Fault prediction-fused SRv6 network dynamic path optimization method and system
By constructing a heterogeneous graph and fault prediction model for SRv6 networks and combining it with a comprehensive optimization cost function, the shortcomings of fault prediction and path optimization in SRv6 networks are addressed, achieving intelligent and efficient network management, reducing network outages, and improving reliability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA DATANG GRP DIGITAL TECH CO LTD
- Filing Date
- 2025-12-10
- Publication Date
- 2026-05-01
AI Technical Summary
Existing SRv6 networks have shortcomings in terms of dynamism, cross-layer perception, multi-objective collaboration, and real-time accurate decision-making. They are unable to cope with the challenges brought about by sudden failures and multi-dimensional resource competition in complex network environments. In particular, they suffer from problems such as slow response, oscillation risk, and low integration of decision-making mechanisms in fault prediction and path optimization.
By acquiring real-time performance data and topology information of the SRv6 network, a heterogeneous graph of the physical topology layer and the SRv6 logical layer is constructed. The failure probability and type are output using a pre-trained link failure prediction model. The optimal primary path and backup path are calculated by combining the comprehensive optimization cost function. Real-time performance evaluation and closed-loop self-learning are supported to realize traffic switching and model iterative training.
It enables proactive fault prediction and intelligent path optimization for SRv6 networks, reducing network outages, improving network operation and maintenance efficiency and reliability, supporting rapid path switching and continuous iterative optimization, and enhancing network robustness and reliability.
Smart Images
Figure CN121967303A_ABST
Abstract
Description
A method and system for dynamic path optimization in SRv6 networks that incorporates fault prediction Technical Field
[0001] This application generally relates to the field of network communication technology. More specifically, this application relates to a dynamic path optimization method and system for SRv6 networks that incorporates fault prediction. Background Technology
[0002] In the current field of network communication technology, SRv6 (Segment Routing over IPv6), as an emerging forwarding mechanism, is gradually gaining widespread application due to its powerful programmability and potential to simplify network architecture. A large amount of research has been dedicated to improving the path selection and load balancing capabilities of SRv6 networks, aiming to improve network performance by optimizing traffic distribution.
[0003] However, existing technologies have significant shortcomings in terms of dynamism, cross-layer perception, multi-objective collaboration, and real-time accurate decision-making, especially in dealing with the challenges posed by sudden failures and multi-dimensional resource competition in complex network environments. Specific limitations are mainly reflected in the following three aspects: First, response lag and oscillation risk: Most methods rely on static or periodic weight calculation mechanisms, making it difficult to respond in real time to rapid fluctuations in network conditions, easily leading to optimization lag or causing paths to oscillate repeatedly between different links.
[0004] Secondly, there is a lack of proactive fault detection: most existing technologies do not fully consider potential network faults and lack predictive capabilities, which may lead to repeated faults after path adjustments, resulting in repeated adjustments and unnecessary resource consumption.
[0005] Third, the decision-making mechanism has low integration: the existing state assessment mechanism is often independent of the path decision-making process and is not deeply integrated into real-time routing calculation, which limits its applicability in large-scale, high-speed dynamic networks.
[0006] In view of this, there is an urgent need to provide a dynamic path optimization scheme for SRv6 networks that integrates fault prediction in order to solve the above problems and effectively cope with the performance and reliability challenges in a rapidly changing network environment. Summary of the Invention
[0007] In order to at least address one or more of the technical problems mentioned above, this application proposes a dynamic path optimization scheme for SRv6 networks that incorporates fault prediction in several aspects.
[0008] In a first aspect, this application provides a dynamic path optimization method for SRv6 networks that integrates fault prediction, comprising: acquiring real-time performance data and topology information of the SRv6 network, and performing anomaly identification and localization based on the real-time performance data; constructing a heterogeneous graph containing a physical topology layer and an SRv6 logical layer based on the real-time performance data and topology information; using the heterogeneous graph as input, outputting the predicted fault probability value and potential fault type of each link through a pre-trained link fault prediction model; defining a comprehensive optimization cost function by combining the real-time performance data of the SRv6 network and the predicted fault probability value of each network link, and calculating the optimal primary path and the optimal backup path based on the comprehensive optimization cost function; obtaining the corresponding Segment list according to the optimal primary path and the optimal backup path, and distributing the Segment list to the source node to realize traffic switching to the new path; evaluating the real-time performance of the new path, and feeding the evaluation result data back to the link fault prediction model for iterative training of the link fault prediction model.
[0009] In some embodiments, during the process of acquiring real-time performance data and topology information of the SRv6 network, and performing anomaly identification and location based on the real-time performance data, the following steps are performed: Real-time performance data of the entire SRv6 network link is collected at a preset frequency, and the topology information of the SRv6 network is also collected. The real-time performance data includes link latency, link bandwidth utilization, port packet loss rate, and bit error rate. Anomaly indicators are confirmed based on the real-time performance data of the entire SRv6 network link using a sliding time window-based judgment mechanism. The process then proceeds to the fault or congestion location stage, wherein the following steps are performed during the fault or congestion location stage: The network is abstracted as a directed graph. Where V is the vertex set, which includes network nodes and logical entities capable of performing specific SRv6 functions; E is the edge set, which includes physical links and virtual SRv6 path segments; all confirmed abnormal indicators are summarized into an abnormal indicator set. , The m-th abnormal indicator; the set of potential fault links is located using a Bayesian probabilistic model. .
[0010] In some embodiments, during the process of constructing a heterogeneous graph including a physical topology layer and an SRv6 logical layer based on the real-time performance data and topology information, the following steps are performed: constructing a physical topology layer graph based on the topology information. The real-time performance data is then assigned to the corresponding elements of the physical topology layer graph, wherein... For physical network nodes, It is a set of physical links; an SRv6 logical layer diagram is constructed based on topology information. ,in, A collection of logical functional entities for deploying SRv6 functionality. A set of virtual path segments defined by the Segment list; construct the association matrix between the physical topology layer diagram and the SRv6 logical layer diagram. The physical topology layer diagram, SRv6 logical layer diagram, and association matrix are merged to construct a two-layer heterogeneous network topology diagram.
[0011] In some embodiments, the link failure prediction model includes a spatiotemporal feature extraction module, which performs the following steps: processing the time series of multidimensional performance indicators in the heterogeneous graph using a gated recurrent unit and a one-dimensional causal convolutional layer in parallel; weighting and merging the output of the gated recurrent unit and the output of the one-dimensional causal convolutional layer using a gated fusion unit to generate a preliminary temporal feature representation for each link; wherein, in the process of processing the time series of multidimensional performance indicators in the heterogeneous graph using a gated recurrent unit, the following calculation formula is used: , Let be the hidden state of link e at time t. Let e be the hidden state of link e at time t-1. Let be the candidate hidden state of link e at time t. , This is the first weight matrix. The multidimensional performance index of link e at time t. for and The splicing matrix, It is the Sigmoid activation function. For the first bias term, , , This is the second weight matrix. For the second bias term, The hyperbolic tangent activation function is used. For Hadama accumulation, for and The splicing matrix; the calculation formula for the preliminary time-series characteristics of each link is: , This represents the preliminary time-series characteristics of link e at time t. Let be the hidden state of link e at time t. The output of a one-dimensional causal convolutional layer. These are the gating parameters.
[0012] In some embodiments, the link failure prediction model further includes a spatial dependency modeling module, which simultaneously processes the physical topology layer and the SRv6 logical layer in the heterogeneous graph. The spatial dependency modeling module performs the following steps: for any node in the heterogeneous graph, its neighboring nodes are divided into different sets of relation types, including physical neighbors, logical function neighbors, and cross-layer associated neighbors; based on any node and its neighboring nodes belonging to the relation types, the corresponding initial attention coefficient is calculated, wherein the formula for calculating the initial attention coefficient is: , It is a non-linear activation function. Let r be a vector of learnable weights for relation r. For learnable weights of relation r, Let i be the feature vector of node i. Let j be the feature vector of node j. for and The concatenated vector; the initial attention coefficients under the same relation are normalized to obtain the final attention coefficients, where the formula for calculating the final attention coefficients is: , Let i be the set of neighbors of node i under relation r. It is an exponential function; the update features of each node are obtained based on the final attention coefficient, where the calculation formula for the update features of each node is: , Let i be the updated feature of node i. This is the Sigmoid activation function.
[0013] In some embodiments, the link failure prediction model further includes a heterogeneous information fusion and prediction module, which performs the following steps: concatenating the updated features of each node with the preliminary temporal features of each link to form an enhanced feature representation for each link; inputting the enhanced feature representation into a multi-head self-attention layer to obtain the predicted failure probability value and potential failure type of each link; wherein, the multi-head self-attention layer includes two parallel prediction branches; the prediction branches include a link failure probability prediction branch and a failure type classification prediction branch, and the link failure prediction model is trained by minimizing the total loss of the output of the link failure probability prediction branch and the output of the failure type classification prediction branch; the formula for calculating the total loss is: , For the total loss, Let this be the first loss hyperparameter. For the second loss hyperparameter, This represents the probability of link failure. For real link failure labels, For the predicted failure type probability distribution, Labels for actual fault types. For binary cross-entropy loss, This represents the cross-entropy loss.
[0014] In some embodiments, the comprehensive optimization cost function is: , The comprehensive optimization cost for path P is... The available bandwidth of link e. The weighting factor is the available bandwidth. For the delay of link e, The weighting factor for time delay. Let e be the packet loss rate. This is the weighting coefficient for the packet loss rate. This is the predicted failure probability value for link e. These are the weighting coefficients for the predicted failure probability values.
[0015] In some embodiments, during the process of calculating the optimal primary path and the optimal backup path based on the comprehensive optimization cost function, the path with the lowest comprehensive cost is taken as the primary path, and the path with the second lowest comprehensive cost is taken as the backup path.
[0016] In some embodiments, during the evaluation of the real-time performance of a new path, the end-to-end performance improvement ratio, latency-sensitive traffic, path stability index, and global load balancing index are calculated based on the real-time performance of the new path.
[0017] In a second aspect, this application provides an SRv6 network dynamic path optimization system with integrated fault prediction, employing the SRv6 network dynamic path optimization method with integrated fault prediction as described in any embodiment of the first aspect. The system includes: a data monitoring module for acquiring real-time performance data and topology information of the SRv6 network, and performing anomaly identification and location based on the real-time performance data; a heterogeneous graph construction module for constructing a heterogeneous graph including a physical topology layer and an SRv6 logical layer based on the real-time performance data and topology information; and a fault prediction module for taking the heterogeneous graph as input and performing pre-trained... The link failure prediction model outputs the predicted failure probability and potential failure type of each link; the path decision module is used to define a comprehensive optimization cost function by combining the real-time performance data of the SRv6 network and the predicted failure probability of each network link, and calculate the optimal primary path and optimal backup path based on the comprehensive optimization cost function; the policy formation and distribution module obtains the corresponding segment list according to the optimal primary path and optimal backup path, and distributes the segment list to the source node to realize the traffic switching to the new path; the evaluation and feedback module evaluates the real-time performance of the new path and feeds the evaluation results back to the link failure prediction model for iterative training of the link failure prediction model.
[0018] By employing the SRv6 network dynamic path optimization scheme with integrated fault prediction as described above, this application embodiment achieves a shift from passive response to proactive fault avoidance by real-time perception of network status and prediction of link faults using pre-trained models. Future fault risks are proactively considered during path optimization, and a comprehensive cost function is defined based on real-time data to calculate more robust and reliable primary / backup paths, effectively reducing network outages. Furthermore, it supports rapid path switching and a closed-loop self-learning mechanism, enabling continuous iterative optimization of fault prediction and path strategies, ultimately achieving more intelligent, efficient, and reliable dynamic management of the SRv6 network.
[0019] Furthermore, in some embodiments, a comprehensive network status view is constructed by frequently collecting multi-dimensional real-time performance data and topology information. A sliding time window mechanism is employed to dynamically identify abnormal indicators, improving the sensitivity and accuracy of anomaly detection. More importantly, by abstracting the network as a physical-logical heterogeneous graph and combining it with a Bayesian probabilistic model, potential faulty links can be accurately located probabilistically, enabling in-depth analysis of the root causes of faults and providing robust inference even in uncertain environments. Together, these elements construct an intelligent, efficient, and accurate anomaly identification and fault location system, providing high-quality, forward-looking fault information for SRv6 network path optimization, significantly improving network operation and maintenance efficiency and reliability.
[0020] Furthermore, in some embodiments, by independently constructing physical topology layer diagrams and SRv6 logical layer diagrams, and utilizing an association matrix to capture the complex mapping relationships between them, a two-layer heterogeneous network topology diagram is ultimately formed. This structured, multi-dimensional unified representation not only provides a deep insight into the interdependencies between physical infrastructure and SRv6 services, significantly enhancing fault location and root cause analysis capabilities, but also provides high-quality input for advanced machine learning models such as graph neural networks, serving as a key foundation for realizing intelligent SRv6 network management and optimization.
[0021] Furthermore, in some embodiments, multi-scale temporal dependencies are captured by deep fusion of temporal features, and spatial dependencies of heterogeneous networks are accurately modeled to comprehensively understand fault propagation at both the physical and logical levels. A multi-head self-attention layer is used to achieve heterogeneous information fusion and multi-task prediction, while simultaneously outputting fault probabilities and types. This innovative design significantly improves the accuracy, comprehensiveness, and practicality of fault prediction, providing crucial forward-looking information for dynamic path optimization in SRv6 networks, ensuring high reliability and performance of network services. Attached Figure Description
[0022] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily understood by reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of this application are illustrated by way of example and not limitation, and the same or corresponding reference numerals denote the same or corresponding parts, wherein: Figure 1 shows an exemplary flowchart of an SRv6 network dynamic path optimization method with fusion fault prediction according to an embodiment of this application; Figure 2 shows an exemplary flowchart of the operation of a spatial dependency modeling module according to an embodiment of this application; Figure 3 shows an exemplary structural block diagram of an SRv6 network dynamic path optimization system with fusion fault prediction according to an embodiment of this application. Detailed Implementation
[0023] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0024] It should be understood that the terms "comprising" and "including" used in the specification and claims of this application indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.
[0025] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this specification and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this specification and claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations.
[0026] Figure 1 shows an exemplary flowchart of an SRv6 network dynamic path optimization method 100 with fusion fault prediction according to an embodiment of this application.
[0027] As shown in Figure 1, in step S110, real-time performance data and topology information of the SRv6 network are obtained, and anomaly identification and location are performed based on the real-time performance data.
[0028] In the embodiments of this application, real-time performance data includes link latency, link bandwidth utilization, port packet loss rate, bit error rate, etc.
[0029] In the embodiments of this application, the actual devices in the SRv6 network of the topology information network, the physical connections connecting these actual devices, the logical services in the SRv6 network, the entities represented by function blocks or segment IDs (SIDs) in the SRv6 network, and the path segments in the SRv6 routes are included.
[0030] In the embodiments of this application, during the execution of step S110, firstly, real-time performance data of the entire SRv6 network link is collected at a preset frequency, and the topology information of the SRv6 network is also collected. Secondly, an abnormal indicator is confirmed based on the real-time performance data of the entire SRv6 network link using a sliding time window-based judgment mechanism. Then, the fault or congestion location stage begins.
[0031] In the embodiments of this application, the sliding time window judgment mechanism works as follows: within a certain time window, if any performance indicator exceeds its preset threshold, it is determined to be an abnormal indicator. For example, when the bandwidth utilization of a certain link continuously exceeds 90% for three consecutive sampling periods, it is determined to be an abnormal indicator and a monitoring alarm is triggered.
[0032] In the embodiments of this application, during the fault or congestion localization phase, the network is first abstracted as a directed graph. Here, V represents the vertex set, which includes network nodes and logical entities capable of performing specific SRv6 functions, and E represents the edge set, which includes physical links and virtual SRv6 path segments. Next, all confirmed abnormal indicators are summarized into an abnormal indicator set. , Let m be the m-th abnormal indicator. Next, a Bayesian probabilistic model is used to locate the set of potential faulty links. .
[0033] In the embodiments of this application, the Bayesian probability model is: Among them, prior probability The likelihood function reflects the fault tendency of each element in the SRv6 network. This represents the probability of anomaly index O when the fault set is F.
[0034] Specifically, prior probability The calculation is based on the independent failure assumption of each link. Failure probability It is determined by a combination of physical attributes, historical fault data, and the output of the link fault prediction model. ,in is the link's feature vector (containing factors such as node SRv6 support capability, number of configured SIDs, number of historical segment processing errors, proportion of SRv6 traffic carried by the link, and reliability score output by the link failure prediction model, etc.), while w and b are model parameters. It is the sigmoid function. The prior probability is... .
[0035] Specifically, the likelihood function The abnormal indicator is given independently by F, for each , ,in It is a link Fault caused The probability of an anomaly is estimated using historical data. The overall likelihood is... .
[0036] Based on step S110, by comprehensively collecting real-time performance data and SRv6 physical-logical topology information, and combining it with a sliding window mechanism, a foundation is laid for dynamic and accurate perception of network anomalies. Its core advantage lies in constructing a heterogeneous directed graph that integrates physical infrastructure and SRv6 logical entities, and innovatively applying a Bayesian probability model for fault location. This model deeply integrates SRv6-specific features (such as SID configuration and traffic ratio) with historical data when calculating prior probabilities, and quantifies the probability of a fault causing anomalies through a likelihood function. This allows for the scientific and accurate inference of potential faulty link sets from surface anomaly indicators in complex and uncertain network environments, greatly improving the intelligence and accuracy of SRv6 network fault root cause analysis.
[0037] After completing step S110, in step S120, a heterogeneous graph containing a physical topology layer and an SRv6 logical layer is constructed based on real-time performance data and topology information.
[0038] In the embodiments of this application, during the execution of step S120, firstly, a physical topology layer diagram is constructed based on the topology information. And assign the real-time performance data to the corresponding elements of the physical topology layer graph, where, For physical network nodes, This is a set of physical links. Secondly, an SRv6 logical layer diagram is constructed based on the topology information. ,in, A collection of logical functional entities for deploying SRv6 functionalities (such as End.DT6, End.B6.Encaps, etc.). This is the set of virtual path segments defined by the Segment list. Next, construct the association matrix between the physical topology layer diagram and the SRv6 logical layer diagram. Among them, the correlation matrix Describes logical entities With physical nodes The bearing relationship, in physical nodes It carries logical entities When, then the correlation matrix The corresponding element in the diagram is 1 if it is not 0 otherwise. Then, the physical topology diagram, the SRv6 logical diagram, and the association matrix are merged to construct a two-layer heterogeneous network topology diagram.
[0039] Based on step S120, a two-layer heterogeneous network topology was successfully established by constructing a physical topology layer diagram that integrates real-time performance data and an SRv6 logical layer diagram that focuses on functional entities and virtual paths, and by using an association matrix to accurately define the bearer mapping between the two. The core advantage of this modeling approach lies in achieving deep cross-layer integration of physical infrastructure and logical services. This not only provides a comprehensive and three-dimensional digital representation of the network, clearly revealing the complex dependencies between underlying resources and upper-layer services, but also provides a structured, multi-dimensional, and information-rich high-quality data foundation for subsequent fault analysis and intelligent decision-making.
[0040] After step S120 is completed, in step S130, the heterogeneous graph is used as input, and the pre-trained link fault prediction model outputs the predicted fault probability value and potential fault type of each link.
[0041] In the embodiments of this application, the link failure prediction model includes a spatiotemporal feature extraction module, a spatial dependency modeling module, and a heterogeneous information fusion and prediction module.
[0042] In the embodiments of this application, during the operation of the spatiotemporal feature extraction module, firstly, a gated recurrent unit and a one-dimensional causal convolutional layer are used in parallel to process the time series of multidimensional performance indicators in the heterogeneous graph. Then, a gated fusion unit weighted and merged the output of the gated recurrent unit and the output of the one-dimensional causal convolutional layer to generate preliminary temporal feature representations for each link.
[0043] In the embodiments of this application, the following calculation formula is used in the process of processing the time series of multidimensional performance indicators in heterogeneous graphs using a gated loop unit: , Let be the hidden state of link e at time t. Let e be the hidden state of link e at time t-1. Let be the candidate hidden state of link e at time t. , This is the first weight matrix. The multidimensional performance index of link e at time t. for and The splicing matrix, It is the Sigmoid activation function. For the first bias term, , , This is the second weight matrix. For the second bias term, The hyperbolic tangent activation function is used. For Hadama accumulation, for and The splicing matrix.
[0044] In the embodiments of this application, the calculation formula for the preliminary timing characteristics of each link is as follows: , This represents the preliminary time-series characteristics of link e at time t. Let be the hidden state of link e at time t. The output of a one-dimensional causal convolutional layer. These are the gating parameters.
[0045] In embodiments of this application, the spatial dependency modeling module simultaneously processes the physical topology layer and the SRv6 logic layer in the heterogeneous graph.
[0046] In the embodiments of this application, the specific working process of the spatial dependency modeling module can be seen in Figure 2.
[0047] Figure 2 shows an exemplary flowchart of the working process of the spatial dependency modeling module in an embodiment of this application.
[0048] As shown in Figure 2, in step S210, for any node in the heterogeneous graph, its neighboring nodes are divided into different relation type sets. In step S220, the corresponding initial attention coefficients are calculated based on any node and its neighboring nodes belonging to the relation type. In step S230, the corresponding initial attention coefficients under the same relation are normalized to obtain the final attention coefficients. In step S240, the update features of each node are obtained based on the final attention coefficients.
[0049] In the embodiments of this application, the aforementioned relationship types include physical neighbors, logical functional neighbors, and cross-layer associated neighbors.
[0050] In the embodiments of this application, the formula for calculating the initial attention coefficient is: , It is a non-linear activation function. Let r be a vector of learnable weights for relation r. For learnable weights of relation r, Let i be the feature vector of node i. Let j be the feature vector of node j. for and The concatenated vector.
[0051] In the embodiments of this application, the formula for calculating the final attention coefficient is as follows: , Let i be the set of neighbors of node i under relation r. It is an exponential function.
[0052] In the embodiments of this application, the calculation formula for the update features of each node is as follows: , Let i be the updated feature of node i. This is the Sigmoid activation function.
[0053] In the embodiments of this application, the heterogeneous information fusion and prediction module first concatenates the updated features of each node with the preliminary temporal features of each link to form an enhanced feature representation for each link. Secondly, the enhanced feature representation is input to the multi-head self-attention layer to obtain the predicted fault probability value and potential fault type for each link.
[0054] In the embodiments of this application, the multi-head self-attention layer captures the most important global interaction information from the fused features of all links and nodes, employing a linear complexity attention mechanism to reduce computational complexity. The calculation formula for this attention mechanism is as follows: In this matrix, Q, K, and V represent the query matrix, key matrix, and value matrix, respectively. E and F are both low-dimensional projection matrices; E is used to project the key matrix, and F is used to project the value matrix V. is the dimension of the key vector.
[0055] In embodiments of this application, the multi-head self-attention layer includes two parallel prediction branches. Specifically, the prediction branches include a link failure probability prediction branch and a failure type classification prediction branch. The link failure prediction model is trained by minimizing the total loss of the outputs of the link failure probability prediction branch and the failure type classification prediction branch.
[0056] In the embodiments of this application, the formula for calculating the total loss is: , For the total loss, Let this be the first loss hyperparameter. For the second loss hyperparameter, This represents the probability of link failure. For real link failure labels, For the predicted failure type probability distribution, Labels for actual fault types. For binary cross-entropy loss, This represents the cross-entropy loss.
[0057] In the embodiments of this application, based on the supervised learning paradigm, a dataset containing historical link performance index data and its corresponding real fault labels (whether a fault occurred and the type of fault) is used to train the link fault prediction model.
[0058] Specifically, during training, the input data is first preprocessed. The historical performance metrics time series data are normalized to have a mean of 0 and a standard deviation of 1. The network topology information is then constructed as a heterogeneous graph structure as described above and stored in the form of adjacency and correlation matrices. The training, validation, and test sets are divided in chronological order.
[0059] Specifically, the training process employs mini-batch gradient descent, using the AdamW optimizer to adaptively adjust the learning rate and incorporate weight decay to prevent overfitting. The training loop iterates through all training batches in each iteration.
[0060] Specifically, for a batch of samples, the forward propagation first executes the spatiotemporal feature extraction module: the independent time series of each link are input into parallel gated recurrent units (GRUs) and causal convolutional layers to obtain preliminary temporal feature representations. .
[0061] Specifically, in the spatial dependency modeling module, message passing and aggregation are performed on the constructed physical and logical heterogeneous graph. Attention weights are calculated based on different neighbor relationship types, and the feature representations of all nodes are updated to capture the spatial dependencies between network elements.
[0062] Specifically, the heterogeneous information fusion and prediction module concatenates the updated node features with the final temporal features of the corresponding links to form an enhanced feature representation for each link. This fused feature is then fed into an improved multi-head self-attention layer, which calculates global attention weights through low-rank projection, focusing on key state information in all links. The output of the self-attention layer is finally fed into the multi-task output layer, simultaneously generating link failure probability prediction and failure type classification prediction.
[0063] Specifically, based on the calculated total loss using the formula described above, the gradients of all parameters are calculated using the backpropagation algorithm, and the gradient descent update rule is applied to adjust the parameters. After training, the final performance of the model is evaluated on an independent test set. The predicted potential fault type classifications are recorded in the database, and the link failure probability is introduced as a weighting factor into the comprehensive optimization cost function for path calculation.
[0064] Based on step S130, by using GRU to capture long-term dependencies and one-dimensional causal convolution to extract short-term patterns in parallel, and intelligently merging them through a gated fusion unit, the model can generate preliminary temporal features that take into account multi-scale time awareness, accurately reflecting the complex dynamics of network performance evolution over time. The spatial dependency modeling module can handle the physical topology layer and the SRv6 logical layer. By classifying neighbors into physical neighbors, logical functional neighbors, and cross-layer related neighbors, and dynamically aggregating different types of neighbor information based on a learnable attention mechanism, it effectively captures diverse and complex spatial dependencies in heterogeneous networks, surpassing the limitations of traditional graph models. The heterogeneous information fusion and prediction module concatenates the spatial features of node updates with the preliminary temporal features of links to form enhanced features, and achieves deep fusion of spatiotemporal information through a multi-head self-attention layer (using a linear complexity attention mechanism). In addition, two parallel branches, link failure probability prediction and failure type classification prediction, are designed. By minimizing the total loss during training, the model can not only predict whether a failure will occur, but also predict what type of failure will occur, providing more comprehensive and instructive prediction results.
[0065] After completing step S130, in step S140, a comprehensive optimization cost function is defined by combining the real-time performance data of the SRv6 network and the predicted failure probability values of each network link, and the optimal primary path and the optimal backup path are calculated based on the comprehensive optimization cost function.
[0066] In the embodiments of this application, the comprehensive optimization cost function is: , The comprehensive optimization cost for path P is... The available bandwidth of link e. The weighting factor is the available bandwidth. For the delay of link e, The weighting factor for time delay. Let e be the packet loss rate. This is the weighting coefficient for the packet loss rate. This is the predicted failure probability value for link e. These are the weighting coefficients for the predicted failure probability values.
[0067] In the embodiments of this application, during the process of calculating the optimal primary path and the optimal backup path based on the comprehensive optimization cost function, the path with the smallest comprehensive cost is taken as the primary path, and the path with the second smallest comprehensive cost is taken as the backup path. The relevant information of the backup path is also pre-calculated and distributed in order to enable rapid switching when the primary path encounters problems again.
[0068] Based on step S140, by incorporating the predicted link failure probability into the comprehensive optimization cost function, potentially risky links can be proactively avoided before a failure occurs, thereby significantly improving the reliability and stability of network paths and transforming passive response into proactive prevention. The comprehensive cost function integrates real-time performance indicators such as available bandwidth, latency, and packet loss rate, ensuring that path selection considers not only failure risk but also network resource utilization, transmission efficiency, and user experience, achieving more comprehensive service quality optimization. It not only calculates the optimal primary path but also simultaneously determines the optimal backup path and pre-calculates and distributes relevant information. This enables rapid and intelligent path switching when the primary path encounters problems, minimizing service interruption time and significantly enhancing the fault tolerance and availability of the SRv6 network. Through configurable weight coefficients… , , , It can flexibly adjust the importance of various indicators in path selection according to different business needs or network strategies, so as to achieve more refined path control.
[0069] After step S140 is completed, in step S150, the corresponding Segment list is obtained according to the optimal primary path and the optimal backup path, and the Segment list is sent to the source node to realize the traffic switching to the new path.
[0070] In the embodiments of this application, the segment list not only includes the SIDs of each node traversed by the path, but also embeds different types of SRv6 segments according to business needs, such as functional segments representing specific forwarding behaviors or custom SIDs bound to specific network services.
[0071] In the embodiments of this application, when creating the segment list, the segment type is selected and combined according to the actual needs of the path. For strictly specified path segments that require explicit specification, the Node SID is used. If it is necessary to instruct intermediate nodes to perform specific operations, such as traffic counting, performance measurement, or service chain processing, the corresponding Function SID can be inserted. In addition, the TLV field in the SRv6 header is used to carry path attribute parameters such as bandwidth reservation and latency limit, enabling intermediate nodes to implement more granular resource allocation and policy execution based on this information.
[0072] In the embodiments of this application, when the Segment list is sent to the source node, it is sent via a southbound protocol such as NETCONF or RESTCONF and encoded using the YANG data model. Upon receiving the policy, the source node writes it into its forwarding table and binds it to a local policy identifier or destination address prefix. When a data packet arrives, the source node adds an SRH header according to the matching policy and forwards it hop-by-hop according to the SID list order, thereby switching the traffic to the new path.
[0073] Based on step S150, the segment list not only includes Node SIDs but also embeds functional segments representing specific forwarding behaviors according to business needs. It also utilizes the TLV field in the SRv6 header to carry path attribute parameters such as bandwidth reservation and latency limits. This enables the network to achieve highly flexible and programmable traffic path control and intermediate node behavior definition, thereby meeting diverse business requirements and stringent quality of service assurance. By distributing the segment list using standard southbound protocols such as NETCONF or RESTCONF combined with the YANG data model, network configuration is automated and standardized, significantly improving path deployment efficiency and reducing errors caused by manual intervention. Upon receiving the segment list, the source node immediately writes it into its forwarding table and binds it, then adds an SRH header for hop-by-hop forwarding according to the matching strategy. This ensures that in the event of changes in network conditions or primary path failure, traffic can be quickly switched to a new primary or backup path, minimizing service interruption time and significantly improving network response speed and overall resilience.
[0074] After step S150 is completed, in step S160, the real-time performance of the new path is evaluated, and the evaluation results are fed back to the link failure prediction model for iterative training of the link failure prediction model.
[0075] In the embodiments of this application, during the evaluation of the real-time performance of the new path, the end-to-end performance improvement ratio, latency-sensitive traffic, path stability index, and global load balancing index are calculated based on the real-time performance of the new path.
[0076] In the embodiments of this application, the following calculation formula is used in calculating the end-to-end performance improvement ratio: , For end-to-end performance improvement, and These represent the effective bandwidth that the new path and the old path can provide, respectively. This represents the delay of the old path.
[0077] In the embodiments of this application, the following calculation formula is used in the calculation of latency-sensitive traffic: , For latency-sensitive traffic, The delay is due to the old path. This is the delay for the new path.
[0078] In the embodiments of this application, the following calculation formula is used in the process of calculating the path stability index: , This is a path stability index. This is the time it takes for the new path to remain stable. The observation period is defined as the time of observation. The path stability index is measured by the percentage of time during which the new internal path itself is not further optimized.
[0079] In the embodiments of this application, the load balancing of all links in the entire network after the formation of a new path is evaluated using a global load balancing index. The following calculation formula is used in the process of calculating the global load balancing index: , This is a global load balancing metric, where N is the total number of links. It is the utilization rate of the i-th link. It is the average utilization rate of the entire network.
[0080] In the embodiments of this application, the above evaluation results are continuously collected and recorded in a database. This data is used to generate intuitive visualization reports for network operations personnel to evaluate the effectiveness of their decisions. Simultaneously, it serves as training data fed back to the link failure prediction model for periodic incremental training and iterative optimization.
[0081] Based on step S160, the actual operational effect of the new path is fed back to the link failure prediction model, forming a closed-loop feedback mechanism of "prediction-execution-evaluation-re-prediction". This enables the link failure prediction model to continuously learn and adapt to the ever-changing network environment and actual operating conditions, thereby continuously improving its prediction accuracy and robustness. A comprehensive and quantitative evaluation of the new path's performance is conducted by calculating multiple key indicators such as end-to-end performance improvement ratio, latency-sensitive traffic, path stability index, and global load balancing. This not only provides network operators with intuitive and rich data support to evaluate the effectiveness of decisions but also provides multi-dimensional goals for model optimization. The visualized reports generated from the evaluation results greatly facilitate network operators' understanding and decision-making regarding network status and optimization effects. Simultaneously, as high-quality training data, it directly drives the enhancement of model capabilities, providing a solid foundation for more intelligent and autonomous network operations in the future.
[0082] In summary, through the SRv6 network dynamic path optimization scheme with integrated fault prediction provided above, this application embodiment achieves a shift from passive response to proactive fault avoidance by real-time perception of network status and prediction of link faults using a pre-trained model. By proactively considering future fault risks during path optimization and defining a comprehensive cost function based on real-time data, a more robust and reliable primary / backup path is calculated, effectively reducing network outages. Furthermore, it supports rapid path switching and a closed-loop self-learning mechanism, enabling continuous iterative optimization of fault prediction and path strategies, ultimately achieving more intelligent, efficient, and reliable dynamic management of the SRv6 network.
[0083] Furthermore, in some embodiments, a comprehensive network status view is constructed by frequently collecting multi-dimensional real-time performance data and topology information. A sliding time window mechanism is employed to dynamically identify abnormal indicators, improving the sensitivity and accuracy of anomaly detection. More importantly, by abstracting the network as a physical-logical heterogeneous graph and combining it with a Bayesian probabilistic model, potential faulty links can be accurately located probabilistically, enabling in-depth analysis of the root causes of faults and providing robust inference even in uncertain environments. Together, these elements construct an intelligent, efficient, and accurate anomaly identification and fault location system, providing high-quality, forward-looking fault information for SRv6 network path optimization, significantly improving network operation and maintenance efficiency and reliability.
[0084] Furthermore, in some embodiments, by independently constructing physical topology layer diagrams and SRv6 logical layer diagrams, and utilizing an association matrix to capture the complex mapping relationships between them, a two-layer heterogeneous network topology diagram is ultimately formed. This structured, multi-dimensional unified representation not only provides a deep insight into the interdependencies between physical infrastructure and SRv6 services, significantly enhancing fault location and root cause analysis capabilities, but also provides high-quality input for advanced machine learning models such as graph neural networks, serving as a key foundation for realizing intelligent SRv6 network management and optimization.
[0085] Furthermore, in some embodiments, multi-scale temporal dependencies are captured by deep fusion of temporal features, and spatial dependencies of heterogeneous networks are accurately modeled to comprehensively understand fault propagation at both the physical and logical levels. A multi-head self-attention layer is used to achieve heterogeneous information fusion and multi-task prediction, while simultaneously outputting fault probabilities and types. This innovative design significantly improves the accuracy, comprehensiveness, and practicality of fault prediction, providing crucial forward-looking information for dynamic path optimization in SRv6 networks, ensuring high reliability and performance of network services.
[0086] This application also provides an SRv6 network dynamic path optimization system with fusion fault prediction. It can use the aforementioned SRv6 network dynamic path optimization method 100 with fusion fault prediction to perform SRv6 network dynamic path optimization, or it can use other methods to perform SRv6 network dynamic path optimization with fusion fault prediction. This application does not limit it here.
[0087] Figure 3 shows an exemplary structural block diagram of the SRv6 network dynamic path optimization system with fusion fault prediction according to an embodiment of this application.
[0088] As shown in Figure 3, the system 300 includes a data monitoring module 310, a heterogeneous graph construction module 320, a fault prediction module 330, a path decision module 340, a strategy formation and distribution module 350, and an evaluation and feedback module 360. In the embodiments of this application, the data monitoring module 310, the heterogeneous graph construction module 320, the fault prediction module 330, the path decision module 340, the strategy formation and distribution module 350, and the evaluation and feedback module 360 can be located in the same controller or can be separate units.
[0089] Specifically, the data monitoring module 310 is used to acquire real-time performance data and topology information of the SRv6 network, and to identify and locate anomalies based on the real-time performance data.
[0090] Specifically, the heterogeneous graph construction module 320 is used to construct a heterogeneous graph containing a physical topology layer and an SRv6 logical layer based on real-time performance data and topology information.
[0091] Specifically, the fault prediction module 330 is used to take the heterogeneous graph as input and output the predicted fault probability value and potential fault type of each link through the pre-trained link fault prediction model.
[0092] Specifically, the path decision module 340 is used to define a comprehensive optimization cost function by combining the real-time performance data of the SRv6 network and the predicted failure probability values of each network link, and to calculate the optimal primary path and the optimal backup path based on the comprehensive optimization cost function.
[0093] Specifically, the strategy formation and distribution module 350 obtains the corresponding Segment list based on the optimal primary path and the optimal backup path, and distributes the Segment list to the source node to achieve traffic switching to the new path.
[0094] Specifically, the evaluation and feedback module 360 is used to evaluate the real-time performance of the new path and feed the evaluation results back to the link failure prediction model for iterative training. When the system 300 uses the aforementioned SRv6 network dynamic path optimization method 100 with fused failure prediction for dynamic path optimization, the data monitoring module 310 executes the aforementioned step S110, the heterogeneous graph construction module 320 executes the aforementioned step S120, the failure prediction module 330 executes the aforementioned step S130, the path decision module 340 executes the aforementioned step S140, the policy formation and distribution module 350 executes the aforementioned step S150, and the evaluation and feedback module 360 executes the aforementioned step S360. The specific execution process can be found above and will not be repeated here.
[0095] While numerous embodiments of this application have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will arise for those skilled in the art without departing from the spirit and intent of this application. It should be understood that various alternatives to the embodiments of this application described herein may be employed in the practice of this application. The appended claims are intended to define the scope of protection of this application and therefore cover equivalents or alternatives within the scope of these claims.
Claims
1. A dynamic path optimization method for SRv6 networks that incorporates fault prediction, characterized in that, include: Acquire real-time performance data and topology information of the SRv6 network, and perform anomaly identification and location based on the real-time performance data; Based on the real-time performance data and topology information, a heterogeneous graph containing a physical topology layer and an SRv6 logical layer is constructed. Using the heterogeneous graph as input, a pre-trained link failure prediction model outputs the predicted failure probability and potential failure type of each link; a comprehensive optimization cost function is defined by combining the real-time performance data of the SRv6 network and the predicted failure probability of each network link, and the optimal primary path and optimal backup path are calculated based on the comprehensive optimization cost function; the corresponding Segment list is obtained according to the optimal primary path and optimal backup path, and the Segment list is sent to the source node to realize the traffic switching to the new path; The real-time performance of the new path is evaluated, and the evaluation results are fed back to the link failure prediction model for iterative training.
2. The SRv6 network dynamic path optimization method with fusion fault prediction as described in claim 1, characterized in that, In the process of acquiring real-time performance data and topology information of the SRv6 network, and identifying and locating anomalies based on the real-time performance data, the following steps are performed: Real-time performance data of the entire SRv6 network link is collected at a preset frequency, and the topology information of the SRv6 network is also collected. The real-time performance data includes link latency, link bandwidth utilization, port packet loss rate, and bit error rate. An anomaly indicator is confirmed based on the real-time performance data of the entire SRv6 network link using a sliding time window-based judgment mechanism. The process then proceeds to the fault or congestion location stage, where the following steps are performed: The network is abstracted as a directed graph. Where V is the vertex set, which includes network nodes and logical entities capable of performing specific SRv6 functions; E is the edge set, which includes physical links and virtual SRv6 path segments; all confirmed abnormal indicators are summarized into an abnormal indicator set. ， The m-th abnormal indicator; the set of potential fault links is located using a Bayesian probabilistic model. 。 3. The SRv6 network dynamic path optimization method with fusion fault prediction as described in claim 1, characterized in that, In the process of constructing a heterogeneous graph containing a physical topology layer and an SRv6 logical layer based on the real-time performance data and topology information, the following steps are performed: Constructing a physical topology layer graph based on the topology information. The real-time performance data is then assigned to the corresponding elements of the physical topology layer graph, wherein... For physical network nodes, It is a set of physical links; an SRv6 logical layer diagram is constructed based on topology information. ,in, A collection of logical functional entities for deploying SRv6 functionality. A set of virtual path segments defined by the Segment list; construct the association matrix between the physical topology layer diagram and the SRv6 logical layer diagram. The physical topology layer diagram, SRv6 logical layer diagram, and association matrix are merged to construct a two-layer heterogeneous network topology diagram.
4. The SRv6 network dynamic path optimization method with fusion fault prediction according to claim 1, characterized in that, The link failure prediction model includes a spatiotemporal feature extraction module, which performs the following steps: processing the time series of multidimensional performance indicators in the heterogeneous graph using a gated recurrent unit and a one-dimensional causal convolutional layer in parallel; weighting and merging the output of the gated recurrent unit with the output of the one-dimensional causal convolutional layer using a gated fusion unit to generate preliminary temporal feature representations for each link; wherein, the following calculation formula is used in the process of processing the time series of multidimensional performance indicators in the heterogeneous graph using the gated recurrent unit: ， Let be the hidden state of link e at time t. Let e be the hidden state of link e at time t-1. Let be the candidate hidden state of link e at time t. ， This is the first weight matrix. The multidimensional performance index of link e at time t. for and The splicing matrix, It is the Sigmoid activation function. For the first bias term, ， ， This is the second weight matrix. For the second bias term, The hyperbolic tangent activation function is used. For Hadama accumulation, for and The splicing matrix; the calculation formula for the preliminary time-series characteristics of each link is: ， This represents the preliminary time-series characteristics of link e at time t. Let be the hidden state of link e at time t. The output of a one-dimensional causal convolutional layer. These are the gating parameters.
5. The SRv6 network dynamic path optimization method with fusion fault prediction as described in claim 4, characterized in that, The link failure prediction model also includes a spatial dependency modeling module, which simultaneously processes the physical topology layer and the SRv6 logical layer in the heterogeneous graph. The spatial dependency modeling module performs the following steps: For any node in the heterogeneous graph, its neighboring nodes are divided into different sets of relation types, including physical neighbors, logical functional neighbors, and cross-layer related neighbors; based on any node and its neighboring nodes belonging to the relation types, the corresponding initial attention coefficient is calculated, wherein the formula for calculating the initial attention coefficient is: ， It is a non-linear activation function. Let r be a vector of learnable weights for relation r. For learnable weights of relation r, Let i be the feature vector of node i. Let j be the feature vector of node j. for and The concatenated vector; the initial attention coefficients under the same relation are normalized to obtain the final attention coefficients, where the formula for calculating the final attention coefficients is: ， Let i be the set of neighbors of node i under relation r. It is an exponential function; the update features of each node are obtained based on the final attention coefficient, where the calculation formula for the update features of each node is: ， Let i be the updated feature of node i. This is the Sigmoid activation function.
6. The SRv6 network dynamic path optimization method with fusion fault prediction as described in claim 5, characterized in that, The link failure prediction model also includes a heterogeneous information fusion and prediction module. This module performs the following steps: concatenating the updated features of each node with the preliminary temporal features of each link to form an enhanced feature representation for each link; inputting the enhanced feature representation into a multi-head self-attention layer to obtain the predicted failure probability and potential failure type for each link; wherein the multi-head self-attention layer includes two parallel prediction branches; each prediction branch includes a link failure probability prediction branch and a failure type classification prediction branch; the link failure prediction model is trained by minimizing the total loss of the outputs of the link failure probability prediction branch and the failure type classification prediction branch; the formula for calculating the total loss is: ， For the total loss, Let this be the first loss hyperparameter. For the second loss hyperparameter, This represents the probability of link failure. For real link failure labels, For the predicted failure type probability distribution, Labels for actual fault types. For binary cross-entropy loss, This represents the cross-entropy loss.
7. The SRv6 network dynamic path optimization method with fusion fault prediction according to claim 1, characterized in that, The comprehensive optimization cost function is: ， The comprehensive optimization cost for path P is... The available bandwidth of link e. The weighting factor is the available bandwidth. For the delay of link e, The weighting coefficient for time delay. Let e be the packet loss rate. This is the weighting coefficient for the packet loss rate. This is the predicted failure probability value for link e. These are the weighting coefficients for the predicted failure probability values.
8. The SRv6 network dynamic path optimization method with fusion fault prediction according to claim 7, characterized in that, In the process of calculating the optimal primary path and the optimal backup path based on the comprehensive optimization cost function, the path with the smallest comprehensive cost is selected as the primary path, and the path with the second smallest comprehensive cost is selected as the backup path.
9. The SRv6 network dynamic path optimization method with fusion fault prediction according to claim 1, characterized in that, In the process of evaluating the real-time performance of the new path, the end-to-end performance improvement ratio, latency-sensitive traffic, path stability index, and global load balancing index are calculated based on the real-time performance of the new path.
10. A dynamic path optimization system for SRv6 networks integrating fault prediction, characterized in that, The system employs the SRv6 network dynamic path optimization method with fused fault prediction as described in any one of claims 1-9. The system comprises: a data monitoring module for acquiring real-time performance data and topology information of the SRv6 network, and performing anomaly identification and location based on the real-time performance data; a heterogeneous graph construction module for constructing a heterogeneous graph including a physical topology layer and an SRv6 logical layer based on the real-time performance data and topology information; and a fault prediction module for taking the heterogeneous graph as input and outputting the fault probability of each link through a pre-trained link fault prediction model. The system includes: a prediction value and potential fault type module; a path decision module, which combines real-time performance data of the SRv6 network with the predicted fault probability values of each network link to define a comprehensive optimization cost function, and calculates the optimal primary path and optimal backup path based on the comprehensive optimization cost function; a policy formation and distribution module, which obtains the corresponding Segment list based on the optimal primary path and optimal backup path, and distributes the Segment list to the source node to realize traffic switching to the new path; and an evaluation and feedback module, which evaluates the real-time performance of the new path and feeds the evaluation results back to the link fault prediction model for iterative training of the link fault prediction model.