A method for dynamic orchestration of network paths through multimodal network index fusion
By fusing multimodal network metrics and using a lightweight XGBoost model, the system collects and evaluates the status of network, host, and cloud resources in real time, enabling efficient and real-time path selection in vehicle-road-cloud scenarios. This solves the problems of inaccurate decision-making and time-consuming inference on edge devices, and improves the stability and reliability of service transmission.
Patent Information
- Application Number
- CN202610969591.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-01
- Publication Date
- 2026-11-10
AI Technical Summary
Existing path selection technologies, in the context of vehicle-road-cloud integration, ignore non-network factors such as the computing power bottleneck of edge nodes and the competition for cloud resources, leading to inaccurate decision-making. Furthermore, existing AI solutions are time-consuming to infer on edge devices and are difficult to achieve real-time path selection.
A multimodal network metric fusion approach is adopted, which collects network quality, host load, cloud resources and SD-WAN node status metrics in real time, uses a lightweight XGBoost regression model for path scoring and dynamic switching, and introduces an online incremental learning mechanism to optimize model parameters.
It achieves millisecond-level real-time path selection on edge devices, adapts to rapid changes in network topology, improves the stability and reliability of service transmission, and solves the problem of inaccurate decision-making in complex scenarios in traditional solutions.
Smart Images

Figure CN123074967A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for dynamic orchestration of network paths, and more particularly to a method for dynamic orchestration of network paths that integrates multimodal network metrics. Background Technology
[0002] With the deep integration of vehicle-road-cloud integration and software-defined wide area network (SD-WAN) technology, transportation networks are exhibiting a new form of cloud-edge-device collaboration. Against this backdrop, services such as lidar point cloud transmission on highways and cloud-based digital twins place extremely high demands on network determinism.
[0003] Existing path selection technologies, such as traditional routing, policy-based routing, and basic SD-WAN routing, primarily rely on single-dimensional network quality metrics (QoS), such as latency, packet loss rate, and bandwidth. However, in real-world production environments, service lag is often not caused by network congestion, but rather by edge node computing bottlenecks (CPU / memory overload) or cloud resource contention. Ignoring these non-network-related but experience-impacting metrics leads to inaccurate decision-making in complex scenarios using traditional solutions.
[0004] While existing technologies are beginning to incorporate artificial intelligence (AI) for traffic prediction, most employ complex deep learning models such as LSTM and DNN. These models have a large number of parameters and long inference times, making it difficult to achieve millisecond-level real-time inference on resource-constrained edge gateways. Furthermore, existing solutions lack online incremental learning mechanisms, making them unable to adapt to the rapid evolution of network topologies in vehicle-road-cloud scenarios, resulting in model lag.
[0005] Therefore, there is an urgent need for a lightweight machine learning path orchestration method that can integrate multimodal metrics and adapt to edge computing power limitations in order to solve the above-mentioned technical bottlenecks. Summary of the Invention
[0006] Purpose of the invention: The technical problem to be solved by the present invention is to provide a method for dynamic orchestration of network paths by fusing multimodal network indicators, which addresses the shortcomings of the existing technology.
[0007] To address the aforementioned technical problems, this invention discloses a method for dynamic network path orchestration through multimodal network index fusion, comprising the following steps:
[0008] Step 1: Collect multimodal network metrics for all candidate paths in real time, and obtain standardized multimodal metric data after preprocessing;
[0009] Step 2: Based on multimodal feature index data, combined with the XGBoost regression model, perform path fusion scoring;
[0010] Step 3: Calculate the comprehensive decision value based on the path fusion score and switch paths accordingly;
[0011] Step 4: Collect multimodal index data after path switching, optimize and update the XGBoost regression model, and re-execute Step 1 to achieve dynamic orchestration of network paths for multimodal network index fusion.
[0012] Furthermore, the multimodal network metrics mentioned in step 1 include:
[0013] Network quality indicators Host load metrics Cloud resource indicators and SD-WAN node status ;in,
[0014] Network quality indicators Specifically, this includes latency, jitter, packet loss rate, and available bandwidth;
[0015] Host load metrics Specifically, this includes CPU utilization, memory usage, and disk I / O latency;
[0016] Cloud resource metrics Specifically, this includes cloud link bandwidth quota, available CPU cores, and tenant resource utilization.
[0017] SD-WAN node status Specifically, this includes routing table integrity, link redundancy status, number of sessions, and forwarding queue length;
[0018] The core metrics for defining a Service Level Agreement (SLA) include latency, jitter, packet loss rate, and available bandwidth.
[0019] Furthermore, the standardized multimodal index data obtained through preprocessing in step 1 includes:
[0020] Step 1-1: Collect multimodal network index data for all candidate paths during the data collection process.
[0021] Steps 1-2 involve preprocessing the collected multimodal network index data, specifically including:
[0022] Step 1-2-1, standardize indicators and align time, as follows:
[0023] All indicator data are collected within the same time window and stamped with a unified timestamp; heterogeneous indicator data from different sources are formatted and time-aligned.
[0024] Step 1-2-2, normalization and direction consistency processing, are as follows:
[0025] All indicator data values are mapped to the [0,1] interval; inverse normalization is used for indicators where lower values are better, and forward normalization is used for indicators where higher values are better, so as to achieve consistent directional processing.
[0026] Furthermore, the path fusion scoring described in step 2 includes:
[0027] Step 2-1: Perform inference calculations based on the multimodal index data to obtain the dynamic weights corresponding to each index;
[0028] Step 2-2: Construct a path fusion scoring model based on the dynamic weights and calculate the comprehensive score of each candidate path;
[0029] Steps 2-3: Calculate the switching cost for each candidate path;
[0030] Steps 2-4: Calculate the comprehensive decision value for each candidate path based on the comprehensive score and switching cost.
[0031] Furthermore, step 2-1, which involves performing inference calculations based on the multimodal index data, includes:
[0032] Step 2-1-1, initial weight allocation, specifically includes:
[0033] Construct input feature vectors based on multimodal index data. The initial weights are obtained by predicting using a pre-trained XGBoost regression model.
[0034] Step 2-1-2: Based on multimodal indicator data, calculate the Service Level Agreement (SLA) completion rate and SLA completion rate deviation for each candidate path in real time.
[0035] Step 2-1-3: Based on the SLA completion rate deviation, dynamically adjust the weights of each indicator according to the initial weights to obtain the dynamic weights of each indicator.
[0036] Furthermore, step 2-1-2 involves real-time calculation of the Service Level Agreement (SLA) completion rate and SLA completion rate deviation for each candidate path. The SLA completion rate is calculated based on key indicators selected according to the service type of each candidate path, specifically including:
[0037] For latency-sensitive services, the latency compliance rate is calculated based on whether the real-time latency, jitter, and packet loss rate meet the preset SLA thresholds. Shaking compliance rate and packet loss rate target rate ;
[0038] For bandwidth-sensitive services, the bandwidth compliance rate is calculated based on whether the available bandwidth meets the preset SLA threshold. ;
[0039] For services that are sensitive to stability, the resource compliance rate is calculated based on whether CPU utilization, memory usage, cloud link bandwidth quota, number of sessions, and forwarding queue length are within preset ranges. and node status compliance rate ;
[0040] The SLA completion rate of the current candidate path is calculated by weighting the achievement rates of each criterion. The details are as follows:
[0041]
[0042] in, As an indicator Weighting coefficients in SLA completion rate calculation Indicators The compliance rate;
[0043] Compare the SLA completion rate with the preset target value By comparison, the SLA deviation is obtained. , means as follows:
[0044] .
[0046] Furthermore, step 2-1-3, which describes dynamically adjusting the weights of each indicator based on the SLA completion rate deviation on the initial weights, includes:
[0047] Only when SLA deviation If the weight adjustment continues for more than a preset time window, it will be triggered; otherwise, the current weight will remain unchanged. The specific method for weight adjustment is as follows:
[0048] Determine the set of core metrics based on the current business type, and assign weighting coefficients to the SLA completion rate calculation. The key indicators are the core indicators, and the rest are non-core indicators.
[0049] Calculate the weight adjustment for each core indicator k. The details are as follows:
[0050]
[0051] in, It is a weight-adjusted learning rate. It is a core indicator The current weight;
[0052] Using weight adjustment Update the temporary weights of core metrics , means as follows:
[0053]
[0054] Calculate the total weight increment of all core metrics. It is expressed as follows:
[0055]
[0056] For non-core indicators, their weights will be proportionally reduced based on their current weight relative to the total weight of all non-core indicators to compensate for the total increase in the weight of core indicators. To ensure that the sum of all weights is 1, the dynamic weights corresponding to each indicator are obtained. Let the weight of the first indicator be... The dynamic weights corresponding to each indicator are: .
[0057] Furthermore, step 2, which involves calculating the comprehensive decision value for all candidate paths, includes:
[0058] For any candidate path Its overall score Defined as:
[0059]
[0060] in: Representing a path The first Multimodal indicator data corresponding to each indicator This indicates the total number of indicators involved in the scoring calculation.
[0061] Furthermore, step 3, which involves calculating a comprehensive decision value based on the path fusion score and then switching paths, includes:
[0062] Step 3-1, set the route switching trigger conditions, as follows:
[0063] When the current main path Overall rating Lower than alternative paths Overall rating ,Right now:
[0064]
[0065] And the duration is greater than the preset decision time threshold. When this occurs, a path switch is triggered;
[0066] Step 3-2: When a path switch is triggered, calculate the current main path. Switch to candidate path Switching cost The details are as follows:
[0067]
[0068] in, Indicates the target candidate path The normalized value of the connection delay; This represents the normalized value of active session migration cost; This represents the normalized value of the estimated packet loss rate during the handover process; This represents the normalized value of the difference in transmission costs before and after the handover; These are the weighting coefficients;
[0069] Step 3-3: Calculate the combined decision value of the current primary path and candidate paths. , means as follows:
[0070]
[0071] in, The penalty coefficient is the cost.
[0072] Steps 3-4, when the comprehensive decision value of the candidate path When the maximum value among all candidate paths is reached, a dual-mode path switching is performed.
[0073] Steps 3-5: After the path switch is executed, the preset rollback monitoring window will be activated. Continuously calculate the service level agreement deviation value of the path after the handover. , means as follows:
[0074]
[0075] in: Indicates the first Each service level agreement indicator at time The deviation;
[0076] For indicators where smaller is better, the deviation is defined as:
[0077]
[0078] For indicators where larger is better, the deviation is defined as:
[0079]
[0080] in, This is the real-time measurement value of the k-th indicator. This refers to the service level agreement threshold corresponding to this indicator;
[0081] When any preset condition is met, the execution path is rolled back; wherein the preset conditions include:
[0082] Preset condition one, during continuous monitoring duration Within, Service Level Agreement Deviation Value Continuously greater than the preset deviation threshold ;
[0083] Preset condition two: Overall score of the path after switching Continuous monitoring duration The internal score is lower than the original path score before the switch. Subtract the preset tolerance threshold ;
[0084] Preset condition three: A preset abnormal event has been detected;
[0085] Before executing the path rollback, determine whether the original path is available. If the original path is available, roll back the business to the original path; otherwise, select the path with the largest comprehensive decision value from the remaining candidate paths as the alternative rollback path.
[0086] Cooldown period starts after rollback is complete. During the cooling-off period, switching back to the path where the previous rollback failed is prohibited.
[0087] Furthermore, the optimization and updating of the XGBoost regression model described in step 4 includes:
[0088] Step 4-1: Collect multimodal index data and construct input feature vectors. The input feature vector includes: service features, network features, resource features, and handover features.
[0089] Step 4-2, utilize the pre-trained XGBoost regression model The predicted initial weights for the next period are expressed as follows:
[0090]
[0091] in, The predicted weight vector output by the model; For the XGBoost regression model, by Composed of 10 regression trees; For the first The predicted output of the regression tree;
[0092] Step 4-3: In the offline phase, using historical path switching records as the training set and the SLA compliance rate after path switching as the label, train the XGBoost regression model to learn the input feature vector. With optimal weight Nonlinear mapping relationship between them;
[0093] Step 4-4: At preset intervals or when the amount of newly collected data reaches a preset threshold, incrementally train the model based on the latest historical data; calculate the gradient of the loss function. and the model parameters Fine-tuning is performed as follows:
[0094]
[0095] in, For learning rate, These are the current model parameters. These are the fine-tuned model parameters.
[0096] Beneficial effects:
[0097] 1. This invention achieves fine-grained perception of integrated computing and network capabilities, avoiding the blind spots of decision-making that rely solely on network metrics. This invention not only collects traditional network quality indicators such as latency and packet loss, but also incorporates the host load of edge nodes, such as CPU / memory and SD-WAN tunnel status, into the decision-making system. In vehicle-road-cloud scenarios, this can effectively identify service delays caused by network overload despite smooth operation, solving the technical pain point of inaccurate decision-making in traditional path selection algorithms in "vehicle-road-cloud collaborative" scenarios.
[0098] 2. This invention employs a lightweight XGBoost model, ensuring AI inference accuracy while meeting the real-time requirements of edge computing. Addressing the high inference latency issue of deep learning models (such as LSTM) on edge gateways, this invention uses a lightweight XGBoost regression model. This model has a small number of parameters and low computational complexity, enabling weight calculation and path scoring to be completed in milliseconds. This solves the problem of AI algorithms being difficult to implement on resource-constrained edge devices, achieving truly real-time intelligent routing.
[0099] 3. This invention introduces an online incremental learning mechanism, endowing the system with a "self-healing" capability to dynamically adapt to changes in business operations. This invention is not a one-time deployed model, but rather continuously fine-tunes model parameters using the latest historical data through incremental learning based on a sliding time window. This enables the system to automatically adapt to dynamic scenarios such as peak-hour traffic changes and seasonal traffic flow variations, without requiring manual intervention or retraining, thus extending the effective lifespan of the model.
[0100] 4. This invention provides highly reliable service transmission assurance, and is particularly suitable for critical services in the transportation industry. Through the aforementioned multimodal fusion decision-making, this invention can dynamically guide high-value service traffic such as LiDAR point clouds and cloud-based digital twins to the optimal path, significantly reducing transmission jitter and packet loss rates for critical services, and providing a deterministic network transmission foundation for autonomous driving and intelligent traffic management. Attached Figure Description
[0101] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.
[0102] Figure 1 This is a schematic diagram of the Unified Data Model (UDM) structure in this invention.
[0103] Figure 2 This is a schematic diagram of the overall process of the present invention.
[0104] Figure 3 This is a schematic diagram of the decision-making and path selection process in this invention. Detailed Implementation
[0105] This invention is applicable to vehicle-to-everything (V2X), transportation private networks, and SD-WAN scenarios, specifically involving a multi-dimensional decision-making method that integrates network quality, host load, cloud resources, and edge node status. This invention is particularly suitable for critical transportation industry scenarios that demand high levels of both computing power and network capabilities, such as highway LiDAR point cloud backhaul, cloud-based digital twin rendering, and emergency command and dispatch. The proposed method for dynamic network path orchestration, which integrates multi-modal network indicators, comprehensively considers multi-dimensional data such as network quality indicators, host load indicators, cloud resource indicators, and SD-WAN node status indicators during the decision-making process. It uses a fusion model to comprehensively score the path, thereby achieving intelligent selection of the optimal path and solving the problems of inconsistent service experience, low resource utilization, and policy lag caused by existing single-indicator path selection methods.
[0106] To address the shortcomings of current path selection mechanisms, such as insufficient accuracy, delayed response, and singular decision-making in the context of diversified services, complex networks, and multi-domain resource distribution, this invention proposes a method that can continuously and stably collect multi-source heterogeneous network and system operation indicators within a second or even sub-second timeframe. Through multimodal data fusion technology, these indicator data from different sources, dimensions, units, and update frequencies are uniformly modeled, normalized, and dynamically weighted to form a comprehensive scoring system that reflects the overall availability and service carrying capacity of paths in real time. Based on this scoring system, all available paths are sorted, filtered, and intelligently switched to ensure end-to-end performance stability and high overall system availability throughout the service lifecycle. The multi-source indicators focused on in this invention are not limited to the network quality parameters used in traditional path selection; they also incorporate host load indicators reflecting computing resource carrying capacity, cloud resource availability reflecting cross-domain cloud service operation capabilities, and real-time status of SD-WAN nodes reflecting the health of edge nodes and transmission links. These metrics come from multiple sources, including local physical link monitoring systems, virtualization platform monitoring interfaces, cloud service provider APIs, and internal statistical modules of SD-WAN edge gateways. They are characterized by large differences in collection frequency, inconsistent data formats, and high timeliness requirements. This means that relying solely on one type of metric will inevitably lead to distorted or even erroneous decision-making.
[0107] The core idea of this invention is as follows: Figure 2The diagram illustrates a closed-loop decision-making system for cross-domain real-time metric collection, heterogeneous data normalization, multimodal weight fusion, dynamic path scoring, and intelligent switching. This system first requires designing a unified cross-domain metric collection and transmission protocol to synchronously collect data from different sources within a unified timestamp framework and push it to the decision engine of the SD-WAN edge gateway via a standardized interface. Second, it requires establishing a multimodal metric data normalization method to map the original metric values to the same comparable intervals according to their physical meanings, ensuring directional consistency for subsequent weighted calculations. Then, it dynamically calculates the weights of each metric based on service type, current network status, historical performance data, and SLA achievement rate, implementing an adaptive weight adjustment mechanism to ensure the scoring model remains aligned with business objectives. On this basis, the decision engine scores each candidate path in real-time according to a formula and selects the primary path in descending order of scores. A stability judgment mechanism is introduced during switching to avoid frequent path changes caused by instantaneous fluctuations. Finally, the post-switching service performance is fed back to the weight adjustment module, forming a closed-loop optimization that gradually improves the adaptability of the scoring model under different services and network environments. Through the above mechanism, this invention can achieve a high-precision, high-real-time, and highly adaptive intelligent path selection in SD-WAN scenarios, solving the technical bottlenecks of existing solutions in multi-index fusion, cross-domain information synchronization, dynamic weight adjustment, and switching stability, fundamentally ensuring the stability of end-to-end service performance and the long-term availability of the entire system.
[0108] In summary, the proposed method for dynamic orchestration of network paths through multimodal network index fusion consists of three parts, as detailed below:
[0109] Part 1: Unified data modeling based on multimodal feature engineering, such as Figure 1 As shown.
[0110] Traditional SD-WAN routing mechanisms select paths based solely on network layer QoS metrics (such as latency, jitter, packet loss rate, and bandwidth), ignoring the impact of host node load status, cloud resource availability, and SD-WAN node operating status. This results in inconsistent path selection results with actual business experience, especially in cross-domain services and multi-cloud deployment scenarios.
[0111] The specific technical solution of this invention is as follows:
[0112] 1. Multi-source indicator collection
[0113] Network quality indicators Latency, jitter, packet loss rate, and available bandwidth are measured with a sampling period of 100ms to 2s. The corresponding indicator data are obtained through active detection and link monitoring protocols.
[0114] Host load metrics CPU utilization, memory usage, and disk I / O latency are collected in real time by the node agent.
[0115] Cloud resource metrics Cloud link bandwidth quota, available CPU cores, and tenant resource utilization rate are obtained through the cloud platform API.
[0116] SD-WAN node status metrics Routing table integrity, link redundancy status, number of sessions, forwarding queue length, etc., are collected through the internal statistics and monitoring module of the edge gateway.
[0117] Four types of multimodal metrics are collected simultaneously. Among them, the core metrics directly related to the Service Level Agreement (SLA) include: latency, jitter, packet loss rate, and available bandwidth, corresponding to the pre-agreed SLA thresholds for the service. Other auxiliary metrics include host load, cloud resource usage, and SD-WAN node status. The timestamps of each metric are recorded simultaneously during collection to provide a basis for subsequent time alignment and deviation calculation.
[0118] 2. Standardization of indicators and alignment with timeframes
[0119] All indicator data are collected within the same time window and stamped with a unified timestamp.
[0120] Heterogeneous data from different sources are formatted and time-aligned through a data processing layer to ensure comparability.
[0121] 3. Normalization and directional consistency
[0122] All index values are mapped to the [0,1] interval, with high scores indicating excellent performance.
[0123] Inverse normalization is used for metrics that are better when the value is lower (latency, packet loss rate, jitter), while forward normalization is used for metrics that are better when the value is higher (bandwidth, CPU utilization).
[0124] This mechanism ensures that path scores not only reflect network link quality but also the system's end-to-end service carrying capacity.
[0125] Part Two: Path Fusion Scoring Based on Multimodal Feature Index Data.
[0126] The weights of existing routing mechanisms are mostly statically configured and cannot be dynamically adjusted according to service type, operating environment and network status. As a result, when multiple services are running concurrently or the network status changes frequently, the path selection results cannot well match the service SLA requirements.
[0127] The specific technical solution is as follows:
[0128] 1. Initial weight allocation
[0129] The system obtains the initial weights by predicting them using a pre-trained lightweight XGBoost regression model.
[0130] The system collects multi-source data in real time and constructs input feature vectors. The vector includes:
[0131] Business characteristics: Current business type (One-hot encoding), business traffic burst coefficient;
[0132] Network characteristics: Near Each sampling period (e.g.) The average delay of ) jitter variance Packet loss rate trend ;
[0133] Resource characteristics: Edge node CPU utilization rate change rate, memory usage fluctuation rate;
[0134] Switching characteristics: The cost of the last switch and the benefits of the last switch (the extent of SLA improvement).
[0135] The above feature vectors The input is fed into a pre-trained lightweight XGBoost regression model to obtain the initial weights.
[0136] 2. Adaptive weight adjustment
[0137] Based on the preprocessed indicator data, the Service Level Agreement (SLA) completion rate for each candidate path is calculated in real time. The SLA completion rate is calculated by selecting corresponding key indicators according to the service type. The key indicators include at least network quality indicators such as latency, jitter, packet loss rate, and available bandwidth, and can be further combined with host load indicators, cloud resource indicators, and SD-WAN node status indicators for comprehensive judgment.
[0138] For latency-sensitive services, the latency compliance rate is calculated based on whether the real-time latency, jitter, and packet loss rate meet the preset SLA thresholds. Shaking compliance rate and packet loss rate target rate For bandwidth-sensitive services, the bandwidth compliance rate is calculated based on whether the available bandwidth meets the preset SLA threshold. For services with high requirements for service stability, the corresponding resource compliance rate is calculated by considering whether CPU utilization, memory usage, cloud link bandwidth quota, number of sessions, and forwarding queue length are within preset ranges. and node status compliance rate .
[0139] The SLA completion rate for the current business is obtained by weighting the achievement rates of various targets. ,in This represents the weighting coefficient of indicator k in the SLA assessment. In the three different business types mentioned above, the respective compliance rates are represented, satisfying... The SLA completion rate is compared with the preset target value. By comparison, the SLA deviation is obtained. .
[0140] Among them, the resource compliance rate The weighted average of the compliance rates of the host and cloud resource sub-indicators is calculated using the following formula:
[0141]
[0142] In the formula: , , , The weighting coefficients for each sub-indicator satisfy the following conditions: ; , , These are the compliance rates for CPU utilization, memory usage, and disk I / O latency, respectively. The overall compliance rate of cloud resources is obtained by weighting the compliance rates of cloud link bandwidth quota, available CPU cores, and tenant resource utilization rate.
[0143] Node status compliance rate The weighted average of the compliance rates of each core sub-indicator of the SD-WAN node is calculated using the following formula:
[0144]
[0145] In the formula: , , , The weighting coefficients for each sub-indicator satisfy the following conditions: Preset according to the business requirements for node stability; Routing table integrity compliance rate: The percentage of sampling points within the statistical time window where the routing table entry integrity rate is ≥ the preset threshold; Link redundancy compliance rate: The percentage of sampling points in which both primary and backup links of a node are available within the statistical time window; Session count compliance rate: The percentage of sampling points where the number of active sessions on a node is ≤ the preset maximum capacity threshold within the statistical time window; Forwarding queue length compliance rate: The percentage of sampling points where the node forwarding queue length is ≤ the preset congestion threshold within the statistical time window.
[0146] The weights of relevant indicators are dynamically adjusted based on the SLA completion rate deviation: only when the SLA completion rate deviation... If the deviation continues for more than the preset time window, a weight adjustment is triggered; if the deviation is negative or the duration is insufficient, the current weight remains unchanged.
[0147] when When the SLA is not met, first determine the set of core indicators based on the current business type: adjust the weighting coefficients in the SLA completion rate calculation. The indicator k is determined to be a core indicator, and the remaining indicators are non-core indicators. A weight adjustment is calculated for each core indicator k. :
[0148]
[0149] in, It is the weighted learning rate (with a value range of 0.01~0.1, preset according to the real-time requirements of the business). It is a core indicator The current weights are updated, and the temporary weights of the core metrics are updated. :
[0150]
[0151] Calculate the total weight increment for all core metrics:
[0152]
[0153] For non-core indicators, their weights are proportionally reduced based on their current weight relative to the total weight of all non-core indicators to compensate for the total increase in the weight of core indicators, ensuring that the sum of all weights always equals 1. This process is ultimately applied to the [specific indicator / indicator]. The weight coefficients corresponding to each indicator are denoted as follows: .
[0154] 3. Fusion Scoring Formula
[0155] After completing the index normalization process and weight allocation, the edge gateway constructs a path comprehensive scoring model based on multimodal indicators and calculates the score for each candidate path.
[0156] Specifically, for any candidate path Its overall score Defined as:
[0157]
[0158] in: Representing a path In the The normalized values for each metric range from [0,1], with larger values indicating better performance. Indicates the first The weight coefficients corresponding to each indicator satisfy the following conditions: ; This indicates the total number of indicators involved in the scoring calculation.
[0159] Combining the four types of multimodal indicator systems defined above in this invention, all indicators are categorized as network quality indicators. Host load metrics Cloud resource indicators and SD-WAN node status metrics By grouping the data, the general comprehensive scoring formula described above can be further expanded into a weighted sum of four sub-scorings:
[0160]
[0161] in: Representing a path The network quality sub-score is obtained by weighting the normalized results of latency, jitter, packet loss rate, and available bandwidth. The sub-score of the host load corresponding to the path is obtained by weighting the normalized results of CPU utilization, memory usage and disk I / O latency. The sub-score of the cloud resources corresponding to the path is obtained by weighting the normalized results of cloud link bandwidth quota, available CPU cores and resource utilization.
[0162] The sub-score representing the status of the SD-WAN node corresponding to the path is obtained by weighting the normalized results of routing table integrity, link redundancy status, number of sessions, and forwarding queue length. These represent the category weights corresponding to the four categories of indicators, and satisfy the following conditions:
[0163]
[0164] The weights of the major categories and the weights of various sub-indicators are dynamically adjusted based on the business type, service level agreement completion rate, and real-time network status, so that the comprehensive score can reflect the importance of various indicators in the current business scenario.
[0165] All candidate paths are ranked based on the comprehensive scores mentioned above, providing a basis for subsequent path selection and switching decisions.
[0166] Part Three: Dynamic Network Path Switching and Stability Control Mechanisms, such as Figure 3 As shown.
[0167] If the triggering conditions for dynamic path switching are too sensitive, it will lead to frequent switching (path jitter); if the triggering conditions are too conservative, it will delay the switching and reduce the effectiveness of business performance assurance. The existing mechanism lacks an effective balance between the smoothness and stability of the switching.
[0168] The specific technical solution is as follows:
[0169] 1. Switch trigger determination
[0170] When the current main path Overall rating continuously lower than alternative paths Overall rating And the difference between the two scores satisfies:
[0171]
[0172] And the duration is greater than the preset decision time threshold. The path switch is only triggered at that time.
[0173] in, and All scores are comprehensive path scores calculated using the aforementioned fusion scoring formula. The threshold for the score difference. To continuously determine the time window.
[0174] Preferably, It can be adaptively adjusted according to the type of service; for latency-sensitive services, a smaller value is used. and shorter For bandwidth-sensitive services, a larger [size / weight] is used. and longer .
[0175] To avoid frequent or high-cost switching due to relying solely on path score for switching, this invention quantifies and evaluates the switching cost of candidate paths. This switching cost characterizes the combined impact on service continuity, transmission efficiency, and resource overhead during the switching process from the current primary path to the target candidate path.
[0176] Specifically, for the current main path Switch to candidate path Switching cost is defined as:
[0177]
[0178] in:
[0179] Indicates the target candidate path The normalized value of the connection establishment delay is used to characterize the time overhead required to establish a new link, initialize the tunnel, achieve route convergence, or make the forwarding table effective.
[0180] This represents the normalized value of active session migration cost, used to characterize the number of business sessions that need to be migrated during path switching, the session type weight, and the migration processing overhead.
[0181] This represents the normalized value of the estimated packet loss rate during the handover process, used to characterize the risk of data packet loss that may occur at the moment of handover.
[0182] This represents the normalized value of the difference in transmission costs before and after the switch, used to characterize the cost increment caused by changes in leased line fees, cross-cloud network fees, bandwidth usage fees, or traffic unit prices.
[0183] Let be the weight coefficients of each component, and satisfy .
[0184] Each weighting coefficient is pre-configured or dynamically adjusted according to the business type and service level agreement requirements.
[0185] Among them, the connection delay The active session migration cost can be obtained through historical link establishment time, tunnel negotiation time, route convergence time, and control plane transmission delay data of the target path; The estimated packet loss rate can be calculated based on metrics such as the number of currently active sessions, session duration, session priority, and session state synchronization overhead. The transmission cost difference can be predicted based on historical handover records, current buffer queue length, dual transmission duration, and link quality fluctuation amplitude; This is the difference between the cost per unit time or per unit traffic of the target path after the switch and the cost of the current path.
[0186] Preferably, the combined decision value for the current primary path and candidate paths is defined as:
[0187]
[0188] in, Candidate paths Overall score For the current main path Switch to candidate path The cost of switching This is the penalty coefficient. The edge gateway selects a comprehensive decision value from candidate paths that satisfy the service level agreement constraints. The longest path is used as the target switching path.
[0189] Preferably, when the comprehensive score of the candidate path is higher than the score of the current main path and the score difference exceeds a preset switching threshold, and the comprehensive decision value... Path switching is triggered only when the value is the maximum among all candidate paths.
[0190] This mechanism ensures timely and stable switching, avoiding frequent jitter and repeated performance fluctuations.
[0191] 2. Smooth execution transition
[0192] During the handover process, a dual-transmission mode is used, meaning that the old and new paths are forwarded simultaneously for 100-500ms to ensure uninterrupted service and that data packets are not out of order. The duration is adjusted according to the service type: shorter durations are used for video backhaul and emergency dispatch services with high real-time requirements, and longer durations are used for ETC data services involving large file transfers.
[0193] 3. Rollback mechanism
[0194] After the path switch is completed, the edge gateway will monitor the fallback window. The system continuously monitors the Service Level Agreement (SLA) metrics of the path after the switchover to determine whether the switchover has truly improved service performance. These SLA metrics include at least one or more of latency, jitter, packet loss rate, and available bandwidth, and key metrics can be selected based on the service type.
[0195] To quantify the fallback conditions, based on the Service Level Agreement (SLA) metrics collected in step 1, such as latency, jitter, packet loss rate, and available bandwidth, the SLA deviation value for the path after the handover is defined as follows:
[0196]
[0197] in:
[0198] Indicates the first Each service level agreement metric (i.e., one of latency, jitter, packet loss rate, or available bandwidth) at time... The deviation;
[0199] Indicates the first The weights corresponding to each service level agreement indicator, and satisfying the following conditions:
[0200]
[0201] The weights are configured based on the business SLA priority.
[0202] For metrics where smaller values are better, including latency, jitter, and packet loss rate, the deviation is defined as:
[0203]
[0204] For metrics where larger is always better, including available bandwidth, the deviation is defined as:
[0205]
[0206] in, This is the real-time measurement value of the k-th indicator. This is the service level agreement threshold corresponding to this indicator.
[0207] The system will roll back the execution path if any of the following conditions are met:
[0208] Continuous monitoring duration Within, Service Level Agreement Deviation Value Continuously greater than the preset deviation threshold ;
[0209] Overall score of the route after switching Continuous monitoring duration The internal score is lower than the original path score before the switch. Subtract the preset tolerance threshold ;
[0210] Real-time abnormal events are detected, including a sudden increase in packet loss rate exceeding the preset packet loss threshold, a sudden increase in latency exceeding the preset latency threshold, path interruption, or a service level agreement compliance rate falling below the preset lower limit.
[0211] Preferably, before executing the path rollback, it should be determined whether the original path is available; if the original path is available, the business is rolled back to the original path; if the original path is unavailable, the path with the highest comprehensive decision value is selected from the remaining candidate paths as the alternative path.
[0212] Preferably, to avoid frequent back-and-forth jitter between switching and rollback, a cooling-off period is initiated after the rollback is completed. During the cooling period, switching back to the path that just failed to revert is prohibited unless the path meets the preset stability condition again.
[0213] Part 4: Online Incremental Learning Loop Based on Sliding Time Window
[0214] To enable the comprehensive scoring model to continuously adapt to changes in business type, network status fluctuations, and resource load variations, this invention employs a dynamic weight adjustment mechanism based on a lightweight machine learning model. This mechanism utilizes historical path switching records and real-time performance feedback to construct feature vectors, and then uses a pre-trained model to predict the optimal weights for the next cycle, replacing traditional static configuration or simple linear adjustment algorithms.
[0215] 1. Feature Engineering and Dataset Construction
[0216] The system collects multi-source data in real time and constructs input feature vectors. The vector includes:
[0217] Business characteristics: Current business type (One-hot encoding), business traffic burst coefficient;
[0218] Network characteristics: Near Each sampling period (e.g.) The average delay of ) jitter variance Packet loss rate trend ;
[0219] Resource characteristics: Edge node CPU utilization rate change rate, memory usage fluctuation rate;
[0220] Switching characteristics: The cost of the last switch and the benefits of the last switch (the extent of SLA improvement).
[0221] 2. Weighted Prediction Model
[0222] This invention utilizes a pre-trained lightweight XGBoost (Extreme Gradient Boosting) regression model. The basic dynamic weights for predicting the next cycle are the initial weights mentioned earlier. The model's prediction logic is expressed as follows:
[0223]
[0224] in: The predicted weight vector output by the model; For XGBoost ensemble learning models, by Composed of 10 regression trees; For the first The predicted output of the regression tree; The input feature vector.
[0225] 3. Model Training and Update
[0226] During the offline phase, historical path switching records were used as the training set, and the XGBoost model was trained to learn features using "SLA compliance rate after switching" as the label. With optimal weight The nonlinear mapping relationship between them. During online operation, the edge gateway only needs to perform lightweight model inference, inputting the current feature vector. This allows us to obtain the optimal weights for the current scenario. .
[0227] 4. Incremental update mechanism
[0228] To ensure the model's timeliness, the system performs incremental training on the model every 24 hours (or when the amount of newly collected data reaches a preset threshold) based on the latest historical data. The system calculates the gradient of the loss function. And according to the formula:
[0229]
[0230] For model parameters Fine-tuning was carried out. Among other things, The learning rate is extremely small. This mechanism allows the model to adapt to long-term evolution of network topology without having to be trained from scratch, thus achieving a balance between high accuracy and low computational cost in edge computing scenarios.
[0231] Example:
[0232] This embodiment takes a "vehicle-road-cloud integrated" collaborative scenario on a highway in an intelligent transportation system as an example. In this scenario, the roadside edge computing node (MEC) needs to transmit the real-time perceived traffic flow data (including LiDAR point cloud, high-definition camera video stream, and V2X signaling) to the cloud-based digital twin platform for real-time rendering and analysis.
[0233] The hardware architecture of this embodiment uses an SD-WAN edge gateway with NPU (Neural Processing Unit) computing power as the decision-making entity to support real-time inference of lightweight machine learning models. Assume there are three candidate paths in the current network environment:
[0234] Path A (5G-A uRLLC slice): Ultra-reliable low-latency communication slice based on 5G-Advanced.
[0235] Path B (SRv6 backbone leased line): Government / transportation backbone network based on SRv6 (segmented routing) technology.
[0236] Path C (Ordinary Broadband Internet): Ordinary public network access link.
[0237] The edge gateway executes the path selection method proposed in this invention with a sampling period of 200ms.
[0238] Step 1: Cross-domain multimodal data acquisition and preprocessing.
[0239] The edge gateway collects metric data for each candidate path in real time. The raw data collected for point cloud synchronization services is shown in the table below:
[0240] Indicator Type Specific indicators Path A (5G-A) Path B (SRv6) Path C (Broadband) Network quality Transmission delay 8ms 20ms 50ms Network quality Packet loss rate 0.01% 0.05% 5% Host load Edge node CPU 95% 40% 30% SD-WAN status Tunnel queue depth 150 20 10
[0241] Normalization processing: Based on the method described in the article, inverse normalization is used for indicators such as latency, packet loss rate, and CPU utilization, which are better when they are smaller.
[0242] Calculation example: Normalized delay value of path A The calculated result is 0.98; however, the normalized CPU utilization value for path A is... The calculation result is 0.05, and the tunnel queue depth has overflowed, with a normalized value of 0.00.
[0243] Time alignment: All metrics are assigned a uniform timestamp 't' to form a standardized metric vector. .
[0244] Step 2: Based on the weighted inference of the lightweight XGBoost model, perform fusion scoring and comprehensive decision-making to ultimately achieve path switching.
[0245] (i) In this embodiment, the lightweight XGBoost regression model built into the edge gateway is responsible for calculating the initial weights and making adaptive adjustments. This model has been trained based on massive historical switching data before leaving the factory, and only the decision tree structure is retained to adapt to the resources of the edge devices.
[0246] Feature input: The model inputs the feature vector at the current time step. (Including one-hot encoding of business type, latency jitter variance, CPU utilization trend, etc.)
[0247] Reasoning process: The model calculates the suggested weights for each indicator by traversing the pre-stored decision tree (Booster).
[0248] Scenario Analysis: The model identifies the current business as point cloud transmission, and the CPU utilization of the edge node is as high as 95%.
[0249] Output: The model outputs a weight vector. The host load weight was increased to 0.6, while the network quality weight was compressed to 0.4. This indicates that the model determines the current bottleneck lies in the computing power of edge nodes, rather than network bandwidth.
[0250] The initial weight reference values for typical business scenarios are as follows:
[0251] High-definition video real-time transmission: latency 0.4, jitter 0.3, bandwidth 0.2, node status 0.1;
[0252] Batch upload of ETC gantry transaction data: bandwidth 0.5, packet loss rate 0.2, node status 0.2, latency 0.1;
[0253] Emergency command and dispatch / vehicle-road cooperative signaling: latency 0.5, reliability 0.3, link redundancy 0.2;
[0254] General real-time video / audio: latency 0.4, jitter 0.3, bandwidth 0.2, others 0.1;
[0255] General large file transfer: bandwidth 0.5, latency 0.2, node health 0.2, and everything else 0.1.
[0256] (ii) The edge gateway calculates the comprehensive score of each path.
[0257] 1. Calculation of Integrated Scores (Formula) ):
[0258] Path A: Although the network score is extremely high (0.95), the CPU weight is boosted to 0.6 by the model, resulting in a very poor overall score. It is 0.42.
[0259] Path B: Network metrics are good, host load is moderate, and the calculated overall score is [not specified]. It is 0.85.
[0260] Path C: Packet loss rate is too high, resulting in immediate elimination.
[0261] 2. Calculation of comprehensive decision value (formula) ):
[0262] Switching Cost Path B is an SRv6 leased line with low connection establishment latency, but cross-domain handover involves session migration of large files (point cloud data), with estimated costs... .
[0263] Calculation results: Comprehensive decision value for path B .
[0264] (III) Path switching and stability control.
[0265] Since the overall decision value of path B (0.835) is much higher than that of path A (0.42), the system determines that the switching conditions are met.
[0266] 1. Smooth switching: The system initiates a dual transmission mechanism for 300ms to ensure that point cloud data packets are not out of order.
[0267] 2. Rollback Monitoring: After switching, enter the monitoring window.
[0268] (1) Monitoring results: After switching to path B, the CPU load of the edge node dropped (due to traffic offloading), and the rendering frame rate of the cloud digital twin recovered from 15fps to 60fps, and the SLA was met.
[0269] (2) Judgment: The rollback condition was not triggered, so path B is locked.
[0270] Step 3: Online incremental learning based on sliding time windows.
[0271] The closed-loop feedback mechanism in this embodiment uses a gradient descent algorithm based on a sliding time window.
[0272] 1. Data Recording: The system records the characteristics of this handover. With results .
[0273] 2. Weight Fine-tuning: The system calculates the gradient of the loss function in a silent background thread. .
[0274] (1) Logic: The switch to path B was successful under high CPU load (SLA improvement) The system determined that the XGBoost model's weight prediction in the "host load" dimension was accurate.
[0275] (2) Update: According to the formula Fine-tuning was performed on the split nodes in the model that were related to CPU utilization.
[0276] 3. Results: After this iteration, when the model encounters scenarios where edge node CPU usage is >90%, it will be able to more sensitively increase the weight of host load metrics, thereby avoiding congested paths in advance.
[0277] Therefore, this invention can organically combine multi-source indicator collection, adaptive weight adjustment, path comprehensive scoring, handover trigger determination, handover cost evaluation and rollback control in actual business scenarios, so as to realize intelligent path selection for different business types and improve the accuracy of path selection and the stability of system operation.
[0278] In its specific implementation, this application provides a computer storage medium and a corresponding data processing unit. The computer storage medium is capable of storing a computer program, which, when executed by the data processing unit, can run the invention's content regarding a multimodal network index fusion method for dynamic network path orchestration, as well as some or all of the steps in various embodiments. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.
[0279] Those skilled in the art will clearly understand that the technical solutions in the embodiments of the present invention can be implemented using computer programs and their corresponding general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of computer programs, i.e., software products. These computer program software products can be stored in a storage medium and include several instructions to cause a device containing a data processing unit (which may be a personal computer, server, microcontroller, MCU, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.
[0280] This invention provides a concept and method for dynamic network path orchestration based on multimodal network index fusion. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.
Claims
1. A method for dynamic orchestration of network paths through multimodal network index fusion, characterized in that, Includes the following steps: Step 1: Collect multimodal network metrics for all candidate paths in real time, and obtain standardized multimodal metric data after preprocessing; Step 2: Based on multimodal feature index data, combined with the XGBoost regression model, perform path fusion scoring; Step 3: Calculate the comprehensive decision value based on the path fusion score and switch paths accordingly; Step 4: Collect multimodal index data after path switching, optimize and update the XGBoost regression model, and re-execute Step 1 to achieve dynamic orchestration of network paths for multimodal network index fusion.
2. The method for dynamic orchestration of network paths through multimodal network index fusion according to claim 1, characterized in that, The multimodal network metrics mentioned in step 1 include: Network quality indicators Host load metrics Cloud resource indicators and SD-WAN node status ;in, Network quality indicators Specifically, this includes latency, jitter, packet loss rate, and available bandwidth; Host load metrics Specifically, this includes CPU utilization, memory usage, and disk I / O latency; Cloud resource metrics Specifically, this includes cloud link bandwidth quota, available CPU cores, and tenant resource utilization. SD-WAN node status Specifically, this includes routing table integrity, link redundancy status, number of sessions, and forwarding queue length; The core metrics for defining a Service Level Agreement (SLA) include latency, jitter, packet loss rate, and available bandwidth.
3. The method for dynamic orchestration of network paths through multimodal network index fusion according to claim 2, characterized in that, The standardized multimodal index data obtained through preprocessing in step 1 includes: Step 1-1: Collect multimodal network index data for all candidate paths during the data collection process. Steps 1-2 involve preprocessing the collected multimodal network index data, specifically including: Step 1-2-1, standardize indicators and align time, as follows: Collect all indicator data within the same time window and assign a unified timestamp; standardize the format and align the time for heterogeneous indicator data from different sources. Step 1-2-2, normalization and direction consistency processing, are as follows: All indicator data values are mapped to the [0,1] interval; inverse normalization is used for indicators where lower values are better, and forward normalization is used for indicators where higher values are better, so as to achieve consistent directional processing.
4. The method for dynamic orchestration of network paths through multimodal network index fusion according to claim 3, characterized in that, Step 2, which involves performing path fusion scoring, includes: Step 2-1: Perform inference calculations based on the multimodal index data to obtain the dynamic weights corresponding to each index; Step 2-2: Construct a path fusion scoring model based on the dynamic weights and calculate the comprehensive score of each candidate path; Steps 2-3: Calculate the switching cost for each candidate path; Steps 2-4: Calculate the comprehensive decision value for each candidate path based on the comprehensive score and switching cost.
5. The method for dynamic network path orchestration based on multimodal network index fusion according to claim 4, characterized in that, Step 2-1, which involves inference calculations based on the multimodal index data, includes: Step 2-1-1, initial weight allocation, specifically includes: Construct input feature vectors based on multimodal index data. The initial weights are obtained by predicting using a pre-trained XGBoost regression model. Step 2-1-2: Based on multimodal indicator data, calculate the Service Level Agreement (SLA) completion rate and SLA completion rate deviation for each candidate path in real time. Step 2-1-3: Based on the SLA completion rate deviation, dynamically adjust the weights of each indicator according to the initial weights to obtain the dynamic weights of each indicator.
6. The method for dynamic orchestration of network paths through multimodal network index fusion according to claim 5, characterized in that, Step 2-1-2 describes the real-time calculation of the Service Level Agreement (SLA) completion rate and SLA completion rate deviation for each candidate path. The SLA completion rate is calculated based on key indicators selected according to the service type of each candidate path, specifically including: For latency-sensitive services, the latency compliance rate is calculated based on whether the real-time latency, jitter, and packet loss rate meet the preset SLA thresholds. Shaking compliance rate and packet loss rate target rate ; For bandwidth-sensitive services, the bandwidth compliance rate is calculated based on whether the available bandwidth meets the preset SLA threshold. ; For services that are sensitive to stability, the resource compliance rate is calculated based on whether CPU utilization, memory usage, cloud link bandwidth quota, number of sessions, and forwarding queue length are within preset ranges. and node status compliance rate ; The SLA completion rate of the current candidate path is calculated by weighting the achievement rates of each criterion. The details are as follows: ; in, As an indicator Weighting coefficients in SLA completion rate calculation Indicators The compliance rate; Compare the SLA completion rate with the preset target value By comparison, the SLA deviation is obtained. , means as follows: 。 7. The method for dynamic orchestration of network paths through multimodal network index fusion according to claim 6, characterized in that, Step 2-1-3, which describes dynamically adjusting the weights of each indicator based on the SLA completion rate deviation and the initial weights, includes: Only when SLA deviation If the weight adjustment continues for more than a preset time window, it will be triggered; otherwise, the current weight will remain unchanged. The specific method for weight adjustment is as follows: Determine the set of core metrics based on the current business type, and assign weighting coefficients to the SLA completion rate calculation. The key indicators are the core indicators, and the rest are non-core indicators. Calculate the weight adjustment for each core indicator k. The details are as follows: ; in, It is a weight-adjusted learning rate. It is a core indicator The current weight; Using weight adjustment Update the temporary weights of core metrics , means as follows: ; Calculate the total weight increment of all core metrics. It is expressed as follows: ; For non-core indicators, their weights will be proportionally reduced based on their current weight relative to the total weight of all non-core indicators to compensate for the total increase in the weight of core indicators. To ensure that the sum of all weights is 1, the dynamic weights corresponding to each indicator are obtained. Let the weight of the first indicator be... The dynamic weights corresponding to each indicator are: .
8. The method for dynamic orchestration of network paths through multimodal network index fusion according to claim 7, characterized in that, Step 2, which involves calculating the comprehensive decision value for all candidate paths, includes: For any candidate path Its overall score Defined as: ; in: Representing a path The first Multimodal indicator data corresponding to each indicator This indicates the total number of indicators involved in the scoring calculation.
9. The method for dynamic orchestration of network paths through multimodal network index fusion according to claim 8, characterized in that, Step 3, which involves calculating a comprehensive decision value based on the path fusion score and then switching paths, includes: Step 3-1, set the route switching trigger conditions, as follows: When the current main path Overall rating Lower than alternative paths Overall rating ,Right now: ; And the duration is greater than the preset decision time threshold. When this occurs, a path switch is triggered; Step 3-2: When a path switch is triggered, calculate the current main path. Switch to candidate path Switching cost The details are as follows: ; in, Indicates the target candidate path The normalized value of the connection delay; This represents the normalized value of active session migration cost; This represents the normalized value of the estimated packet loss rate during the handover process; This represents the normalized value of the difference in transmission costs before and after the handover; These are the weighting coefficients; Step 3-3: Calculate the combined decision value of the current primary path and candidate paths. , means as follows: ; in, The penalty coefficient is the cost. Steps 3-4, when the comprehensive decision value of the candidate path When the maximum value among all candidate paths is reached, a dual-mode path switching is performed. Steps 3-5: After the path switch is executed, the preset rollback monitoring window will be activated. Continuously calculate the service level agreement deviation value of the path after the handover. , means as follows: ; in: Indicates the first Each service level agreement indicator at time The deviation; For indicators where smaller is better, the deviation is defined as: ; For indicators where larger is better, the deviation is defined as: ; in, This is the real-time measurement value of the k-th indicator. This refers to the service level agreement threshold corresponding to this indicator; When any preset condition is met, the execution path is rolled back; wherein the preset conditions include: Preset condition one, during continuous monitoring duration Within, Service Level Agreement Deviation Value Continuously greater than the preset deviation threshold ; Preset condition two: Overall score of the path after switching Continuous monitoring duration The internal score is lower than the original path score before the switch. Subtract the preset tolerance threshold ; Preset condition three: A preset abnormal event has been detected; Before executing the path rollback, determine whether the original path is available. If the original path is available, roll back the business to the original path; otherwise, select the path with the largest comprehensive decision value from the remaining candidate paths as the alternative rollback path. Cooldown period starts after rollback is complete. During the cooling-off period, switching back to the path where the previous rollback failed is prohibited.
10. The method for dynamic orchestration of network paths through multimodal network index fusion according to claim 9, characterized in that, Step 4, which involves optimizing and updating the XGBoost regression model, includes: Step 4-1: Collect multimodal index data and construct input feature vectors. The input feature vector includes: service features, network features, resource features, and handover features. Step 4-2, utilize the pre-trained XGBoost regression model The predicted initial weights for the next period are expressed as follows: ; in, The predicted weight vector output by the model; For the XGBoost regression model, by Composed of 10 regression trees; For the first The predicted output of the regression tree; Step 4-3: In the offline phase, using historical path switching records as the training set and the SLA compliance rate after path switching as the label, train the XGBoost regression model to learn the input feature vector. With optimal weight Nonlinear mapping relationship between them; Step 4-4: At preset intervals or when the amount of newly collected data reaches a preset threshold, incrementally train the model based on the latest historical data; calculate the gradient of the loss function. and the model parameters Fine-tuning is performed as follows: ; in, For learning rate, These are the current model parameters. These are the fine-tuned model parameters.
