Communication network fault rapid positioning and recovery method based on self-supervised learning

Abstract

The invention discloses a communication network fault rapid positioning and recovery method based on self-supervised learning. The method comprises the following steps: S1, collecting and preprocessing communication network operation data; s2, constructing a communication network topological graph, and mapping the preprocessed data into node and edge attributes; s3, executing feature mask and structure disturbance, and constructing positive and negative comparison samples; s4, constructing a graph neural network model and pre-training by using positive and negative comparison samples; s5, calculating reconstruction errors of the nodes and the edges, and marking elements higher than a threshold value as abnormal candidate areas; s6, performing graph segmentation to extract continuous abnormal sub-graphs, and calculating sub-graph embedding and historical feature center similarity; and S7, matching an instruction sequence according to a positioning result, executing link scheduling, node reconfiguration and topology updating, and iteratively training the model. According to the method, high-precision rapid positioning and automatic recovery of communication network faults are realized under the condition of lack of a large number of labeled samples, and the method has a continuous optimization capability.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and in particular to a method for rapid fault location and recovery in communication networks based on self-supervised learning. Background Technology

[0002] With the widespread deployment of cloud computing, the Internet of Things (IoT), and 5G communication technologies, the scale and complexity of communication networks are constantly increasing, and network topologies are characterized by a large number of nodes, complex connections, and diverse service types. In large-scale network environments, problems such as link interruptions, node failures, configuration errors, and traffic bursts occur frequently. These problems not only affect the normal operation of services but also cause widespread network congestion and performance degradation. Therefore, accurately locating network faults and quickly restoring network operation in the shortest possible time has become a core requirement in the field of communication network operation and maintenance.

[0003] Existing methods for fault detection and localization in communication networks primarily rely on rule-based monitoring mechanisms and human experience-based analysis. These methods monitor network operating parameters by setting pre-defined alarm rules and fixed thresholds, triggering alarms when parameters exceed these thresholds. However, due to dynamic changes in network traffic patterns and topology, fixed thresholds often fail to accurately adapt to different business scenarios, leading to a large number of false alarms and missed alarms. Furthermore, rule-based alarm mechanisms lack a deep understanding of complex fault modes, making it difficult to accurately analyze the root cause of faults when multiple points, links, and types of anomalies occur simultaneously.

[0004] Supervised learning is widely used in existing intelligent fault diagnosis methods. It involves constructing training sets to classify or regress network operation data, thereby identifying fault types and locations. However, these methods rely heavily on large-scale, high-quality, and fully labeled training data. In real communication networks, the number of fault samples is limited and unevenly distributed. Obtaining comprehensive labeled data is extremely costly, and in some novel network architectures, sufficient historical labeled data may not be available, resulting in insufficient model generalization ability and poor transferability.

[0005] Existing technologies generally suffer from the following drawbacks: First, they rely heavily on labeled samples, making it difficult for models to adapt to network environments lacking training samples; second, the fault diagnosis and recovery processes are disconnected, lacking automation and closed-loop optimization capabilities; third, the graph structure modeling is insufficient, failing to effectively address issues such as feature loss, structural disturbances, and multi-source dynamic data fusion; and fourth, the recovery strategy execution method is rigid, unable to adaptively generate the optimal recovery instruction sequence based on the fault location results.

[0006] Therefore, how to provide a method for rapid fault location and recovery in communication networks based on self-supervised learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose a method for rapid fault location and recovery in communication networks based on self-supervised learning. This invention fully utilizes self-supervised learning, graph neural network modeling, and intelligent instruction mapping and execution techniques. It details how, in the absence of labeled data, positive and negative comparison samples are constructed through feature masks and structural perturbations, and model parameters are optimized to achieve accurate fault location. Furthermore, it automatically generates and executes link scheduling, node reconfiguration, and topology updates based on fault type. This method has the advantages of not requiring a large number of labeled samples, high location accuracy, fast recovery speed, and continuous optimization.

[0008] A method for rapid fault location and recovery in communication networks based on self-supervised learning according to an embodiment of the present invention includes the following steps:

[0009] S1. Collect and preprocess the operational data of the communication network;

[0010] S2. Using network devices as nodes and physical or logical connections as edges, construct a communication network topology graph and map the preprocessed running data to the attribute features of the corresponding nodes or edges in the graph.

[0011] S3. Perform feature masking and structural perturbation operations on the communication network topology graph to generate masked views and perturbation views, and construct positive and negative comparison samples at the graph level and node level.

[0012] S4. Construct a graph neural network model, pre-train the graph neural network model using positive and negative contrast samples, optimize the model parameters through the contrastive loss function, and complete the self-supervised pre-training process.

[0013] S5. Apply the communication network topology graph to the graph neural network model, calculate the reconstruction error distribution of each node and edge, and mark the nodes or edges with reconstruction errors higher than the preset threshold as abnormal candidate regions.

[0014] S6. Based on the abnormal candidate region, perform graph structure segmentation, extract subgraph regions with continuous abnormal nodes and edges, calculate the similarity between the subgraph embedding vector and the historical feature center, and generate fault location results.

[0015] S7. Match the corresponding instruction sequence based on the fault location result, execute link scheduling, node reconfiguration and topology update operations, and continuously iterate and train the graph neural network model.

[0016] Optionally, the operational data includes link status information, node forwarding tables, traffic statistics, and historical alarm logs.

[0017] Optionally, the preprocessing involves performing field normalization, timestamp alignment, missing value completion, and outlier removal on the running data.

[0018] Optionally, S2 specifically includes:

[0019] S21. Extract each network device with a unique identifier in the communication network, define the network device as a node in the graph structure, and set a node attribute field for each node. The node attribute field includes device identifier, device type code, device operating status identifier and device network level number.

[0020] S22. Extract connection paths based on the physical and logical connection relationships between network devices, treat the connection paths as edges in the graph structure, and set edge attribute fields for each edge. The edge attribute fields include starting node identifier, target node identifier, connection type code, connection availability status, and connection capacity value.

[0021] S23. Match the preprocessed running data with the network device identifier according to the timestamp, and map the node-related numerical attributes to the target fields of the corresponding nodes. The target fields include processing capacity indicators, resource consumption indicators, forwarding load indicators and alarm frequency indicators.

[0022] S24. Match the preprocessed running data with the connection path identifier according to the timestamp, and map the numerical attributes related to the edge to the attribute fields of the corresponding edge. The attribute fields include communication quality indicators, transmission stability indicators, traffic change magnitude and path forwarding delay.

[0023] S25. The graph structure after mapping the attribute features of nodes and edges is used as the topology graph of the communication network.

[0024] Optionally, S3 specifically includes:

[0025] S31. Represent the communication network topology as a graph structure consisting of a set of nodes, a set of edges, and a node feature matrix, wherein the node feature matrix contains the multi-dimensional numerical features corresponding to each node.

[0026] S32. Perform a feature masking operation, randomly select some feature elements in the node feature matrix according to a set ratio, replace these feature elements with constant zero to form a new mask feature matrix, and construct a mask view while retaining the original topology. The mask view is missing some node feature information while maintaining the same structure.

[0027] S33. Perform a structural perturbation operation, delete some existing edges in the original topology according to a set ratio and insert an equal number of new edges to generate a perturbation view. The perturbation view changes some connection relationships while keeping the node characteristics unchanged.

[0028] S34. The communication network topology graph and the masked view are used to form graph-level positive contrast samples, and the communication network topology graph and the perturbation view are used to form graph-level negative contrast samples for graph contrast loss calculation in graph embedding representation training.

[0029] S35. For each node in the communication network topology graph, extract its embedding representation in the communication network topology graph, mask view, and perturbation view. Construct node-level positive contrast samples by combining the embeddings in the communication network topology graph and the mask view, and construct node-level negative contrast samples by combining the embeddings in the communication network topology graph and the perturbation view. These samples are used for calculating the node contrast loss in the node embedding representation training.

[0030] Optionally, S4 specifically includes:

[0031] S41. Construct a graph neural network model consisting of an encoder and a projector. Input the communication network topology, mask view and perturbation view into the encoder in sequence, and output the corresponding node embedding matrix respectively. The node embedding matrix extracts the embedding vector of each node according to the pairing relationship between node-level positive sample pairs and node-level negative sample pairs, maintaining a one-to-one correspondence.

[0032] S42. Construct a projector to perform feature mapping on node embeddings to obtain node projection vectors. Use cosine similarity and temperature coefficient to calculate similarity. All calculations are based on the fixed pairing relationship between node-level positive sample pairs and node-level negative sample pairs.

[0033] S43. Calculate the similarity between the projection vectors of corresponding nodes in the communication network topology diagram and the masked view, and compare it with the similarity between the projection vectors of corresponding nodes in the communication network topology diagram and the perturbation view to form a node-level contrast loss.

[0034] S44. Combining graph-level positive sample pairs and graph-level negative sample pairs, the node embedding matrix is ​​used to generate a graph representation vector through average pooling. The similarity between the communication network topology graph and the masked view is calculated and compared with the similarity between the communication network topology graph and the perturbation view to form a graph-level contrast loss.

[0035] S45. The node-level contrastive loss and the graph-level contrastive loss are weighted and fused to form the total loss function. The encoder parameters and projector parameters are iteratively updated by minimizing the total loss function until the training converges, thus completing the self-supervised pre-training process.

[0036] Optionally, S5 specifically includes:

[0037] S51. Apply the communication network topology graph to the graph neural network model, and use a projector to reconstruct node features and edge features to generate node feature reconstruction matrices and edge feature reconstruction matrices.

[0038] S52. For each node, calculate the absolute difference between the true feature vector and the reconstructed feature vector in each dimension, and average the differences in all feature dimensions to obtain the reconstruction error value of the node features.

[0039] S53. For each edge, calculate the absolute difference between the true feature vector and the reconstructed feature vector in each dimension, and average the differences in all feature dimensions to obtain the reconstruction error value of the edge feature.

[0040] S54. Based on the preset node error threshold and edge error threshold, nodes with reconstruction error values ​​higher than the threshold are marked as nodes in the abnormal candidate region, and edges with reconstruction error values ​​higher than the threshold are marked as edges in the abnormal candidate region, for use in communication network fault location.

[0041] Optionally, S6 specifically includes:

[0042] S61. Based on the set of nodes and the set of edges of the abnormal candidate regions, perform connectivity analysis on the communication network topology graph, and divide nodes and edges with direct or indirect connections into the same connected component to form multiple subgraph regions.

[0043] S62. Input each subgraph into the encoder in the graph neural network model to obtain the corresponding subgraph encoding representation, and perform average pooling operation on the subgraph encoding representation to obtain the subgraph embedding vector representing the overall features of the subgraph.

[0044] S63. Perform similarity calculation between the subgraph embedding vector and the pre-stored historical feature centers to determine the fault category label corresponding to the maximum similarity value, and combine the fault type label with the topological location range of the subgraph to form a fault location result set containing location range and category information.

[0045] Optionally, S7 specifically includes:

[0046] S71. Input the fault location result set into the instruction mapping function, and match the corresponding ordered instruction sequence according to the fault category label of each sub-map;

[0047] S72. Execute each instruction in the ordered instruction sequence in sequence. If the instruction type is link scheduling, adjust the bandwidth allocation ratio and priority of the target link. If the instruction type is node reconfiguration, update the routing table, resource allocation table and processing queue threshold of the target node. If the instruction type is topology update, perform add and delete operations in the node set and edge set to form the updated communication network topology.

[0048] S73. Re-input the updated communication network topology and running data into the graph neural network model, and continue to iteratively optimize the model parameters according to the self-supervised training method until the training process meets the convergence condition.

[0049] Optionally, the instruction mapping function retrieves a preset instruction mapping table based on the fault category label of each subgraph in the fault location result set, outputs the matched ordered instruction sequence to the execution queue, and maintains a one-to-one correspondence between the category label and the corresponding instruction sequence to ensure that different fault types can trigger corresponding link scheduling, node reconfiguration, and topology update operations.

[0050] The beneficial effects of this invention are:

[0051] First, this invention introduces a self-supervised learning method based on feature masks and structural perturbations, which can still construct effective positive and negative comparison samples even in the absence of large-scale labeled samples. It also uses a graph neural network model to extract robust node and graph structure representation features, thereby improving the accuracy and generalization ability of network fault location in dynamic topology environments.

[0052] Secondly, this invention deeply integrates fault diagnosis results with recovery strategies. Through the instruction mapping function, it automatically generates corresponding link scheduling, node reconfiguration and topology update operations according to different fault types. During the execution process, it optimizes network resource allocation and routing paths in real time, effectively shortening the time from fault detection to recovery completion.

[0053] Finally, this invention introduces a continuous iterative training mechanism for the model in the recovery phase, which re-inputs the updated network topology and operational data into the self-supervised training process, realizing a closed loop of fault location and model optimization. This improves the model's adaptability to new fault modes and its long-term stability, thereby significantly enhancing the availability, reliability, and automation level of the communication network. Attached Figure Description

[0054] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0055] Figure 1 This is a flowchart of a method for rapid fault location and recovery in communication networks based on self-supervised learning proposed in this invention;

[0056] Figure 2 This is a flowchart of the self-supervised comparative pre-training process for a fast fault location and recovery method for communication networks based on self-supervised learning proposed in this invention.

[0057] Figure 3This is a flowchart illustrating the closed-loop process of anomaly detection, localization, and recovery for a self-supervised learning-based method for rapid fault location and recovery in communication networks proposed in this invention. Detailed Implementation

[0058] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0059] refer to Figure 1-3 A method for rapid fault location and recovery in communication networks based on self-supervised learning includes the following steps:

[0060] S1. Collect and preprocess the operational data of the communication network;

[0061] S2. Using network devices as nodes and physical or logical connections as edges, construct a communication network topology graph and map the preprocessed running data to the attribute features of the corresponding nodes or edges in the graph.

[0062] S3. Perform feature masking and structural perturbation operations on the communication network topology graph to generate masked views and perturbation views, and construct positive and negative comparison samples at the graph level and node level.

[0063] S4. Construct a graph neural network model, pre-train the graph neural network model using positive and negative contrast samples, optimize the model parameters through the contrastive loss function, and complete the self-supervised pre-training process.

[0064] S5. Apply the communication network topology graph to the graph neural network model, calculate the reconstruction error distribution of each node and edge, and mark the nodes or edges with reconstruction errors higher than the preset threshold as abnormal candidate regions.

[0065] S6. Based on the abnormal candidate region, perform graph structure segmentation, extract subgraph regions with continuous abnormal nodes and edges, calculate the similarity between the subgraph embedding vector and the historical feature center, and generate fault location results.

[0066] S7. Match the corresponding instruction sequence based on the fault location result, execute link scheduling, node reconfiguration and topology update operations, and continuously iterate and train the graph neural network model.

[0067] This invention achieves a fully automated closed loop from fault detection to recovery by assembling a complete processing flow consisting of communication network operation data acquisition, topology modeling, feature perturbation, self-supervised graph neural network training, anomaly detection, subgraph similarity calculation, and instruction-based recovery execution, which significantly improves the efficiency and accuracy of fault handling.

[0068] In this embodiment, the operational data includes link status information, node forwarding table, traffic statistics information, and historical alarm logs. The link status information includes link bandwidth utilization, packet loss rate, latency value, and link activation status within the collection time window. The node forwarding table includes routing and forwarding rules, priority queue mapping, and port forwarding records for each communication node within a specified period. The traffic statistics information includes inbound traffic, outbound traffic, flow direction distribution, and traffic peak identifiers for each communication node and link within equally spaced sampling periods. The historical alarm logs include records of abnormal events triggered by network devices within a known past time period, alarm level codes, event type labels, and processing response timestamps.

[0069] This invention enriches the model input feature dimensions and improves the accuracy and generalization ability of fault mode recognition by introducing multi-dimensional operational data such as link status information, node forwarding table, traffic statistics information and historical alarm logs.

[0070] In this embodiment, the preprocessing process includes the following four steps: First, data type unification is performed on each type of field in the running data; string types are converted to discrete integer encoding, timestamps are uniformly converted to Coordinated Universal Time (UTC) format, and decimal points are uniformly retained and units are standardized for numerical fields. Second, the running data is grouped into windows using a sliding time window, and timestamp alignment is performed on the data within each window to ensure that link status, forwarding table information, traffic statistics, and alarm logs have corresponding matching records at the same sampling time point. Third, data fields with null values, missing values, or logical conflicts are filled using mean filling, median filling, or interpolation based on adjacent time-series values, with the filling method automatically selected according to the data type. Fourth, outlier detection is performed, and records with field values ​​that significantly deviate from the normal distribution range are removed. The criteria for outlier removal are sample points outside three standard deviations in the statistical distribution or outliers with a confidence level below a set threshold determined based on the isolated forest model.

[0071] This invention ensures the temporal consistency and integrity of input data through preprocessing operations such as field normalization, timestamp alignment, missing value completion, and outlier removal, thereby reducing the impact of data noise on model training and fault diagnosis.

[0072] In this embodiment, S2 specifically includes:

[0073] S21. Extract each network device with a unique identifier in the communication network, define the network device as a node in the graph structure, and set a node attribute field for each node. The node attribute field includes device identifier, device type code, device operating status identifier and device network level number.

[0074] S22. Extract connection paths based on the physical and logical connection relationships between network devices, treat the connection paths as edges in the graph structure, and set edge attribute fields for each edge. The edge attribute fields include starting node identifier, target node identifier, connection type code, connection availability status, and connection capacity value.

[0075] S23. Match the preprocessed running data with the network device identifier according to the timestamp, and map the node-related numerical attributes to the target fields of the corresponding nodes. The target fields include processing capacity indicators, resource consumption indicators, forwarding load indicators and alarm frequency indicators.

[0076] S24. Match the preprocessed running data with the connection path identifier according to the timestamp, and map the numerical attributes related to the edge to the attribute fields of the corresponding edge. The attribute fields include communication quality indicators, transmission stability indicators, traffic change magnitude and path forwarding delay.

[0077] S25. The graph structure after mapping the attribute features of nodes and edges is used as the topology graph of the communication network.

[0078] This invention precisely defines the attribute fields of nodes and edges when constructing a communication network topology graph, and accurately maps the preprocessed running data to the target fields of the corresponding elements, realizing the organic combination of topology structure and running features, and providing high-quality input for subsequent graph neural network model analysis.

[0079] In this embodiment, S3 specifically includes:

[0080] S31. Let the communication network topology be G=(V,E,X), where V represents the set of nodes, E represents the set of edges, and X represents the node feature matrix with dimension N×D, where N is the number of nodes and D is the node feature dimension.

[0081] S32. In the node feature matrix X, randomly select a set of feature elements M according to a preset ratio m. f , set M f The elements in the matrix are replaced with constant zero to obtain the mask feature matrix X. m And construct a masked view G m =(V,E,X) m The masked view weakens the feature representation of some nodes while keeping the topology unchanged.

[0082] S33. Perform a structural perturbation operation, deleting and inserting edges into the edge set E according to the perturbation ratio p, to obtain the perturbed edge set E. p Constructing the perturbation view G p =(V,E) pThe edge insertion (X) is a random insertion into non-connected node pairs. The edge deletion refers to the random deletion of edges from the edge set E. Edge;

[0083] S34. Construct graph-level positive and negative comparison samples, comparing the communication network topology graph G with the mask view G. m Defined as positive sample pairs (G, G) m ), compare the communication network topology graph G with the disturbance view G p Defined as negative sample pairs (G, G) p ), used for calculating graph contrastive loss in graph embedding representation training;

[0084] S35. Construct node-level positive and negative comparison samples for each node v in the node set V. i Extract its value in G and G m G p The corresponding embedding vector in the communication network topology graph is denoted as h. i Embedded in the mask view as Embedded in the perturbation view as Construct node-level positive sample pairs as Node-level negative sample pairs Used for calculating node contrastive loss during node embedding representation training.

[0085] This invention generates different views through feature masks and structural perturbations, and constructs positive and negative contrast samples at the node and graph levels, which effectively enhances the robustness of the model under the conditions of missing features and structural changes, and improves the sensitivity of self-supervised training to abnormal patterns.

[0086] In this embodiment, S4 specifically includes:

[0087] S41. Construct a graph neural network model consisting of an encoder and a projector, and combine the communication network topology graph G and the mask view G. m and perturbation view G p The encoder f of the graph neural network is input sequentially. θ (·), output the corresponding node embedding matrices H and H respectively. m and H p The node embedding matrix is ​​derived from G and G m and G p The corresponding embedding vector h i , and Where H = f θ (G), H m =f θ (G m ), H p =f θ(G p ), where θ is the encoder parameter in the graph neural network model;

[0088] S42, Combining node-level positive sample pairs as and node-level negative sample pairs Build a projector g φ (·) Perform feature mapping on the node embeddings to obtain different projection vectors, where z i =g φ (h i ), φ is the projector parameter in the graph neural network model;

[0089] S43. Based on the projection vector z i , Construct a node-level contrastive loss function L node :

[0090]

[0091] Where N is the number of nodes. Depend on The projection is obtained and the pairing is kept consistent. Depend on The projection obtains and maintains the pairing consistency, sim(·) represents the cosine similarity function, τ is the temperature coefficient, exp(·) represents the natural exponential function with base e, and log2(·) represents the logarithmic function;

[0092] S44, Combining graph-level positive sample pairs (G,G) m ) and graph-level positive sample pairs (G,G p For each graph, embedding matrices H and H' are used to embed the nodes. m and H p Perform average pooling separately to obtain the graph representation vector g. G , and And construct a graph-level contrastive loss function L graph :

[0093]

[0094] S45. Weighted fusion of node-level contrastive loss and graph-level contrastive loss yields the total loss function:

[0095] L total =αL node +(1-α)L graph ;

[0096] Where α is the weighting coefficient for node-level and graph-level losses, and the model parameters θ and φ are updated via gradient descent with a learning rate η. and Until L total Convergence completes the self-supervised pre-training process.

[0097] This invention optimizes encoder and projector parameters by jointly optimizing the contrast loss at the node level and graph level, achieving multi-level feature alignment and differential representation, and significantly improving the model's ability to understand complex network structures under unlabeled conditions.

[0098] In this embodiment, S5 specifically includes:

[0099] S51. Apply the communication network topology graph G to the graph neural network model, using the projector g. φ (·) Perform node feature reconstruction and edge feature reconstruction to generate a node feature reconstruction matrix. Reconstruction matrix of edge features Where P is the edge index matrix, which records the starting node index and the target node index corresponding to each edge, θ is the encoder parameter, φ is the projector parameter, n represents the node, and e represents the edge;

[0100] S52, Computing the reconstruction error vector of node features

[0101]

[0102] Among them, X ij ∈X represents node v i The true value in the j-th feature dimension. This represents the corresponding reconstructed value, where D is the node feature dimension;

[0103] S53. Calculate the reconstruction error vector of edge features.

[0104]

[0105] Among them, A kl Representing edge e k The true value in the l-th feature dimension. This represents the corresponding reconstructed value, where F is the edge feature dimension;

[0106] S54, Set the node error threshold T n With the edge error threshold T e Identify candidate regions for anomalies:

[0107]

[0108] Among them, V anom E represents the set of nodes marked as candidate regions for anomalies. anomThis represents the set of edges marked as candidate regions for anomalies, used for fault location in communication networks.

[0109] This invention uses the reconstruction error distribution of node features and edge features to determine abnormal regions, enabling accurate screening in the case of multiple types and scales of anomalies, and providing high-confidence candidate regions for subsequent fault location.

[0110] In this embodiment, S6 specifically includes:

[0111] S61, Based on node set V anom With edge set E anom Perform connectivity analysis on the communication network topology graph G, and partition nodes and edges with direct or indirect connections into the same connected component to obtain a subgraph set {G1, G2, ..., G}. K}, where K is the number of subgraphs;

[0112] S62, each subgraph G k The input is fed into the encoder f in the graph neural network model. θ (·), to obtain the corresponding subgraph encoding representation H k =f θ (G k ), and for H k Perform average pooling to generate subgraph embedding vectors:

[0113]

[0114] Among them, s k For subgraph G k The embedding vector, V k For subgraph G k The set of nodes, h i For node v i The embedding vector, |V k | represents the number of nodes in the subgraph;

[0115] S63. Let the set of historical feature centers be {c1, c2, ..., c...} M} Calculate the similarity between the subgraph embedding vector and the historical feature centers:

[0116]

[0117] Where M is the number of known historical fault types, c j Let represent the feature center vector of the j-th type of historical fault, ‖·‖ represent the vector magnitude, and sim(·) represent the cosine similarity function;

[0118] The c corresponding to the maximum similarity j Category labels as subgraph G kThe fault type labels are generated, and their location range in the topology is combined with the category information to generate a fault location result set R = {(G k ,label k )}, where label k For subgraph G k Fault type label.

[0119] This invention achieves fault classification and location determination based on pattern matching by embedding calculations into subgraphs and performing similarity analysis with historical feature centers, effectively improving the accuracy and interpretability of the location results.

[0120] In this embodiment, S7 specifically includes:

[0121] S71. Input the fault location result set R into the command mapping function:

[0122] Φ:label k →Q k ;

[0123] Where, label k Subgraph G k Fault type label, Q k ={q k1 ,q k2 ,…,q ku} represents an ordered sequence of instructions corresponding to the fault type label, where u is the number of instructions;

[0124] S72, execute the instruction sequence Q sequentially. k Each instruction q in ki ∈Q k :

[0125] If q ki If it is a link scheduling instruction, then the target link bandwidth allocation ratio β is adjusted in the edge set E. e and link priority π e ;

[0126] If q ki If a node reconfiguration command is given, the routing table, resource allocation table, and processing queue threshold of the target node in the node set V will be updated.

[0127] If q ki If the instruction is to update the topology, then add or delete nodes V and edges E to generate the updated communication network topology graph: G′=(V′,E′,X′), where V′, E′, and X′ are the updated node set, edge set, and node feature matrix, respectively.

[0128] S73. Re-input the updated communication network topology graph G′ and the running data into the graph neural network model, and continue to iteratively optimize the model parameters using a self-supervised training method. The updated loss function is:

[0129]

[0130] Where t represents the iteration round, until the convergence condition is met.

[0131] This invention maps fault location results to corresponding link scheduling, node reconfiguration, and topology update instructions, and continues to iteratively optimize the model after execution, achieving deep integration of diagnosis and recovery and shortening network recovery time.

[0132] In this embodiment, the instruction mapping function retrieves a preset instruction mapping table based on the fault category label of each subgraph in the fault location result set, outputs the matched ordered instruction sequence to the execution queue, and maintains a one-to-one correspondence between the category label and the corresponding instruction sequence to ensure that different fault types can trigger corresponding link scheduling, node reconfiguration, and topology update operations.

[0133] This invention uses a precise matching mechanism of instruction mapping functions to ensure that different fault types can trigger specific recovery actions, thereby improving the targeting of fault recovery and the degree of automation of execution.

[0134] Example 1:

[0135] To verify the feasibility of this invention in practice, it was applied to a large-scale multi-service bearer communication network environment. This network consists of a core layer, an aggregation layer, and an access layer, with over 200 nodes and nearly 500 links. Service types cover real-time voice and video transmission, data synchronization, and IoT terminal access. During peak hours, this network experiences high concurrency and link utilization, and has a complex cross-domain routing and virtualization deployment environment. Once a fault occurs, the location is wide-ranging, recovery is difficult, and manual intervention is frequent, easily leading to service interruptions and a degraded user experience. Existing threshold-based alarms and manual diagnostic methods frequently result in missed or false alarms in such networks, and the average time from fault occurrence to recovery exceeds 15 minutes, failing to meet the continuity requirements of critical services.

[0136] In the implementation of this invention, the network management platform first collects operational data in real time, including link status information, node forwarding tables, traffic statistics, and historical alarm logs. After preprocessing operations such as field normalization, timestamp alignment, missing value completion, and outlier removal, the collected data is mapped onto a communication network topology graph with network devices as nodes and physical or logical connections as edges, forming high-precision structured graph data.

[0137] After the graph data is constructed, masked views and perturbation views are generated using feature masks and structural perturbations. Positive and negative contrast samples at the graph level and node level are also constructed and input into the graph neural network model for self-supervised pre-training. During training, node-level and graph-level contrastive losses are jointly used to optimize model parameters, thereby improving the model's robustness under conditions of feature loss and topological changes.

[0138] When an anomaly occurs during network operation, a pre-trained graph neural network reconstructs the node and edge features of the topology graph, calculates the reconstruction error distribution for each node and edge, and marks elements with reconstruction errors exceeding a threshold as anomaly candidate regions. Subsequently, the system performs connectivity analysis on the subgraph, extracting consecutive anomaly nodes and edges as independent subgraph regions and calculating their similarity with historical feature centers, thereby quickly determining the type and precise location of the fault.

[0139] Once a fault is located, the system automatically matches the corresponding instruction sequence based on the fault category label and executes link scheduling, node reconfiguration, and topology update operations. For example, in a core layer link outage event, the system automatically issues instructions to switch affected services to backup paths, reconfigures the forwarding table entries of relevant nodes, and simultaneously deletes the failed link and adds backup link connections in the topology. The entire process requires no manual intervention and execution time is controlled within a few seconds.

[0140] During implementation, this invention demonstrated significant advantages in numerous network failure cases. Table 1 shows that in networks of the same size and with the same service load, the average fault location time of this invention is 3.8 seconds, and the average recovery time is 6.5 seconds, reducing the overall fault handling time by approximately 73% compared to traditional methods. The fault location accuracy rate reaches 98.7%, and the recovery success rate reaches 99.2%, representing improvements of 11.4 percentage points and 8.9 percentage points respectively compared to the original methods.

[0141] In a simulated large-scale multi-link failure test, this invention maintained a location accuracy rate of over 95% and a recovery success rate of over 90% even with a node feature loss rate of 20% and a link topology disturbance rate of 15%, verifying the robustness and adaptability of the model. Even when network peak traffic exceeded normal values ​​by 40%, this invention still maintained stable fault handling performance.

[0142] Table 1 Comparison and verification data of the fault handling performance of the present invention in multi-service communication networks.

[0143]

[0144] As can be seen from the above embodiments, the present invention not only achieves high-precision fault location in the absence of a large number of labeled samples, but also realizes a closed loop of diagnosis and repair through automated recovery strategies and continuous iterative training mechanisms, which greatly improves the availability, reliability and automation level of communication networks.

[0145] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for rapid fault location and recovery in communication networks based on self-supervised learning, characterized in that, Includes the following steps: S1. Collect and preprocess the operational data of the communication network; S2. Using network devices as nodes and physical or logical connections as edges, construct a communication network topology graph and map the preprocessed running data to the attribute features of the corresponding nodes or edges in the graph. S3. Perform feature masking and structural perturbation operations on the communication network topology graph to generate masked views and perturbation views, and construct positive and negative comparison samples at the graph level and node level. S4. Construct a graph neural network model, pre-train the graph neural network model using positive and negative contrast samples, optimize the model parameters through the contrastive loss function, and complete the self-supervised pre-training process. S5. Apply the communication network topology graph to the graph neural network model, calculate the reconstruction error distribution of each node and edge, and mark the nodes or edges with reconstruction errors higher than the preset threshold as abnormal candidate regions. S6. Based on the abnormal candidate region, perform graph structure segmentation, extract subgraph regions with continuous abnormal nodes and edges, calculate the similarity between the subgraph embedding vector and the historical feature center, and generate fault location results. S7. Match the corresponding instruction sequence based on the fault location result, execute link scheduling, node reconfiguration and topology update operations, and continuously iterate and train the graph neural network model.

2. The method for rapid fault location and recovery in communication networks based on self-supervised learning according to claim 1, characterized in that, The operational data includes link status information, node forwarding tables, traffic statistics, and historical alarm logs.

3. The method for rapid fault location and recovery in communication networks based on self-supervised learning according to claim 1, characterized in that, The preprocessing involves standardizing fields, aligning timestamps, filling in missing values, and removing outliers from the running data.

4. The method for rapid fault location and recovery in communication networks based on self-supervised learning according to claim 1, characterized in that, S2 specifically includes: S21. Extract each network device with a unique identifier in the communication network, define the network device as a node in the graph structure, and set a node attribute field for each node. The node attribute field includes device identifier, device type code, device operating status identifier and device network level number. S22. Extract connection paths based on the physical and logical connection relationships between network devices, treat the connection paths as edges in the graph structure, and set edge attribute fields for each edge. The edge attribute fields include starting node identifier, target node identifier, connection type code, connection availability status, and connection capacity value. S23. Match the preprocessed running data with the network device identifier according to the timestamp, and map the node-related numerical attributes to the target fields of the corresponding nodes. The target fields include processing capacity indicators, resource consumption indicators, forwarding load indicators and alarm frequency indicators. S24. Match the preprocessed running data with the connection path identifier according to the timestamp, and map the numerical attributes related to the edge to the attribute fields of the corresponding edge. The attribute fields include communication quality indicators, transmission stability indicators, traffic change magnitude and path forwarding delay. S25. The graph structure after mapping the attribute features of nodes and edges is used as the topology graph of the communication network.

5. The method for rapid fault location and recovery in communication networks based on self-supervised learning according to claim 1, characterized in that, S3 specifically includes: S31. Represent the communication network topology as a graph structure consisting of a set of nodes, a set of edges, and a node feature matrix, wherein the node feature matrix contains the multi-dimensional numerical features corresponding to each node. S32. Perform a feature masking operation, randomly select some feature elements in the node feature matrix according to a set ratio, replace these feature elements with constant zero to form a new mask feature matrix, and construct a mask view while retaining the original topology. The mask view is missing some node feature information while maintaining the same structure. S33. Perform a structural perturbation operation, delete some existing edges in the original topology according to a set ratio and insert an equal number of new edges to generate a perturbation view. The perturbation view changes some connection relationships while keeping the node characteristics unchanged. S34. The communication network topology graph and the masked view are used to form graph-level positive contrast samples, and the communication network topology graph and the perturbation view are used to form graph-level negative contrast samples for graph contrast loss calculation in graph embedding representation training. S35. For each node in the communication network topology graph, extract its embedding representation in the communication network topology graph, mask view, and perturbation view. Construct node-level positive contrast samples by combining the embeddings in the communication network topology graph and the mask view, and construct node-level negative contrast samples by combining the embeddings in the communication network topology graph and the perturbation view. These samples are used for calculating the node contrast loss in the node embedding representation training.

6. The method for rapid fault location and recovery in communication networks based on self-supervised learning according to claim 1, characterized in that, S4 specifically includes: S41. Construct a graph neural network model consisting of an encoder and a projector. Input the communication network topology, mask view and perturbation view into the encoder in sequence, and output the corresponding node embedding matrix respectively. The node embedding matrix extracts the embedding vector of each node according to the pairing relationship between node-level positive sample pairs and node-level negative sample pairs, maintaining a one-to-one correspondence. S42. Construct a projector to perform feature mapping on node embeddings to obtain node projection vectors. Use cosine similarity and temperature coefficient to calculate similarity. All calculations are based on the fixed pairing relationship between node-level positive sample pairs and node-level negative sample pairs. S43. Calculate the similarity between the projection vectors of corresponding nodes in the communication network topology diagram and the masked view, and compare it with the similarity between the projection vectors of corresponding nodes in the communication network topology diagram and the perturbation view to form a node-level contrast loss. S44. Combining graph-level positive sample pairs and graph-level negative sample pairs, the node embedding matrix is ​​used to generate a graph representation vector through average pooling. The similarity between the communication network topology graph and the masked view is calculated and compared with the similarity between the communication network topology graph and the perturbation view to form a graph-level contrast loss. S45. The node-level contrastive loss and the graph-level contrastive loss are weighted and fused to form the total loss function. The encoder parameters and projector parameters are iteratively updated by minimizing the total loss function until the training converges, thus completing the self-supervised pre-training process.

7. The method for rapid fault location and recovery in communication networks based on self-supervised learning according to claim 1, characterized in that, S5 specifically includes: S51. Apply the communication network topology graph to the graph neural network model, and use a projector to reconstruct node features and edge features to generate node feature reconstruction matrices and edge feature reconstruction matrices. S52. For each node, calculate the absolute difference between the true feature vector and the reconstructed feature vector in each dimension, and average the differences in all feature dimensions to obtain the reconstruction error value of the node features. S53. For each edge, calculate the absolute difference between the true feature vector and the reconstructed feature vector in each dimension, and average the differences in all feature dimensions to obtain the reconstruction error value of the edge feature. S54. Based on the preset node error threshold and edge error threshold, nodes with reconstruction error values ​​higher than the threshold are marked as nodes in the abnormal candidate region, and edges with reconstruction error values ​​higher than the threshold are marked as edges in the abnormal candidate region, for use in communication network fault location.

8. The method for rapid fault location and recovery in communication networks based on self-supervised learning according to claim 1, characterized in that, S6 specifically includes: S61. Based on the set of nodes and the set of edges of the abnormal candidate regions, perform connectivity analysis on the communication network topology graph, and divide nodes and edges with direct or indirect connections into the same connected component to form multiple subgraph regions. S62. Input each subgraph into the encoder in the graph neural network model to obtain the corresponding subgraph encoding representation, and perform average pooling operation on the subgraph encoding representation to obtain the subgraph embedding vector representing the overall features of the subgraph. S63. Perform similarity calculation between the subgraph embedding vector and the pre-stored historical feature centers to determine the fault category label corresponding to the maximum similarity value, and combine the fault type label with the topological location range of the subgraph to form a fault location result set containing location range and category information.

9. The method for rapid fault location and recovery in communication networks based on self-supervised learning according to claim 1, characterized in that, Specifically, S7 includes: S71. Input the fault location result set into the instruction mapping function, and match the corresponding ordered instruction sequence according to the fault category label of each sub-graph; S72. Execute each instruction in the ordered instruction sequence in sequence. If the instruction type is link scheduling, adjust the bandwidth allocation ratio and priority of the target link. If the instruction type is node reconfiguration, update the routing table, resource allocation table and processing queue threshold of the target node. If the instruction type is topology update, perform add and delete operations in the node set and edge set to form the updated communication network topology. S73. Re-input the updated communication network topology and running data into the graph neural network model, and continue to iteratively optimize the model parameters according to the self-supervised training method until the training process meets the convergence condition.

10. A method for rapid fault location and recovery in communication networks based on self-supervised learning according to claim 9, characterized in that, The instruction mapping function retrieves a preset instruction mapping table based on the fault category label of each subgraph in the fault location result set, outputs the matched ordered instruction sequence to the execution queue, and maintains a one-to-one correspondence between the category label and the corresponding instruction sequence to ensure that different fault types can trigger corresponding link scheduling, node reconfiguration, and topology update operations.

Images