Software-defined wide area network data transmission method and system based on edge computing

Abstract

The application relates to the technical field of software-defined wide area network data transmission, and discloses a software-defined wide area network data transmission method and system based on edge computing. The method collects network state data through a distributed probe arranged at an edge node, calculates a dynamic transmission path between nodes after data cleaning and feature extraction, generates a transmission optimization strategy by analyzing historical and real-time data through a collaborative filtering algorithm, generates a dynamic routing strategy in combination with a fuzzy logic decision maker, predicts future congestion states based on network features and the routing strategy, and adjusts flow scheduling parameters accordingly. Finally, the network state data, the routing strategy and the scheduling parameters are visually displayed and stored in time sequence. The method can effectively reduce transmission delay and improve network resource utilization.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of software-defined wide area network data transmission, in particular to a software-defined wide area network data transmission method and system based on edge computing. BACKGROUND

[0002] Software-defined wide area network (SD-WAN) technology separates the control plane from the data plane, enabling centralized and flexible network management. Traditional SD-WAN solutions mainly use centralized controllers for path calculation and policy distribution. This approach faces many challenges in large-scale network environments. In a centralized control architecture, all network state data needs to be transmitted to the central controller for processing, resulting in additional transmission overhead and decision-making delays. As the network size expands and business demands grow, this delay significantly impacts real-time applications.

[0003] Existing SD-WAN systems typically use static or semi-static routing strategies, making it difficult to adapt to dynamic changes in network state. Although some solutions introduce prediction mechanisms based on historical data, due to the complexity and variability of network environments, these predictions often have limited accuracy. In addition, traditional congestion control methods are mostly based on reactive mechanisms, adjusting only after congestion occurs. This post-processing approach cannot effectively prevent congestion from occurring.

[0004] In terms of network state monitoring, existing solutions mostly use periodic polling or simple threshold alarm mechanisms, lacking deep extraction and analysis of network characteristics. This results in the system being unable to timely discover potential network problems and also being unable to provide accurate network state predictions. In the data transmission process, fixed traffic scheduling parameters cannot adapt to dynamic changes in network state, often resulting in uneven resource allocation or low utilization.

[0005] In addition, existing SD-WAN systems also have deficiencies in data visualization and storage. Most systems only provide simple network status displays, lacking deep visualization capabilities for the entire data transmission process. In terms of data storage, relational databases are often used, making it difficult to effectively handle the efficient storage and fast retrieval of time series data. SUMMARY

[0006] The present application aims to provide a software-defined wide area network data transmission method and system based on edge computing to solve the problems raised in the background.

[0007] To achieve the above-mentioned purpose, the present application provides a software-defined wide area network data transmission method and system based on edge computing, the method comprising:

[0008] collecting network status data through distributed probes deployed at SD-WAN edge nodes, cleaning and feature extraction of the network status data;

[0009] calculating dynamic transmission paths between SD-WAN nodes according to the extracted network feature data;

[0010] generating transmission optimization strategies by analyzing historical transmission data and real-time network feature data using a collaborative filtering algorithm;

[0011] generating dynamic routing strategies through a fuzzy logic decision maker based on the transmission optimization strategies and network feature data;

[0012] predicting future network congestion states according to the dynamic routing strategies and network feature data;

[0013] adjusting traffic scheduling parameters of SD-WAN nodes in combination with the future network congestion states and dynamic routing strategies;

[0014] visualizing the network status data, dynamic routing strategies, and traffic scheduling parameters;

[0015] storing the network status data, dynamic routing strategies, and traffic scheduling parameters in a time series database according to timestamps.

[0016] Preferably, the collecting network status data through distributed probes deployed at SD-WAN edge nodes, cleaning and feature extraction of the network status data comprises:

[0017] deploying a probe cluster at SD-WAN edge nodes and aligning the clocks of all probes using a time synchronization protocol;

[0018] the probe cluster continuously collects link delay, packet loss rate, and node resource utilization rate indicators;

[0019] cleaning outliers in the link delay, packet loss rate, and node resource utilization rate indicators through a median filtering algorithm;

[0020] extracting delay fluctuation features from the cleaned link delay data, extracting packet loss burst features from the packet loss rate data, and extracting resource load trend features from the node resource utilization rate data.

[0021] Preferably, the calculating dynamic transmission paths between SD-WAN nodes comprises:

[0022] obtaining SD-WAN network-wide topology data and node location data;

[0023] analyzing the SD-WAN network-wide topology data and node location data using a graph neural network to generate initial weights of inter-node transmission paths;

[0024] According to the time delay fluctuation characteristics and resource load trend characteristics, the initial weight of the inter-node transmission path is dynamically corrected;

[0025] Based on the corrected inter-node transmission path weight, a dynamic transmission path matrix is constructed.

[0026] Preferably, the use of collaborative filtering algorithm to analyze historical transmission data and real-time network feature data to generate transmission optimization strategy includes:

[0027] Extract historical transmission success rate data and historical network feature data from the time series database;

[0028] Establish a similarity mapping relationship between real-time network feature data and historical network feature data;

[0029] According to the similarity mapping relationship, match the data slice with the highest historical transmission success rate;

[0030] Extract the transmission parameter configuration corresponding to the matched data slice as the transmission optimization strategy basic template;

[0031] Incrementally adjust the transmission optimization strategy basic template in combination with the packet loss burst characteristics.

[0032] Preferably, the dynamic routing strategy generated by the fuzzy logic decision maker includes:

[0033] Define the path weight of the dynamic transmission path matrix as a fuzzy input variable;

[0034] Define the bandwidth allocation requirement in the transmission optimization strategy as a fuzzy rule constraint;

[0035] Use the barycenter method to solve the fuzzification process, and output the optimal path selection probability distribution;

[0036] Generate a dynamic routing strategy table according to the optimal path selection probability distribution.

[0037] Preferably, the prediction of the future network congestion state includes:

[0038] Construct a long short-term memory neural network model, and the input dimension includes resource load trend characteristics, dynamic routing strategy table and historical congestion records;

[0039] Input the resource load trend characteristics and dynamic routing strategy table of the current period into the long short-term memory neural network model which has been trained;

[0040] Output the congestion risk coefficient of each SD-WAN node in the future preset period.

[0041] Preferably, the adjustment of the traffic scheduling parameters of the SD-WAN node includes:

[0042] when the congestion risk coefficient exceeds a preset threshold, starting adjustment of a traffic scheduling parameter;

[0043] redistributing the traffic proportion of the high-congestion-risk node according to the dynamic routing strategy table;

[0044] dynamically setting a smooth transition window of the traffic scheduling parameter based on the packet loss burst feature.

[0045] Preferably, the visualized display of the network state data, the dynamic routing strategy and the traffic scheduling parameter comprises:

[0046] converting the dynamic transmission path matrix into a topology thermodynamic diagram;

[0047] mapping the congestion risk coefficient into a node color gradient;

[0048] superimposedly displaying a real-time adjustment track of the traffic scheduling parameter.

[0049] Preferably, the timestamp-based classified storage of the network state data, the dynamic routing strategy and the traffic scheduling parameter into a time sequence database comprises:

[0050] attaching a collection terminal device identifier to each piece of network state data;

[0051] attaching a strategy version identifier to the dynamic routing strategy;

[0052] establishing a bidirectional index of the collection terminal device identifier and the strategy version identifier in the order of timestamps.

[0053] Preferably, the present application further comprises an edge computing-based SD-WAN low-delay data transmission system, comprising a memory, a processor and a computer program stored in the memory and running on the processor, characterized in that the processor, when executing the computer program, implements the steps of the above-mentioned edge computing-based SD-WAN low-delay data transmission method.

[0054] Compared with the prior art, the present application has the following beneficial effects:

[0055] The present application realizes the localized collection and processing of network state data through distributed probe deployment, reduces data transmission delay and improves the real-time performance of state awareness. The feature extraction process can deeply analyze network operation features to provide accurate data support for path calculation.

[0056] The dynamic transmission path calculation mechanism dynamically adjusts path weights according to real-time network features, so that the data transmission path can adapt to network state changes. This method improves the accuracy of path selection and avoids the limitations of traditional static path allocation methods.

[0057] The application of collaborative filtering algorithm enables the transmission optimization strategy to fully utilize historical experience data while combining real-time network features for strategy optimization. This data-driven strategy generation method improves the scientificity and effectiveness of the transmission strategy.

[0058] The introduction of fuzzy logic decision maker effectively handles the uncertainty and fuzziness in the network environment, and can generate more dynamic routing strategies that meet actual needs. This decision-making method overcomes the applicability limitations of traditional binary logic in complex network environments.

[0059] The network congestion prediction function can identify potential congestion risks in advance by analyzing multi-dimensional network features and routing strategies. This prediction capability enables the system to take preventive measures to avoid passive adjustments after congestion occurs.

[0060] The dynamic adjustment mechanism of traffic scheduling parameters automatically optimizes the traffic distribution strategy based on prediction results and real-time network status. This adaptive adjustment capability improves network resource utilization and ensures the transmission quality of critical services.

[0061] The visual display function provides an intuitive view of network status, helping operations personnel quickly grasp the network operating conditions. The use of time series databases ensures the complete preservation and efficient retrieval of historical data, providing data support for network performance analysis and optimization decisions.

[0062] The entire method adopts an edge computing architecture, distributing computing tasks to the network edge, reducing the load pressure on the central controller. This distributed processing approach improves the scalability and reliability of the system, enabling it to adapt to different network environment requirements. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 The working principle diagram of the edge computing-based software-defined wide area network data transmission method and system described in the present invention;

[0064] Figure 2 The flowchart for network state data collection and feature extraction;

[0065] Figure 3 The flowchart for dynamic transmission path calculation;

[0066] Figure 4 The flowchart for transmission optimization strategy generation;

[0067] Figure 5 The flowchart for future network congestion state prediction. DETAILED DESCRIPTION

[0068] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0069] With reference to Figure 1 The present application provides a software-defined wide area network data transmission method and system based on edge computing. The method collects network status data through distributed probes deployed in SD-WAN edge nodes, cleanses and extracts features from these data to remove noise and extract key features. Based on the extracted network feature data, dynamic transmission paths between SD-WAN nodes are calculated to ensure that path selection adapts to network changes. Collaborative filtering algorithm is used to analyze historical transmission data and real-time network feature data to generate transmission optimization strategies, thereby optimizing the data transmission process. Fuzzy logic decision maker combines transmission optimization strategies and network feature data to generate dynamic routing strategies to cope with network uncertainty. According to the dynamic routing strategy and network feature data, the future network congestion state is predicted to identify potential problems in advance. Combining the future network congestion state and the dynamic routing strategy, the traffic scheduling parameters of the SD-WAN nodes are adjusted to achieve traffic optimization. Network status data, dynamic routing strategies and traffic scheduling parameters are visualized to provide an intuitive network status view. Finally, these data are stored in a time series database according to the timestamp for subsequent query and analysis.

[0070] Embodiment 1: With reference to Figure 2 In the process of deploying a probe cluster in the SD-WAN edge node, each edge node is installed with dedicated probe software or hardware devices. These probes are configured to run in a distributed manner, covering key paths and nodes in the network. The deployment of the probe cluster takes into account the complexity of the network topology and the distribution of nodes to ensure that there are no blind spots, thereby comprehensively monitoring the network status. The time synchronization protocol uses standard protocols such as NTP or PTP, and the clocks of all probes are aligned to the same time source to avoid time deviation during data collection. This synchronization ensures the accuracy of subsequent data processing, as the timestamps of network status data must be consistent for effective comparison and analysis. The initialization of the probe cluster includes configuring IP addresses, setting sampling intervals and defining data formats. The sampling interval is usually set to milliseconds to adapt to high-speed network environments, but can be dynamically adjusted according to network load, such as increasing the sampling frequency during peak periods and reducing it during low periods to save resources. Communication between probes is carried out through encrypted channels to ensure secure transmission of data to the central processing unit.

[0071] The probe cluster continuously collects link latency, packet loss rate, and node resource utilization indicators. Link latency is measured by sending ICMP echo requests or TCP handshake packets to measure round-trip time. Packet loss rate is calculated based on the difference between the number of sent and received packets. Node resource utilization includes CPU usage, memory usage, and bandwidth utilization. These indicators are obtained in real time through system APIs or custom scripts. The collection process is automated, with the probe polling each node and link in a loop. Data is temporarily stored in local buffers and then uploaded in batches to avoid network congestion. The collection frequency can be adapted based on network conditions, such as automatically increasing the sampling rate when anomalies are detected, ensuring real-time and representative data. Indicator data is packaged in JSON or binary format, including timestamps, node IDs, and measurement values, for easy parsing and processing.

[0072] Anomalies in link latency, packet loss rate, and node resource utilization indicators are cleaned up using a median filtering algorithm. The median filtering algorithm works by selecting a sliding window, typically with a size of 5 to 10 data points. The median of the data within the window is calculated and used to replace the value at the center of the window, smoothing the data and removing sudden noise or outliers. For example, if a link latency data point suddenly spikes, the median filtering will replace it with the median of the surrounding points, reducing false fluctuations. The cleaning process is performed locally on the probe or on the central server, depending on the distribution of computing resources, to avoid delays. Efficient data structures such as double-ended queues are used to optimize performance and ensure real-time processing of high-speed data streams. Cleaned data is marked as valid, while invalid data is logged for subsequent auditing but not used for feature extraction, ensuring the reliability of subsequent analysis.

[0073] Time delay fluctuation features are extracted from cleaned link latency data. Time delay fluctuation features quantify the stability of the delay by calculating the moving standard deviation and the mean absolute deviation of the delay data. The moving standard deviation is calculated within a time window, with the window size set according to network dynamics, such as 30 seconds or 1 minute, reflecting the degree of delay fluctuation. The mean absolute deviation provides another measure for identifying anomalies in long-term trends. The extraction process involves time series analysis, with data processed in segments corresponding to time intervals. Feature values are calculated and stored as vectors for use by machine learning models. Time delay fluctuation features help identify the reliability of network paths, with high fluctuations possibly indicating congestion or link problems.

[0074] Packet loss burst features are extracted from packet loss rate data by detecting the rate of change of packet loss rate within consecutive time windows, such as calculating the first derivative of packet loss rate or using a sliding window to compare the difference between adjacent windows. When the rate of change exceeds a pre-set threshold, it is marked as a burst event. The threshold is dynamically adjusted based on historical data to adapt to different network environments. The extraction process also includes identifying the duration and intensity of the burst, such as through integration or peak detection algorithms. Packet loss burst features are encoded as binary flags or numerical scores representing the immediacy of packet loss risk for subsequent policy adjustments.

[0075] Resource load trend features are extracted from node resource utilization data by analyzing long-term changes in resource utilization using moving averages or exponential smoothing methods. The moving average window size is set to several minutes to several hours, depending on the prediction requirements. Exponential smoothing gives higher weights to recent data to capture trend changes. The extraction process also includes calculating the slope and curvature of the trend to predict future load direction, such as by linear regression fitting resource usage over time to derive load growth rate. Resource load trend features are represented as aggregated values of time series, such as mean, variance, or trend line equation, for network capacity planning and congestion prediction. The entire feature extraction module is integrated into the data processing pipeline and runs automatically without human intervention. Extracted features are cached and transmitted to downstream modules to ensure low-latency processing. Maintenance of the probe cluster includes regular calibration and software updates to maintain the accuracy of data collection. Network administrators can adjust probe parameters such as sampling rate and filtering settings through a configuration interface to adapt to specific network requirements. Algorithm selection for data cleaning and feature extraction is based on a balance between computational efficiency and accuracy, such as median filtering to remove outliers while preserving real changes. Feature extraction methods are optimized to reduce computational overhead, suitable for edge computing environments. Finally, all processed data is structured and stored, ready for dynamic path computation and policy generation.

[0076] In the implementation process of deploying a probe cluster in an SD-WAN edge node, a lightweight data collection agent is installed on each edge node. These agent software are deployed in the form of Docker containers, which occupy few resources and support hot updates. A certain enterprise's SD-WAN network has deployed such probes in 12 edge nodes in three regions of North America, Europe, and Asia, forming a distributed monitoring system. Probe configuration information is issued through a centralized management platform, including parameters such as sampling frequency, data format, and transmission protocol. After deployment, the probes automatically register with the regional coordinator to form a device list and topology mapping relationship. The clocks of all probes are aligned using the Precision Time Protocol (PTP) to achieve microsecond-level synchronization. The master clock server is deployed in the core data center, and boundary clock devices are deployed in each region to reduce the impact of network delay. The synchronization process performs clock calibration every 30 minutes, and probes with a drift exceeding the threshold will automatically trigger a re-synchronization. In a certain deployment, it was found that the clock deviation of the Asian node was large, and this problem was solved by increasing the local time agent server. The clock synchronization state is monitored in real time, and the data of probes that are not synchronized will be marked but still participate in collection to ensure data continuity. The probe cluster continuously collects link delay, packet loss rate, and node resource utilization indicators using a combination of active probing and passive listening. Active probing measures link delay by sending ICMP probe packets, with 10 probe packets sent per second to calculate the average round-trip time. Passive listening captures traffic data through a mirror port, and counts the number of lost packets and bytes per second. Node resource utilization is obtained through system APIs, including CPU usage, memory occupancy, disk IO, and network interface throughput. In a certain financial enterprise's SD-WAN, the data collected by the probes contains more than 50 dimensions, forming a rich monitoring data set. The collection frequency is dynamically adjusted according to network load, with one collection per second in normal periods and five collections per second during peak periods.

[0077] The median filtering algorithm is used to clean up abnormal values in link delay, packet loss rate and node resource utilization indicators. A sliding window processing method is used. The window size is set to 5 sampling points. The median value is taken after sorting the data in the window to replace the current value. During network jitter, the delay data appears a short peak, and the median filtering effectively smooths these abnormal fluctuations. The cleaning algorithm uses a double buffering mechanism. Real-time data is stored in the buffer, and an independent thread is used for processing, which does not affect the data acquisition performance. The cleaned data and the original data are stored together, and the abnormal markers are retained for subsequent analysis. The standard deviation and the coefficient of variation are used to calculate the delay fluctuation characteristics from the cleaned link delay data. The standard deviation of the delay data per minute is calculated to measure the absolute fluctuation degree. The coefficient of variation (ratio of standard deviation to mean) is calculated to evaluate the relative fluctuation level. In the SD-WAN of an e-commerce platform, it is found that the delay fluctuation characteristics of cross-border links are significantly higher than those of domestic links. These characteristic values are quantized to a score value of 0-100. The feature extraction is performed every 5 minutes, and a rolling time window is used to analyze the data trend in the last 30 minutes.

[0078] The change rate detection and duration analysis are used to extract the packet loss burst feature from the packet loss rate data. The packet loss rate change slope between adjacent sampling points is calculated, and when the slope exceeds the threshold, it is marked as the start of the burst. In the monitoring of a video conference system, it is found that the packet loss burst usually lasts for 3-5 sampling periods. These characteristics are recorded as a binary tuple of burst intensity and duration. The feature extraction algorithm sets a minimum burst duration threshold to avoid misjudging noise as a burst event. The moving average and linear regression analysis are used to extract the resource load trend feature from the node resource utilization data. The 5-minute moving average of CPU, memory and bandwidth usage is calculated to eliminate short-term fluctuations. In the SD-WAN of a manufacturing enterprise, linear regression is used to fit the hourly resource usage trend, and the slope value is used as a load change direction indicator. The feature data is updated every 15 minutes, and the historical trend data of the last 24 hours is retained for comparative analysis.

[0079] The entire data collection and feature extraction process forms an automated pipeline. The raw data is cleaned, feature extracted and formatted before being published to the message queue. The downstream system subscribes to these processed feature data for path calculation and policy generation. The system sets up data quality monitoring, which triggers an alarm and starts a backup collection scheme when the feature data is missing. All processing steps are logged for fault diagnosis and performance optimization.

[0080] Example 2: see Figure 3, obtaining SD-WAN full network topology data and node location data is the basis for building a dynamic transmission path. The full network topology data comes from the configuration information of the network management system, including the connection relationship of all nodes, the bandwidth capacity of each link, the type of physical transmission medium, and the protocol support situation. The node location data contains the geographic coordinate information of each node, the logical grouping label, and the network area identifier. These data are collected through a combination of automatic discovery protocol and manual configuration to form a complete network map. After data acquisition, format standardization processing is performed to convert the data into a unified JSON or XML format, which facilitates subsequent graph neural network analysis. At the same time, a data update mechanism is established to capture and update the dataset in real time when the network topology changes, ensuring the timeliness of the analysis results.

[0081] The process of analyzing SD-WAN full network topology data and node location data using graph neural networks requires the construction of a specialized graph structure model. Network nodes are mapped to graph nodes, and links are mapped to edges. Node features include geographic location, device type, and performance indicators, while edge features include bandwidth, delay, and reliability parameters. The graph neural network model uses a multi-layer graph convolution structure to aggregate neighbor node information through a message passing mechanism, learning the latent representation of nodes and edges. Initial training uses historical network state data to obtain the embedding representation of the network topology through unsupervised learning. The model outputs the initial weight of the transmission path for each node pair. These weights are calculated based on the topology structure and historical performance data, reflecting the inherent transmission capacity of the path. The weight calculation considers factors such as link bandwidth, node processing capacity, and physical distance to form a preliminary path evaluation benchmark.

[0082] Dynamic correction of the initial weight of the transmission path between nodes based on delay fluctuation characteristics and resource load trend characteristics is a continuous optimization process. Delay fluctuation characteristics come from real-time monitoring data, and the delay coefficient of variation is calculated through a sliding window. Resource load trend characteristics reflect the dynamic changes in node processing capacity, and the exponential smoothing method is used to predict short-term load conditions. The correction method uses a weighted adjustment algorithm to combine the initial weight with real-time feature data. High delay fluctuation will cause the path weight to increase, indicating a decrease in path stability. High resource load trend will correspondingly increase the path weight through the node, guiding traffic to avoid nodes with high load. The correction process is executed every 30 seconds to ensure that the weight value can timely reflect changes in network state. At the same time, a weight change threshold is set to avoid excessive frequent adjustments that can cause routing oscillation.

[0083] Building a dynamic transmission path matrix based on the corrected inter-node transmission path weight requires organizing the weight data into a two-dimensional matrix structure, with the rows and columns of the matrix corresponding to the node numbers in the network, and the matrix element values representing the comprehensive transmission cost between the corresponding node pairs. During matrix construction, sparse matrix storage technology is used to store only the actual existing link weights, reducing memory usage, while a fast indexing mechanism is established to support efficient query and update operations. Matrix data is regularly persisted to a distributed storage system, retaining historical weight information for trend analysis and troubleshooting. Each weight update generates a new matrix version, and the version control mechanism manages the change history.

[0084] The entire dynamic transmission path calculation process is integrated into the path management module of the SD-WAN controller, forming a closed-loop control system with the data collection and feature extraction module. The calculation process uses a distributed computing framework to group network nodes for processing, improving the calculation efficiency in large-scale networks. The system sets up an exception handling mechanism that uses interpolation to estimate weight values when some node data is missing, ensuring the integrity of the matrix. A weight verification process is also established to test the rationality of the weight settings through simulation transmission testing, ensuring that the path calculation results conform to the actual network conditions.

[0085] Specific implementation examples include an SD-WAN network of a multinational enterprise that contains 28 nodes and 76 interconnected links distributed globally. Topology data is collected from network devices through the NETCONF protocol, containing structured information such as device models, interface configurations, and connection relationships. Node location data is obtained from the asset management system, including latitude and longitude coordinates, machine room numbers, and network region identifiers. After format unification, these data are stored as a graph structure, with node attributes including device capacity and geographic location, and edge attributes including link bandwidth and delay baseline values. Data updates are achieved through an event-driven mechanism, automatically triggering topology updates when network configurations change, ensuring that the data remains consistent with the actual network state.

[0086] Using a graph neural network to analyze SD-WAN global topology data and node location data requires building a graph structure model suitable for network topology. The enterprise uses a GraphSAGE-based neural network architecture, representing each node as a vector containing device attributes and location features. Model training uses historical network state data to learn the underlying relationships between nodes and generate initial path weights. In specific implementation, the model input includes a feature matrix of 28 nodes and an adjacency matrix of 76 edges, processed through three layers of graph convolution layers to output the transmission cost estimate of node pairs. The initial weight calculation considers factors such as physical distance, link bandwidth, and device performance to form a baseline reference for path selection.

[0087] The initial weight of the inter-node transmission path is dynamically corrected according to the delay fluctuation characteristics and resource load trend characteristics, and a periodic adjustment mechanism is adopted. The latest delay fluctuation index and resource load score are obtained every 5 minutes, and the initial weight is adjusted through a weighting formula. A certain monitoring shows that the delay fluctuation index of the inter-node link in Europe rises from 0.2 to 0.8, and the system correspondingly increases the weight of the path by 40%. At the same time, it is found that the CPU load of a certain core node continuously exceeds 80%, and the weight of all paths passing through this node is increased by 25%. The correction process sets an upper limit for the weight change, and the single adjustment does not exceed 50% of the original weight, preventing route oscillation caused by excessive adjustment. The corrected weight value takes effect immediately and is updated to the path calculation engine.

[0088] Based on the corrected inter-node transmission path weight, a dynamic transmission path matrix is constructed by organizing the weight data into a two-dimensional data structure. The enterprise uses a 28x28 sparse matrix to store the inter-node transmission cost. The matrix row and column correspond to the node number, and the element value represents the comprehensive transmission weight calculated in real time. The matrix construction adopts a compressed storage format, which only stores valid link data to reduce memory occupation. The matrix data is updated in full every 5 minutes, and historical versions are retained for backtracking analysis. The matrix is generated and published to the message bus for consumption by the downstream routing strategy module.

[0089] The entire dynamic transmission path calculation process is implemented in a distributed computing framework. The enterprise uses SparkGraphX for graph calculation, and distributes 28 nodes to multiple computing nodes for parallel processing. The system sets an exception handling mechanism, which uses nearest neighbor interpolation method to estimate the weight value when part of the node data is missing, to ensure the integrity of the matrix. The calculation process produces detailed log records of weight adjustment reasons and value changes, which are used for subsequent algorithm optimization and network planning. The finally generated dynamic transmission path matrix is used as the core input of SD-WAN intelligent routing, driving traffic scheduling decisions.

[0090] Example 3: see Figure 4 To extract historical transmission success rate data and historical network characteristic data from the time series database, a multi-dimensional query mechanism needs to be established. The historical transmission success rate data records the transmission success rate of different types of business traffic at different time points, and the historical network characteristic data includes the delay fluctuation, packet loss burst and resource load trend characteristic values in the past period. The data extraction process adopts a combination of time range filtering and feature label matching to ensure that the data set obtained is representative and complete. The query results are sorted by time sequence to form a structured training sample set for subsequent similarity analysis.

[0091] The similarity mapping relationship between real-time network feature data and historical network feature data needs to define a representation method for feature vectors. The real-time collected network feature data is organized into a multi-dimensional vector form, with each dimension corresponding to a feature type. The historical data is also converted into a vector representation. The cosine similarity algorithm is used for similarity calculation. The similarity degree is evaluated by calculating the cosine value of the angle between the real-time vector and the historical vector. During the calculation process, the normalization processing is performed on each feature dimension to eliminate the influence of the dimension difference. After the mapping relationship is established, a similarity matrix is formed, which records the matching degree of real-time data and each historical data point. The matrix data is stored in the memory for fast query.

[0092] According to the similarity mapping relationship, the data slice with the highest historical transmission success rate needs to be matched. A similarity threshold is set to select only the historical records with a similarity exceeding the threshold for matching. The data segment with the highest transmission success rate is selected from the records that meet the conditions. The matching process uses a sliding window mechanism to find the most similar continuous time period in the historical data to the current network state. The length of the data slice is set according to the business requirements and usually contains feature data of multiple consecutive sampling points. The matching result generates a candidate slice list, which is sorted from high to low according to the transmission success rate, providing a data basis for subsequent strategy generation.

[0093] Extracting the transmission parameter configuration corresponding to the matched data slice as the transmission optimization strategy basic template needs to parse the configuration information stored in the historical data. The transmission parameters include bandwidth allocation ratio, routing priority setting, retransmission mechanism parameters, and traffic shaping strategy, etc. During the template generation process, the parameters are verified for effectiveness, and the expired or invalid configuration items are removed, and the effective parameter combination verified by practice is retained. The basic template is stored in a standardized format, including parameter name, value range, and applicable conditions, etc. metadata, which is convenient for subsequent adjustment and application.

[0094] Incremental adjustment of the transmission optimization strategy basic template combined with the characteristics of packet loss burst needs to analyze the intensity and duration of real-time packet loss burst characteristics. The packet loss burst characteristics are obtained by monitoring the packet loss rate change in a continuous time window, which is represented by burst index and severity level. The adjustment process optimizes the retransmission strategy and routing selection parameters in the template. When there is a high packet loss burst, the redundancy transmission ratio is increased, and the dependence on fragile links is reduced. The adjustment amplitude is dynamically determined according to the severity of the burst characteristics, and a gradual adjustment strategy is adopted to avoid network shock caused by parameter mutation.

[0095] The transmission optimization strategy is input into a fuzzy logic decision maker for processing. The fuzzy logic decision maker first defines the path weight of the dynamic transmission path matrix as a fuzzy input variable. The path weight is divided into multiple fuzzy sets, such as "low weight", "medium weight", "high weight", etc. Each set corresponds to a corresponding membership function. The membership function adopts a trapezoidal or triangular function form to define the degree of weight value belonging to each fuzzy set. The fuzzification process converts the precise weight value into a fuzzy language variable to provide input for subsequent reasoning.

[0096] The bandwidth allocation requirement in the transmission optimization strategy is defined as a fuzzy rule constraint. The bandwidth requirement needs to be converted into a fuzzy rule form. The bandwidth allocation requirement is determined according to the service type and service level agreement and is represented as a combination of bandwidth requirement level and priority. The fuzzy rule is written in the form of "IF-THEN", for example, "IF path weight is low AND bandwidth requirement is high THEN select probability is high". The rule base contains decision logic under various network conditions. The number of rules is configured according to the complexity of the network and usually contains tens to hundreds of rules. The rule setting considers the comprehensive influence of multiple factors to ensure the comprehensiveness and adaptability of the decision.

[0097] The barycenter method is used to solve the fuzzification process, which needs to convert the fuzzy reasoning result into a precise output value. The fuzzy reasoning process is based on the input fuzzy variable and the rule base for logical operation to obtain the fuzzy set of output variables. The barycenter method determines the final output value by calculating the centroid coordinates of the fuzzy set to realize the conversion from fuzzy to precise. The output value is the optimal path selection probability distribution, which represents the recommended degree of each path being selected. The defuzzification process ensures the continuity and smoothness of the output value, avoiding sharp changes in the probability distribution.

[0098] The barycenter method is used to solve the fuzzification process, which needs to convert the fuzzy reasoning result into a precise output value. The fuzzy reasoning process is based on the input fuzzy variable and the rule base for logical operation to obtain the fuzzy set of output variables. The barycenter method determines the final output value by calculating the centroid coordinates of the fuzzy set to realize the conversion from fuzzy to precise. The output value is the optimal path selection probability distribution, which represents the recommended degree of each path being selected. The defuzzification process ensures the continuity and smoothness of the output value, avoiding sharp changes in the probability distribution.

[0099] A specific implementation example is that a certain cloud service provider's SD-WAN network maintains historical running data for the past 18 months, containing more than 20 billion transmission records. The query process uses a distributed SQL engine to filter data by time range, node group, service type, and other dimensions. Historical transmission success rate data includes the success transmission rate and retransmission number statistics of each transmission path, and historical network feature data includes the delay fluctuation index, packet loss burst intensity, and resource load score at the corresponding time point. After data extraction, normalization processing is performed to form a standardized training sample set, each sample containing 150 feature dimensions and 30 minutes of continuous monitoring data. The similarity mapping relationship between real-time network feature data and historical network feature data uses an improved cosine similarity algorithm, the real-time data vector contains the feature data of the previous 30 minutes at the current time point, and the historical data vector comes from the preprocessed sample library. Each feature dimension is weighted during calculation, with a delay fluctuation feature weight of 0.4, a packet loss burst feature weight of 0.3, and a resource load feature weight of 0.3. The similarity threshold is set to 0.85, and only historical records exceeding this value are considered valid matches. In a certain actual operation, the system compared the current network state with 2 million records in the historical library and found the top 500 records with the highest similarity as the candidate set.

[0100] According to the similarity mapping relationship, the data slice with the highest historical transmission success rate is matched by using a hierarchical screening strategy. First, the top 10% of the records with the highest similarity are selected from the candidate set, and then the continuous time period with the highest transmission success rate is screened from the selected records. In a certain network congestion event, the system matches a data slice recorded six months ago, which records the configuration parameters with a success rate of 98% under similar network conditions. The length of the data slice is fixed at 15 minutes, and it contains complete feature data and transmission logs within that period. The matching process also considers the business type matching degree to ensure that the selected slice is consistent with the current business features. The transmission parameter configuration corresponding to the matched data slice is extracted as the basis for the transmission optimization strategy template. This template includes 32 configuration items such as bandwidth allocation ratio, QoS strategy, and retransmission mechanism parameters. The template obtained from a certain match shows that the video conference service is allocated 45% of the bandwidth, TCP acceleration is enabled, and the retransmission timeout is set to 200 ms. After the template is extracted, its effectiveness is verified by checking the compatibility of the configuration items with the current network architecture. Incompatible configuration items are automatically filtered, and the core optimization parameters are retained to form the basis template. The transmission optimization strategy basis template is incrementally adjusted based on the characteristics of the packet loss burst. When the current packet loss burst intensity exceeds the historical level, the number of retransmission attempts and the redundancy coding ratio are increased accordingly. Real-time monitoring shows that the packet loss burst intensity reaches 0.8 (historical average 0.3), and the system increases the number of retransmissions from 3 to 5 and the forward error correction redundancy from 10% to 25%. The adjustment amplitude is calculated based on the deviation between the real-time features and the historical features, and the specific parameter values are determined using linear interpolation. The adjusted parameters take effect immediately but are set to an observation period for subsequent optimization based on actual effects.

[0101] After the transmission optimization strategy is generated, it is input into the fuzzy logic decision maker for processing. First, the path weights of the dynamic transmission path matrix are converted into fuzzy variables. Path weights 0-0.3 are defined as "low", 0.3-0.7 as "medium", and 0.7-1.0 as "high". Each interval corresponds to a trapezoidal membership function. Real-time calculation shows that the weight of a certain core path is 0.85, with a membership degree of 0.8 in the "high" weight category and a membership degree of 0.2 in the "medium" weight category. The bandwidth allocation requirement in the transmission optimization strategy is defined as a fuzzy rule constraint, which requires the establishment of a rule base containing 128 IF-THEN rules. For example: "IF path weight is high AND business type is real-time video THEN selection probability is low", "IF packet loss burst intensity is high AND path weight is low THEN selection probability is medium". The rule weights are dynamically adjusted based on historical effects, with higher weights for rules with better effects. Rule matching uses the max-min composition method to calculate the output fuzzy set of all applicable rules.

[0102] The output fuzzy set is converted into accurate probability values by using the barycenter method to solve the fuzzification process, and the numerical value corresponding to the barycenter of the fuzzy set is calculated as the final output. The selection probability of path A is 0.75, path B is 0.15, and path C is 0.10 obtained by a certain calculation. The de-fuzzification process ensures that the sum of all path probabilities is 1, and the probability distribution is smooth and continuous, avoiding sharp jumps. According to the optimal path selection probability distribution, a dynamic routing strategy table is generated, which needs to map the probability value to the actual routing configuration. The strategy table contains path identifier, selection probability, effective time and service type filtering condition. The generated strategy table is pushed to the SD-WAN controller through API, and the controller allocates traffic according to the probability distribution using a random weighted selection algorithm. The strategy table is updated every 5 minutes, and historical versions are retained for auditing and rollback.

[0103] Example 4: see Figure 5 , Building a long short-term memory neural network model requires sufficient historical data as a training basis. The training data comes from network state records stored in the time series database, including resource load trend features, dynamic routing strategy tables, and historical congestion records, etc. For example, a certain SD-WAN network with three nodes has a training data collection time span of three months, with resource utilization rate data collected every second, routing strategy changes recorded every minute, and congestion event occurrence marked every hour. In the data preprocessing stage, the data of different time granularities need to be aligned and normalized to form a unified input vector format. The model structure is designed as a three-layer LSTM network layer plus a fully connected output layer, with the number of hidden units set according to the network size, which is set to 128 units in this example. The model training uses time series cross-validation method, which divides the data into training set and validation set according to time sequence, optimizes network parameters through back propagation algorithm, and uses early stopping mechanism to prevent overfitting during training.

[0104] When the resource load trend features and dynamic routing strategy table of the current period are input into the trained model, the real-time data needs to be converted into the input format acceptable by the model. Assuming that the current time point is T time, the system will obtain the resource load trend data of the previous 60 minutes at T time, including the moving average of CPU utilization, memory usage and bandwidth occupancy of each node, and extract the currently effective dynamic routing strategy table and convert it into a numerical matrix. After standardization, these data form an input vector, which is input into the trained LSTM model. The model internally processes time series features layer by layer through the gating mechanism, the forget gate determines which historical information needs to be retained, the input gate controls the addition of new information, and the output gate generates the hidden state at the current time. After multiple layers of processing, the model finally outputs the congestion risk prediction of each node in the future preset period.

[0105] The congestion risk coefficient of each SD-WAN node in the future preset period is output as a continuous value between 0 and 1, and the higher the value, the greater the congestion risk. Table 1 below shows an example of the model output at a certain time point:

[0106] Table 1: Node Congestion Risk Coefficient Table

[0107]

[0108] When the congestion risk coefficient exceeds the preset threshold, the system starts the flow scheduling parameter adjustment process. The preset threshold is set according to the actual network carrying capacity. In this example, the threshold is set to 0.7. When the risk coefficient of node A at T+25 minutes reaches 0.73, exceeding the threshold of 0.7, the system immediately triggers the adjustment mechanism. The adjustment process first analyzes the current flow distribution, identifies the types and proportions of business flows passing through the node, and then redistributes the flow according to the business priority. High-priority real-time business traffic is preferentially migrated to low-risk paths, while low-priority background transmission tasks can be appropriately delayed or have their bandwidth allocation reduced.

[0109] According to the dynamic routing strategy table, the flow proportion of the node with high congestion risk needs to be redistributed considering the status of multiple available paths. The system queries the current available alternative paths and their performance indicators, calculates the available bandwidth and expected delay of each path, and selects nodes B and C as the main nodes to handle the overflow traffic of node A, as the model predicts that these two nodes will maintain a low risk level in the next 30 minutes. The flow redistribution adopts a gradual adjustment strategy, adjusting no more than 10% of the total flow per minute to avoid sudden large-scale flow migration causing new network problems. The actual performance indicators of each path are continuously monitored during the adjustment process to ensure that the network conditions are indeed improved after the flow migration.

[0110] The smooth transition window of the flow scheduling parameter based on the characteristics of packet loss burst needs to analyze the network packet loss in real time. The system continuously monitors the change of packet loss rate of each link, and when it detects that the packet loss rate has significantly increased in a short time, it automatically shortens the smooth transition window and speeds up the flow adjustment speed. Conversely, when the network packet loss is stable, a longer transition window is used to maintain the stability of flow scheduling. The length of the transition window is dynamically calculated according to the intensity of the packet loss burst characteristics. The higher the intensity, the shorter the window time, which can be set to a minimum of 30 seconds and a maximum of 5 minutes. This adaptive adjustment mechanism ensures that the system can quickly respond to sudden network events, while avoiding unnecessary flow shocks during stable network periods.

[0111] The entire prediction and adjustment process forms a closed-loop control system, the prediction model is re-run every 5 minutes, the risk coefficient prediction for the next 30 minutes is updated, and the flow scheduling parameters are optimized according to the latest prediction results. All adjustment operations are recorded in the operation log, including adjustment time, adjustment content, adjustment reason and execution result, etc. These log data are used for subsequent model optimization and system improvement. The system also sets an abnormal handling mechanism, when the prediction result deviates greatly from the actual situation, the model retraining process is automatically triggered, the model parameters are updated using the latest data, and the prediction accuracy is maintained. Through this continuous learning and optimization mechanism, the system can gradually adapt to changes in the network environment, improve the accuracy and reliability of congestion prediction and flow scheduling.

[0112] In Example 5, converting the dynamic transmission path matrix into a topology heat map requires mapping the matrix data into visual elements, where each weight value in the matrix corresponds to the transmission cost of a path in the network. These numerical values are converted through a color coding scheme: lower weight values are mapped to cool tones (such as blue), indicating paths with higher transmission efficiency; higher weight values are mapped to warm tones (such as red), indicating paths with transmission bottlenecks. During the conversion process, a linear interpolation algorithm is used to determine the specific color values, ensuring that the color gradient is directly proportional to the change in weight values. The generated heat map is superimposed on the network topology map, with colored lines connecting the nodes. The line color reflects the real-time path status, and the line thickness represents the bandwidth capacity. Users can adjust the display range and time scale of the heat map through the interactive interface to observe the state changes of network paths at different time periods.

[0113] Mapping the congestion risk coefficient to the node color gradient requires establishing a correspondence between risk levels and colors. Risk coefficients from 0 to 0.3 are defined as low risk, corresponding to green; 0.3 to 0.7 are medium risk, corresponding to yellow; and above 0.7 are high risk, corresponding to red. The display icon of each node is filled with the corresponding color according to the real-time calculated risk coefficient, forming an intuitive risk indication. The color gradient processing uses HSL color space conversion to achieve a smooth color transition effect. The node size also reflects the importance of the device, with core nodes displaying larger icons and edge nodes displaying smaller icons. When the node risk state changes, the color gradient transition lasts for 0.5 seconds to avoid sudden changes affecting observation.

[0114] The real-time adjustment trajectory of the flow scheduling parameter displayed in the superimposed manner needs to capture the flow redirection path. The system records the source node, target node and timestamp information of each flow adjustment, and displays the flow migration path in the form of dynamic arrows on the topology graph. The arrow trajectory is drawn using a Bezier curve, the curve color is consistent with the target path state, and the arrow size is proportional to the migrated flow. The adjustment trajectory is presented in an animated manner, showing the flow adjustment history in the last 5 minutes. The user can pause the animation to view the adjustment at a specific time in detail. The trajectory data is updated every 10 seconds to ensure real-time display.

[0115] The visualization is implemented through a Web monitoring platform, which uses hierarchical rendering technology to improve display efficiency. The bottom layer displays the network topology framework, the middle layer renders the heat map and node color, and the top layer displays the dynamic adjustment trajectory. The platform supports multiple view modes, allowing users to view the global situation or focus on a specific area. All visualization elements are drawn using vector graphics, supporting lossless scaling and resolution adaptation. The user interface provides legend and filtering controls, allowing users to filter the display content by business type or risk level.

[0116] Storing network state data, dynamic routing strategy and flow scheduling parameter by timestamp into time series database requires a reasonable data structure. Each network state data record contains measurement value, timestamp, data source node identifier and probe device ID. Data writing uses batch submission, with data in the cache being written to the database in batches every 30 seconds, reducing the number of I / O operations. The database table structure uses a tagging design, making it easy to retrieve data by multiple dimensions. The storage process includes data integrity checking, and invalid data is recorded in the error log but does not affect the normal storage process.

[0117] To attach a collection terminal device identifier to each network state data requires establishing a device registration mechanism. Each probe device is assigned a unique identifier when deployed, which includes device model, deployment location and network area information. The identifier is automatically added to the data packet header during data collection and transmitted and stored with the measurement data. The device identifier is generated using a hash algorithm to produce a fixed-length code, ensuring uniqueness and saving storage space. The system maintains a device status table that records the online status and last reporting time of each probe device, ensuring traceability of data sources.

[0118] To attach a policy version identifier to the dynamic routing strategy requires establishing a version management mechanism. An incremental version number is assigned each time a policy is generated or modified, and the version identifier includes the generation time, policy type and change description. The policy version information is stored together with the policy content to form a complete policy change history. The version rollback function allows the system to revert to a historical policy version when needed, and the version comparison tool can compare the differences between different versions of the policy. The policy identifier uses a combination of timestamps and random codes to avoid version conflicts.

[0119] The bi-directional indexing of the collection terminal device identifier and the policy version identifier in chronological order requires the design of a composite index structure. The time series database creates two main indexes: a device data index sorted by time range and a policy version index sorted by time range. The index uses a B+ tree structure to achieve efficient range queries, supporting fast retrieval of device data or policy records by time interval. The bi-directional index allows traversal of data from front to back or back to front, meeting the query requirements in different directions. Index data is periodically optimized and reorganized to maintain stable query performance. All index operations are executed asynchronously in the background, without affecting the foreground data writing and query response speed.

[0120] The entire visualization and storage system adopts a microservice architecture, with visualization display, data storage, and index maintenance implemented by independent service modules, and data exchange through a message queue. Lightweight communication protocols are used between services to ensure system scalability and maintainability. The monitoring platform provides RESTful API interfaces to support third-party systems to access visualization data or historical records. The system state is monitored in real time, and any component exception will trigger an alarm to notify the operation and maintenance personnel.

[0121] It should be noted that, in this text, relational terms such as first and second are used only to distinguish one entity or action from another, and do not necessarily require or imply that there is any such actual relationship or order between these entities or actions. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device.

[0122] Although embodiments of the present application have been shown and described, it will be understood by those having ordinary skill in the art that various changes, modifications, substitutions and alterations can be made thereto without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A software-defined wide area network (WAN) data transmission method based on edge computing, characterized in that, include: Network status data is collected by distributed probes deployed at SD-WAN edge nodes, and the network status data is cleaned and features are extracted. Calculate the dynamic transmission path between SD-WAN nodes based on the extracted network feature data; The collaborative filtering algorithm is used to analyze historical transmission data and real-time network feature data to generate transmission optimization strategies. Based on the aforementioned transmission optimization strategy and network characteristic data, a dynamic routing strategy is generated through a fuzzy logic decision-maker. Based on the dynamic routing strategy and network characteristic data, predict the future network congestion status; Based on the future network congestion status and dynamic routing strategy, adjust the traffic scheduling parameters of the SD-WAN nodes; The network status data, dynamic routing strategies, and traffic scheduling parameters are visualized. The network status data, dynamic routing strategies, and traffic scheduling parameters are classified and stored in a time-series database according to timestamps. The dynamic transmission path between SD-WAN nodes includes: Obtain SD-WAN network topology data and node location data; A graph neural network is used to analyze the SD-WAN network topology data and node location data to generate initial weights for the transmission paths between nodes; The initial weights of the inter-node transmission paths are dynamically adjusted based on latency fluctuation characteristics and resource load trend characteristics. A dynamic transmission path matrix is ​​constructed based on the corrected inter-node transmission path weights. The method of using collaborative filtering algorithms to analyze historical transmission data and real-time network feature data to generate transmission optimization strategies includes: Extract historical transmission success rate data and historical network characteristic data from the time-series database; Establish a similarity mapping relationship between real-time network feature data and historical network feature data; Based on the similarity mapping relationship, match the data slice with the highest historical transmission success rate; Extract the transmission parameter configuration corresponding to the matching data slice as the basic template for the transmission optimization strategy; Incremental adjustments were made to the basic template of the transmission optimization strategy based on the characteristics of packet loss bursts. The generation of dynamic routing strategies via a fuzzy logic decision-maker includes: The path weights of the dynamic transmission path matrix are defined as fuzzy input variables; The bandwidth allocation requirements in the transmission optimization strategy are defined as fuzzy rule constraints; The centroid method is used to defuzzify the data and output the probability distribution of the optimal path selection. A dynamic routing strategy table is generated based on the optimal path selection probability distribution. The prediction of future network congestion status includes: Construct a long short-term memory neural network model with input dimensions including resource load trend features, dynamic routing strategy table and historical congestion records; Input the current period's resource load trend characteristics and dynamic routing strategy table into the trained long short-term memory neural network model; Output the congestion risk coefficient of each SD-WAN node within a preset future time period.

2. The software-defined wide area network data transmission method based on edge computing as described in claim 1, characterized in that, The process of collecting network status data through distributed probes deployed at SD-WAN edge nodes, and cleaning and extracting features from the network status data, includes: Deploy probe clusters at SD-WAN edge nodes and use a time synchronization protocol to align the clocks of all probes; The probe cluster continuously collects indicators such as link latency, packet loss rate, and node resource utilization. Outliers in the link delay, packet loss rate, and node resource utilization metrics are removed using a median filtering algorithm. The system extracts latency fluctuation features from cleaned link latency data, packet loss burst features from packet loss rate data, and resource load trend features from node resource utilization data.

3. The software-defined wide area network data transmission method based on edge computing as described in claim 1, characterized in that, The adjustment of traffic scheduling parameters for SD-WAN nodes includes: When the congestion risk coefficient exceeds a preset threshold, traffic scheduling parameter adjustment is initiated. The traffic proportion of nodes with high congestion risk is reallocated according to the dynamic routing policy table; A smooth transition window for dynamically setting traffic scheduling parameters based on the burst characteristics of packet loss.

4. The software-defined wide area network data transmission method based on edge computing as described in claim 3, characterized in that, The visualization of the network status data, dynamic routing strategies, and traffic scheduling parameters includes: Convert the dynamic transmission path matrix into a topology heatmap; Map the congestion risk coefficient to a node color gradient; The real-time adjustment trajectory of the traffic scheduling parameters is overlaid and displayed.

5. The software-defined wide area network data transmission method based on edge computing as described in claim 4, characterized in that, The step of storing the network status data, dynamic routing policies, and traffic scheduling parameters into a time-series database according to timestamps includes: Attach a data acquisition terminal device identifier to each piece of network status data; Attach a policy version identifier to the dynamic routing policy; A bidirectional index is established for the data collection terminal device identifier and policy version identifier in chronological order based on timestamps.

6. A software-defined wide area network (WAN) data transmission system based on edge computing, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the software-defined wide area network data transmission method based on edge computing as described in any one of claims 1 to 5.

Images