the overall success of the rollback.

[0250] Finally, the system state consistency assurance results are integrated into the collaborative execution plan, forming a complete execution framework with fault handling capabilities. Typically, the integration process includes not only incorporating rollback strategies as part of the execution plan but also optimizing the original execution plan based on rollback experience. By analyzing historical rollback data, high-risk operations and failure-prone points are identified, and the execution order and methods are adjusted to reduce the probability of failure. The system sets multiple checkpoints in the execution plan to verify the execution status and confirm whether to continue at critical stages, enhancing the plan's robustness. The system also optimizes resource allocation strategies to reserve sufficient system and network resources for possible rollback operations, ensuring that recovery processes can be executed rapidly when needed.

[0251] Through the aforementioned rollback mechanism, this embodiment can securely execute topology changes in complex network environments, maintaining network consistency and service availability even in the event of partial node failures. This advanced fault tolerance capability is a key foundation for achieving network adaptive optimization, providing necessary security guarantees for dynamically adjusting network topology.

[0252] Example 2: like Figure 2 As shown, this embodiment also provides a network topology adaptive optimization system, including: The network status acquisition module 201 is used to acquire network node status data and link load data based on a distributed network probe and a multi-level acquisition strategy, and to perform data standardization, noise filtering and time-series alignment processing on the network node status data and link load data to obtain a standardized network status data stream. The dynamic topology modeling module 202 is used to receive the standardized network state data stream, and to construct a dynamic network topology model by applying a Stancu type neural network operator. The Stancu type neural network operator includes adjustable higher-order terms and nonlinear transformations. The dynamic network topology model is used to capture the dynamic relationships and topological features between network nodes, and to obtain network topology feature vectors and relationship matrices. Anomaly identification module 203 is used to actively detect network weaknesses and potential fault points based on the network topology feature vector and relationship matrix through a perturbation sampling strategy, identify network bottlenecks and anomaly patterns, conduct a comprehensive risk assessment, and obtain a network risk assessment report and an optimization target list. The optimization decision module 204 is used to generate a set of candidate topology adjustment schemes based on the network risk assessment report and the optimization target list, and to comprehensively evaluate each candidate topology adjustment scheme to obtain the optimal topology adjustment scheme. The topology reconfiguration module 205 is used to receive the optimal topology adjustment scheme, convert the optimal topology adjustment scheme into a network topology change instruction sequence, realize the synchronous execution of configuration changes among multiple nodes through a distributed coordination mechanism, complete the distribution and execution of topology reconfiguration instructions, and simultaneously perform dynamic resource scheduling and route updates to obtain topology adjustment results and performance monitoring data, and continuously optimize the decision model based on the topology adjustment results and performance monitoring data.

[0253] It should be noted that each functional module of the system corresponds one-to-one with the corresponding steps in the aforementioned method embodiment 1. Through the collaborative work of the above modules, this embodiment introduces a Stancu type neural network operator to construct a dynamic network topology model, actively detects network weaknesses by combining a perturbation sampling strategy, and uses a distributed collaborative mechanism to achieve millisecond-level topology reconstruction, thereby improving the accuracy, speed and reliability of network topology optimization.

[0254] The above description is merely a specific implementation of the embodiments of this disclosure, but the protection scope of the embodiments of this disclosure is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the embodiments of this disclosure should be included within the protection scope of the embodiments of this disclosure. Therefore, the protection scope of the embodiments of this disclosure should be determined by the protection scope of the claims.

Claims

1. A network topology adaptive optimization method, characterized in that, include: Based on distributed network probes, network node status data and link load data are acquired through a multi-level acquisition strategy. The network node status data and link load data are then processed by data standardization, noise filtering and time-series alignment to obtain a standardized network status data stream. The standardized network state data stream is received, and a dynamic network topology model is constructed using a Stancu-type neural network operator. The Stancu-type neural network operator includes adjustable higher-order terms and nonlinear transformations. The dynamic network topology model is used to capture the dynamic relationships and topological features between network nodes, and to obtain network topology feature vectors and relationship matrices. Based on the network topology feature vector and relation matrix, network weaknesses and potential fault points are actively detected through a perturbation sampling strategy, network bottlenecks and abnormal patterns are identified, a comprehensive risk assessment is conducted, and a network risk assessment report and optimization target list are obtained. Based on the network risk assessment report and the optimization target list, a set of candidate topology adjustment schemes is generated using a multi-objective optimization algorithm. The optimal topology adjustment scheme is obtained by comprehensively evaluating each candidate topology adjustment scheme. The system receives the optimal topology adjustment scheme, converts it into a network topology change instruction sequence, and uses a distributed coordination mechanism to achieve synchronous execution of configuration changes among multiple nodes. This completes the distribution and execution of topology reconstruction instructions, while simultaneously performing dynamic resource scheduling and route updates. The system obtains topology adjustment results and performance monitoring data, and continuously optimizes the decision model based on these results and data.

2. The method according to claim 1, characterized in that, The distributed network probe-based system acquires network node status data and link load data through a multi-level acquisition strategy. The network node status data and link load data undergo data standardization, noise filtering, and time-series alignment processing to obtain a standardized network status data stream, including: Based on the network topology and node distribution, intelligent algorithms are used to determine the probe deployment locations and deploy monitoring agents on network nodes at different levels to obtain a probe deployment scheme that covers the entire network. Based on the probe deployment scheme, differentiated sampling frequencies and monitoring granularities are set, and the sampling frequency of each node is determined according to the node load characteristics to obtain multi-level network raw monitoring data. The system receives the original monitoring data of the multi-level network and processes it through unit conversion, dimension normalization and feature scaling to unify the network data from different sources and of different types into a standard format, thereby obtaining a standardized network indicator dataset. Based on the standardized network index dataset, noise and outliers are removed by applying sliding window midpoint filtering and wavelet transform methods to obtain cleaned network state data. Based on the cleaned network status data, time alignment is achieved through timestamp correction and interpolation algorithms. Multi-source data fusion technology is used to correlate and analyze the node status and link load data in the cleaned network status data to obtain the standardized network status data stream.

3. The method according to claim 1, characterized in that, The process involves receiving the standardized network state data stream, constructing a dynamic network topology model using a Stancu-type neural network operator (which includes adjustable higher-order terms and nonlinear transformations), and capturing the dynamic relationships and topological features between network nodes through the dynamic network topology model to obtain a network topology feature vector and a relationship matrix, including: Based on the requirements of network topology optimization, a mathematical model of a Stancu-type neural network operator is constructed. The adjustable higher-order terms and nonlinear transformations are introduced to obtain the framework of a Stancu-type neural network operator. Based on the Stancu type neural network operator framework and the standardized network state data stream, a graph structure representation of the network topology is constructed. The embedding vectors of nodes and links in the graph structure representation are learned by the neural network to obtain a low-dimensional feature representation of the network elements. Based on the low-dimensional feature representation of the network elements, a hybrid model of temporal convolutional network and long short-term memory network is applied to capture the temporal change pattern of the network state in the low-dimensional feature representation of the network elements, thereby obtaining the network dynamic temporal features. Based on the dynamic temporal characteristics of the network, the dynamic association strength and influence propagation mode between nodes in the dynamic temporal characteristics of the network are learned through graph attention mechanism and relation reasoning network to obtain the network node relation matrix. Based on the network node relationship matrix, the low-dimensional feature representation of the network elements, and the network dynamic temporal features, a multi-level fusion network is used to integrate the static topology structure and dynamic state features. The computational efficiency is optimized by applying model distillation technology to obtain the network topology feature vector and relationship matrix.

4. The method according to claim 1, characterized in that, Based on the network topology feature vector and relationship matrix, a perturbation sampling strategy is used to actively detect network weaknesses and potential fault points, identify network bottlenecks and abnormal patterns, and conduct a comprehensive risk assessment to obtain a network risk assessment report and an optimization target list, including: Based on the network topology feature vector and relation matrix, a multi-level progressive perturbation sampling strategy is designed, which includes micro-perturbation, moderate perturbation and strong perturbation, to obtain the network perturbation sampling execution plan; Based on the network disturbance sampling execution plan, disturbance injection is executed collaboratively by a distributed controller. The network response data to the disturbance is collected in real time during the execution of the network disturbance sampling execution plan, and the state changes of each node and link are recorded to obtain the disturbance response dataset. Based on the disturbance response dataset and the network topology feature vector and relation matrix, gradient analysis and influence propagation model are used to identify the network nodes and links in the disturbance response dataset that are sensitive to disturbances, assess the vulnerability of the network nodes and links that are sensitive to disturbances, and obtain a network sensitivity map. Based on the network sensitive point map and the standardized network status data stream, clustering algorithms and anomaly detection models are applied to detect and classify network bottlenecks in the network sensitive point map, distinguishing between capacity bottlenecks, configuration bottlenecks, and topology bottlenecks, and obtaining a network bottleneck analysis report. Based on the network bottleneck analysis report and the network sensitive point map, a comprehensive risk assessment is conducted in conjunction with business importance and service level agreement requirements. Priorities are assigned to the risk points identified in the network bottleneck analysis report and the network sensitive point map, and corresponding optimization targets are generated, resulting in the network risk assessment report and the optimization target list.

5. The method according to claim 1, characterized in that, Based on the network risk assessment report and the optimization target list, a set of candidate topology adjustment schemes is generated using a multi-objective optimization algorithm. Each candidate topology adjustment scheme is comprehensively evaluated to obtain the optimal topology adjustment scheme, including: Based on the network risk assessment report and the list of optimization objectives, the optimization objectives in the list of optimization objectives are transformed into mathematical objective functions and weight coefficients are set for each objective to obtain a formalized definition of a multi-objective optimization problem. Based on the formalized definition of the multi-objective optimization problem and the network topology feature vector and relation matrix, network physical constraints, service constraints and resource constraints are identified, and the network physical constraints, service constraints and resource constraints are transformed into mathematical constraints to obtain the optimization problem constraint model; Based on the optimization problem constraint model and the formalized multi-objective optimization problem definition, a multi-objective evolutionary algorithm and Pareto optimal front technique are applied to generate topology adjustment schemes for different combinations of objectives in the formalized multi-objective optimization problem definition, thereby obtaining the candidate topology adjustment scheme set. Based on the candidate topology adjustment scheme set, each scheme in the candidate topology adjustment scheme set is simulated and tested using a network simulation engine. The performance indicators of each scheme under different load conditions are predicted and potential risks and benefits are evaluated to obtain a scheme performance prediction report. Based on the performance prediction report of the proposed scheme and the weight coefficients in the formalized multi-objective optimization problem definition, a multi-criteria decision analysis method is applied to comprehensively evaluate each candidate topology adjustment scheme in the candidate topology adjustment scheme set and select the best balance point, generating a detailed implementation plan to obtain the optimal topology adjustment scheme.

6. The method according to claim 5, characterized in that, The process involves receiving the optimal topology adjustment scheme, converting it into a network topology change instruction sequence, and using a distributed coordination mechanism to achieve synchronous execution of configuration changes among multiple nodes. This completes the distribution and execution of topology reconstruction instructions, while simultaneously performing dynamic resource scheduling and route updates, resulting in topology adjustment results and performance monitoring data, including: Based on the optimal topology adjustment scheme, the optimization decisions in the optimal topology adjustment scheme are converted into a network configuration instruction set to ensure the atomicity and consistency of the instructions in the network configuration instruction set, thereby obtaining the network topology change instruction sequence; Based on the network topology change instruction sequence, a distributed transaction management framework is constructed. A two-phase commit protocol is used to ensure the synchronization of configuration changes among multiple nodes, and a rollback mechanism is designed to deal with the situation of node execution failure, thus obtaining a collaborative execution plan. According to the collaborative execution plan, the instructions in the network topology change instruction sequence are rapidly distributed and executed through optimized message queues and parallel processing mechanisms, and the change risk is controlled by time window technology to obtain execution status feedback data. Based on the execution status feedback data and the network topology change instruction sequence, dynamic scheduling of computing, storage and network resources is performed, and a load balancing algorithm is applied to optimize traffic allocation to obtain a resource scheduling result report. Based on the resource scheduling result report and the execution status feedback data, real-time monitoring points are deployed to collect network performance data after the topology change. The data is then compared and analyzed with the performance indicators predicted in the scheme performance prediction report to verify the optimization effect, thereby obtaining the topology adjustment results and performance monitoring data.

7. The method according to claim 2, characterized in that, The step of determining the sampling frequency of each node based on node load characteristics includes: Based on the node types in the probe deployment scheme, nodes are divided into a first type of node and a second type of node according to their hierarchical position in the network topology and their traffic carrying capacity. The first type of node is a high-load node carrying backbone traffic, and the second type of node is a low-load node carrying access traffic. The sampling frequency for the first type of node is set to the millisecond level, and the sampling frequency for the second type of node is dynamically adjusted to the second level to the minute level according to the changes in node load, thereby obtaining the original monitoring data of the multi-layer network.

8. The method according to claim 3, characterized in that, The method involves learning the dynamic association strength and influence propagation pattern between nodes in the network's dynamic temporal features through graph attention mechanisms and relational reasoning networks, resulting in a network node relation matrix, including: Based on the dynamic temporal characteristics of the network, the attention weights between nodes are calculated using a graph attention mechanism to obtain the node attention weight matrix. Based on the node attention weight matrix, the influence propagation path between nodes is modeled through a relational reasoning network to obtain the network node relation matrix containing dynamic association strength and influence propagation mode.

9. The method according to claim 6, characterized in that, The design rollback mechanism addresses node execution failures and generates a collaborative execution plan, including: Based on the two-phase commit protocol, a node execution status monitoring mechanism is set up. When a node execution failure is detected, a distributed rollback instruction is triggered to restore the node that has undergone changes to its state before the change, thereby obtaining a system state consistency guarantee result. The system state consistency guarantee result is then integrated into the collaborative execution plan.

10. A network topology adaptive optimization system, characterized in that, include: The network status acquisition module is used to acquire network node status data and link load data based on distributed network probes and through a multi-level acquisition strategy. The network node status data and link load data are then processed for data standardization, noise filtering and time-series alignment to obtain a standardized network status data stream. The dynamic topology modeling module is used to receive the standardized network state data stream, apply a Stancu type neural network operator to construct a dynamic network topology model, the Stancu type neural network operator includes adjustable higher-order terms and nonlinear transformations, and capture the dynamic relationships and topological features between network nodes through the dynamic network topology model to obtain network topology feature vectors and relationship matrices. An anomaly identification module is used to actively detect network weaknesses and potential fault points based on the network topology feature vector and relation matrix, identify network bottlenecks and anomaly patterns, conduct a comprehensive risk assessment, and obtain a network risk assessment report and an optimization target list. The optimization decision module is used to generate a set of candidate topology adjustment schemes based on the network risk assessment report and the optimization target list, and to comprehensively evaluate each candidate topology adjustment scheme to obtain the optimal topology adjustment scheme. The topology reconfiguration module is used to receive the optimal topology adjustment scheme, convert the optimal topology adjustment scheme into a network topology change instruction sequence, realize the synchronous execution of configuration changes among multiple nodes through a distributed coordination mechanism, complete the distribution and execution of topology reconfiguration instructions, and simultaneously perform dynamic resource scheduling and route updates to obtain topology adjustment results and performance monitoring data, and continuously optimize the decision model based on the topology adjustment results and performance monitoring data.

Images