Bipartite graph message passing calibration network for diagnosis of uncertain topology

Abstract

This invention discloses a bipartite graph message-passing calibration network tomography diagnostic method for uncertain topologies, belonging to the field of network security. The method includes: constructing a fuzzy path-link correlation matrix and a path temporal observation matrix; performing background estimation on the path temporal observation matrix and extracting path anomaly residual vectors; extracting statistical features of the current observation scenario; performing message passing on the bipartite graph to obtain a link suspicion vector and generating a link penalty weight vector; performing posterior reweighting calibration on the fuzzy path-link correlation matrix; constructing an objective optimization function; establishing and solving a weighted sparse tomography optimization model to obtain a link anomaly state vector; and outputting the anomaly link location results and link anomaly severity results. This invention reduces the dependence on precise network topology, improves the accuracy and stability of link anomaly location and anomaly severity estimation in complex network environments, and reduces the impact of topology bias and noise interference on tomography inference results, thus possessing good practical application value.

Description

Technical Field

[0001] This invention belongs to the field of network security technology, specifically relating to a bipartite graph message passing calibration network layering diagnostic method for uncertain topologies. Background Technology

[0002] With the continuous expansion of the Internet and the increasing complexity of network structures, link status monitoring and fault location in backbone and data center networks have become increasingly important. Network layer analysis (NLT), a method that infers link status based solely on end-to-end path measurement data without requiring additional monitoring equipment within the network, has become one of the important technical means for current network operation and maintenance and fault diagnosis.

[0003] However, network tomography methods still face many challenges in real-world network environments. Traditional tomography methods typically rely on precisely known network topology, assuming that the mapping between paths and links is deterministic and invariant. But in real-world networks, due to factors such as load balancing, dynamic routing strategies, link switching, and measurement errors, path-link relationships often exhibit uncertainty or even time-varying characteristics. This leads to deviations between the constructed topology model and the real network, significantly reducing the accuracy of tomographic inference.

[0004] Because the actual network operating state has dynamic characteristics, network measurement data generally suffers from noise interference and abnormal fluctuations. For example, factors such as sudden congestion, measurement jitter, and equipment errors can cause outliers or unstable fluctuations in path observations, making traditional tomographic methods based on least squares or standard sparse reconstruction quite sensitive to abnormal data and difficult to stably recover the true state of the link in complex environments.

[0005] In actual network operation, link performance anomalies can stem not only from conventional network congestion, equipment failure, or measurement errors, but also from abnormal traffic behavior or potential network attacks. For example, distributed denial-of-service attacks or abnormal traffic injections can cause abnormally high loads on some links, resulting in a significant deviation from normal levels in path observation data. Therefore, accurate estimation of the degree of link anomalies is not only crucial for network performance diagnosis but also provides vital support for the perception and analysis of potential network security incidents.

[0006] To address the aforementioned issues, recent research has attempted to introduce machine learning and neural network methods to improve network tomography performance. However, most methods still rely on fixed topologies or utilize only shallow features, lacking the ability to explicitly model and adaptively calibrate topological uncertainties, making it difficult to achieve stable and reliable diagnostic results in dynamic and complex network environments. Neural network methods, in particular, typically rely on large amounts of high-quality labeled data for training. However, in network tomography scenarios, obtaining relevant training data is difficult and lacks representativeness, thus limiting the sufficiency of training and the actual diagnostic effectiveness of such methods.

[0007] In summary, existing network tomography diagnostic methods face at least the following two core challenges:

[0008] Firstly, how to reasonably model the network topology relationship under the condition that the path and link mapping relationship is uncertain, so as to reduce the impact of topology deviation on the diagnostic results.

[0009] Secondly, how to effectively invert the link status by combining path observation anomaly information when there are deviations in the topology relationship, so as to achieve accurate location of abnormal links and reasonable estimation of the degree of anomaly. Summary of the Invention

[0010] To address the shortcomings of existing network tomography diagnostic methods, such as insufficient diagnostic accuracy under uncertain topological conditions and difficulty in accurately locating abnormal links and effectively estimating the degree of anomaly under topological deviations, this invention provides a bipartite graph message-passing calibration network tomography diagnostic method for uncertain topologies. This invention constructs a fuzzy path-link correlation matrix, combines path anomaly residual vector extraction, bipartite graph message passing, posterior reweighting calibration, and weighted sparse tomography optimization model inversion, to locate abnormal links and estimate their degree of anomaly, thereby improving the accuracy and stability of network tomography diagnostics in complex network environments.

[0011] The technical solution adopted by this invention to solve the technical problem is as follows:

[0012] This invention provides a method for bipartite graph message passing calibration network tomography diagnosis for uncertain topologies, which mainly includes the following steps:

[0013] Step S1: Construct the fuzzy path-link correlation matrix and path time series observation matrix;

[0014] Step S2: Perform background estimation on the path time series observation matrix and extract the path anomaly residual vector; use the fuzzy path-link correlation matrix and the path anomaly residual vector to extract the statistical features of the current observation scene, which are used to characterize the residual fluctuation level, noise level and anomaly intensity of the current observation scene;

[0015] Step S3: Construct a bipartite graph consisting of path nodes and link nodes, perform message passing on the bipartite graph to obtain the link suspicion vector and generate the link penalty weight vector, and perform posterior reweighting calibration on the fuzzy path-link association matrix.

[0016] Step S4: Construct the objective optimization function based on the path anomaly residual vector, link penalty weight, and calibrated fuzzy path-link correlation matrix; establish and solve the weighted sparse tomography optimization model to obtain the link anomaly state vector; output the abnormal link location result and the link anomaly degree result based on the link anomaly state vector.

[0017] Furthermore, in step S1, the topology information and source-destination pair information of the target network are first obtained, then the candidate transmission path set corresponding to each source-destination pair is determined, a fuzzy path-link correlation matrix is ​​constructed based on the candidate transmission path set, and path-level end-to-end measurement data is collected to form a path time series observation matrix.

[0018] Furthermore, in step S1, the mathematical expression for the fuzzy path-link association matrix is: in, The total number of links. The total number of paths, Indicates the first The path and the first The fuzzy association weight or expected occupancy strength between links is used to characterize the degree of association between the path and the link under routing uncertainty.

[0019] Furthermore, in step S1, the mathematical expression for the path time series observation matrix is: in, The total number of paths, The length of the observation period. Indicates the first The path at time The end-to-end delay measurement value, express OK The space of real matrices in the column;

[0020] path time series observation matrix This can be represented as the superposition of the background component matrix and the anomalous perturbation component matrix as follows: ;in, This represents the actual path background delay component under normal network conditions. This refers to the abnormal disturbance component formed at the path layer after the link anomaly, congestion, or failure factors propagate through the path.

[0021] Furthermore, in step S2, the path time series observation matrix is... A path-by-path robust temporal filtering process is performed along the time dimension to obtain the background estimation matrix. It is used to characterize the normal background behavior estimated from the observed data; the residual matrix is ​​calculated. This is used to represent the degree of deviation of the time series observation matrix of each path from the normal background behavior; based on the residual matrix... Extracting path anomaly residual vectors , For the time to be diagnosed, The total number of paths is given; statistical features reflecting the current observation scenario are extracted based on the path anomaly residual vector and the fuzzy path-link correlation matrix.

[0022] Furthermore, in step S2, the statistical characteristics include: residual interquartile range and maximum outlier peak value; based on the statistical characteristics, the parameter configuration of the weighted sparse tomography solver is adaptively determined according to preset rules: the maximum outlier peak value is... With preset peak threshold Compare the residual interquartile ranges. With preset noise threshold When comparing, When using weak sparsity penalty and wide loss boundary parameter configuration; when In the case of strong sparsity penalty and narrow loss boundary, parameter configuration is adopted; otherwise, balanced parameter configuration is adopted.

[0023] Furthermore, in step S3, the mathematical expression for the link suspicion vector is: ;

[0024] in, This is a non-negative constraint mapping used to suppress the negative reward of negative residuals on link suspicion, so that link suspicion can be used to characterize the degree of suspicion of each link for the current path anomaly: This is a fuzzy path-link correlation matrix; This represents the path anomaly residual vector; Total number of links; symbol express A non-negative real vector space.

[0025] Furthermore, in step S3, the link suspicion vector is bounded and normalized to generate a link penalty weight vector; based on the link suspicion, the original non-zero association items in the fuzzy path-link association matrix are reweighted and calibrated posteriorly, and the posterior matrix is ​​obtained under the condition that the overall association quality of each path is basically stable; the posterior matrix is ​​fused and updated with the original fuzzy path-link association matrix according to a preset mixing ratio to obtain the calibrated fuzzy path-link association matrix.

[0026] Furthermore, in step S4, the torsional inversion relationship between the path anomaly residual vector and the link anomaly state vector is established as follows: ; The calibrated fuzzy path-link correlation matrix, This is the abnormal state vector of the link to be restored.

[0027] Furthermore, in step S4, the objective optimization function combines the Huber loss term and the weighted sparse regularization term to construct a weighted sparse tomographic optimization model, the mathematical expression of which is as follows:

[0028]

[0029]

[0030] in, This represents the Huber loss function. The regularization coefficient is . The loss threshold for Huber. For the first The penalty weight corresponding to each link, The abnormal state vector of the link to be restored The first in The component is used to characterize the first component. The degree of abnormality of a link deviating from its normal state. The closer a value is to 0, the closer the link is to a normal state. For path anomaly residual vector The Middle One portion, The calibrated fuzzy path-link correlation matrix, This is the abnormal state vector of the link to be restored.

[0031] The beneficial effects of this invention are:

[0032] This invention provides a bipartite graph message-passing calibration network tomography diagnostic method for uncertain topologies. By constructing a bipartite graph message-passing calibration mechanism oriented towards path nodes and link nodes, it can dynamically correct fuzzy path-link associations based on path anomaly information. Furthermore, it combines this with a weighted sparse tomography optimization model inversion solution to estimate link anomaly states. Compared with existing technologies, this invention reduces the dependence on precise network topology, improves the accuracy and stability of link anomaly location and anomaly severity estimation in complex network environments, and reduces the impact of topology bias and noise interference on tomography inference results, thus possessing significant practical application value. Attached Figure Description

[0033] Figure 1The flowchart of a bipartite graph message passing calibration network tomography diagnostic method for uncertain topology provided by the present invention.

[0034] Figure 2 A schematic diagram of constructing a fuzzy path-link correlation matrix for a single source-destination pair.

[0035] Figure 3 This is a schematic diagram of path-link fuzzy associations calibrated by bipartite graph message passing.

[0036] Figure 4 This invention provides an architecture diagram of a bipartite graph message passing calibration network tomography diagnostic system for uncertain topologies.

[0037] Figure 5 The performance of three methods (RTND, CLINK, MAP) under topological uncertainty is shown. Figure 5 The top-middle figure shows a comparison of the accuracy of the three methods (RTND, CLINK, and MAP) under topological uncertainty, while the bottom figure shows a comparison of the F1 scores of the three methods (RTND, CLINK, and MAP) under topological uncertainty.

[0038] Figure 6 The results show a comparison of the diagnostic accuracy of three methods (RTND, CLINK, and MAP) in scenarios where the complexity of multi-link anomalies gradually increases. Detailed Implementation

[0039] The present invention will be further described in detail below with reference to the accompanying drawings.

[0040] In a first aspect, the present invention provides a bipartite graph message passing calibration network tomography diagnostic method for uncertain topologies.

[0041] This invention provides a bipartite graph message-passing calibration network tomography diagnostic method for uncertain topologies, such as... Figure 1 As shown, the specific implementation process is as follows:

[0042] Step S1: Obtain the topology and source-destination pair information of the target network, determine the candidate transmission path set corresponding to each source-destination pair, construct a fuzzy path-link correlation matrix based on the candidate transmission path set, and collect path-level end-to-end measurement data to form a path time series observation matrix; the specific implementation process is as follows:

[0043] S101: Obtain the topology and source-destination pair information of the target network to determine the set of network nodes. Link set and the set of candidate transmission paths corresponding to each source-destination pair; determine the set of candidate paths based on the actual measured or logical paths corresponding to the source-destination pairs, and set the total number of links to be [value missing]. The total number of paths is The observation time length is .

[0044] S102: For scenarios involving load balancing, path switching, or routing uncertainty in real-world networks, a fuzzy path-link association matrix is ​​constructed by weighted fusion of multiple candidate paths for each source-destination pair. :

[0045]

[0046] in, Indicates the first The path and the first The fuzzy association weight or expected occupancy strength between links is used to characterize the degree of association between the path and the link under routing uncertainty.

[0047] The process of constructing the fuzzy path-link correlation matrix of a single source-destination pair is as follows: Figure 2 As shown, with the source node to the destination node For example, first identify There are 13 candidate paths, each with a different weight: Path1 has a weight of 0.6, Path2 has a weight of 0.3, Path3 has a weight of 0.1, and Link1 through Link13 are different links. Then, the association between paths and links is constructed: for each link in the network, if the link is exclusively used by only one path, the association value of the link is equal to the weight of that path; if the link is shared by multiple paths (e.g., Link4 is traversed by Path1, Path2, and Path3 simultaneously), the association value of the link is the sum of the weights of all paths traversing that link. Based on this, the contributions of multiple candidate paths to the links can be weighted and fused to obtain the fuzzy path-link association between source and destination pairs under uncertain topology conditions.

[0048] Specifically, fuzzy path-link association matrix The construction method is as follows:

[0049] For each source-destination pair, multiple candidate transmission paths are obtained, and the link contributions are weighted and fused according to the path weights corresponding to each candidate transmission path, so that the matrix elements are... Used to characterize the The path under routing uncertainty is the first The degree of correlation between the links.

[0050] S103: Collect end-to-end time delay measurements for each path at multiple consecutive time points to form a path time series observation matrix:

[0051]

[0052] in, Represents the path time series observation matrix; Indicates the first The path at time The end-to-end delay measurement; express OK The space of real matrix columns.

[0053] In fact, the path time series observation matrix Formed by the superposition of the background component matrix and the anomalous disturbance component matrix, it can be expressed in the following form:

[0054] ;

[0055] in, This represents the actual path background delay component under normal network conditions. This represents the abnormal disturbance component formed at the path layer after propagation through the path due to link anomalies, congestion, or failures. It is used when there are no link anomalies in the network. It is a zero matrix.

[0056] Step S2: Based on a robust temporal filtering method, background estimation is performed on the path temporal observation matrix, and path anomaly residual vectors are extracted. Statistical features of the current observation scene are extracted based on the path anomaly residual vectors and the fuzzy path-link correlation matrix, and the parameters of the weighted sparse tomography solver are adaptively determined. The specific implementation process is as follows:

[0057] S201: Path time series observation matrix A path-by-path robust temporal filtering process is performed along the time dimension to obtain the background estimation matrix. ,in, Used to characterize normal background behavior estimated from observation data. Indicates the first The path at time The normal background delay estimate.

[0058] Specifically, a sliding window mid-range filter is preferred for background estimation of the time delay sequence of each path to suppress the impact of isolated spikes, short-term jitter, and sudden anomalies on the background estimation. The filter window length is assumed to be an odd number. , `x` is a non-negative integer used to limit the neighborhood range of the sliding window to the left and right of the current time step. The background estimate is determined by the median of the observations within the neighborhood at the current time step. The path at time The background estimate is:

[0059] ;

[0060] in, Indicates the first The path at time The end-to-end delay observations, For the time index within the sliding window, Indicates the current time A local time window centered on and limited to the effective observation time range.

[0061] S202: Based on the path time series observation matrix With background estimation matrix The difference is used to calculate the residual matrix. :

[0062] ;

[0063] Wherein, residual matrix Used to represent the degree of deviation of each path's time series observation matrix from normal background behavior.

[0064] S203: At the time of diagnosis Extracting path anomaly residual vectors Combined with path anomaly residual vector and fuzzy path-link correlation matrix This involves extracting statistical features that reflect the current observation scenario. These statistical features primarily include: residual interquartile range. and maximum abnormal peak value ,in, Represents the path anomaly residual vector The 75th quantile of each component, Represents the path anomaly residual vector The 25th quantile of each component, Indicates the first The path at the time of diagnosis The abnormal residual values. Statistical features are mainly used to characterize the residual fluctuation level, noise level, and anomaly intensity of the current observation scene.

[0065] S204: Based on statistical characteristics, the parameter configuration of the weighted sparse tomography solver is adaptively determined according to preset rules. The parameters mainly include regularization coefficients. and the threshold parameter of the loss function .

[0066] Specifically, the maximum abnormal peak value With preset peak threshold Compare the residual interquartile ranges. With preset noise threshold When comparing, When using weak sparsity penalty and wide loss boundary parameter configuration; when In the case of strong sparsity penalty and narrow loss boundary, a parameter configuration is adopted; in other cases, a balanced parameter configuration is adopted to preserve the abnormal signal when the sudden anomaly is strong and to enhance the ability to suppress noise disturbance when the noise is strong.

[0067] S205: The parameter configuration of the determined weighted sparse tomography solver is used as the solution condition for subsequent link abnormal state recovery. The weighted sparse tomography solver is provided with constraint strength and fitting boundary adapted to the current observation scenario, so that it can be called when constructing the objective optimization function containing Huber loss term and weighted sparse regularization term.

[0068] Step S3: Message passing is performed on the bipartite graph formed by path nodes and link nodes to obtain link suspicion levels. Link penalty weights are then generated based on these suspicion levels, and the fuzzy path-link association matrix is ​​reweighted and calibrated posteriorly. The specific implementation process is as follows:

[0069] S301: Using path sets and link sets as two types of heterogeneous nodes, construct a bipartite graph composed of path nodes and link nodes, and use a fuzzy path-link association matrix. As the connection weights between path nodes and link nodes in a bipartite graph, the bipartite graph is used to explain how path anomalies are mapped to the link layer via topological relationships.

[0070] S302: The link suspicion vector is obtained by aggregating messages from path nodes to link nodes. Specifically, it is obtained by aggregating the path anomaly residual vector. As input features for path nodes, message aggregation from path nodes to link nodes is performed on the bipartite graph. The path anomaly residual vector is propagated to the link side according to fuzzy association weights and accumulated with weights to obtain the link suspicion vector corresponding to each link:

[0071] ;

[0072] Among them, symbols express A non-negative real vector space The total number of links. This is a non-negativity constraint mapping used to truncate or map negative components in the input vector to non-negativity values. Preferably, it can be taken as... This mapping is used to suppress the negative reward of negative residuals on link suspicion, so that link suspicion is used to characterize the degree of suspicion of each link for the current path anomaly: the higher the link suspicion, the more likely the link is to be the potential cause of the path delay anomaly.

[0073] S303: Based on the link suspicion vector The link penalty weight vector is generated through bounded mapping (preferably exponential decay mapping) and normalization. This results in links with higher suspicion levels being subject to weaker sparsity penalties in subsequent solutions, while links with lower suspicion levels are subject to relatively stronger sparsity penalties, thereby enhancing the sensitivity of the link abnormal state recovery results to key links.

[0074] S304: Analysis of Fuzzy Path-Link Association Matrix Based on Link Suspicion By performing posterior reweighting on the non-zero correlation terms in the original data, and maintaining the overall correlation quality of each path relatively stable, the posterior matrix is ​​obtained. .

[0075] S305: Transform the posterior matrix Compared with the original fuzzy path-link association matrix The fuzzy path-link correlation matrix is ​​obtained by fusion and updating according to the preset mixing ratio. ,in, For preset mixing ratio parameters, preferably, the preset mixing ratio is... satisfy Path-link association matrix This is used to dynamically correct the association strength between paths and links while preserving the overall structure of the original fuzzy association relationship, and to improve the accuracy of subsequent link state inversion.

[0076] like Figure 3 As shown, a bipartite graph is constructed using path sets and link sets as two types of heterogeneous nodes, consisting of path nodes (P1, P2, P3) and link nodes (L1, L2, L3, L4, L5). The left subgraph represents the fuzzy path-link association matrix. The initial bipartite graph structure is described, where the edges between path nodes and link nodes represent non-zero fuzzy associations between the corresponding paths and links, and the edge weights characterize the association strength. The intermediate subgraph represents the message aggregation process, which incorporates the path anomaly residual vectors. The features are used as input features for path nodes and propagated from path nodes to link nodes according to fuzzy association weights to obtain the link suspicion vector. The higher the link suspicion level, the more likely that link is a potential cause of path delay anomalies. The right subgraph represents the posterior reweighted calibration result, based on the link suspicion vector. Generate link penalty weight vector And for the fuzzy path-link correlation matrix The non-zero correlation terms in the matrix are reweighted a posteriori to obtain the calibrated fuzzy path-link correlation matrix. The calibration process, while preserving the original path-link association structure, adjusts the association strength based on the current path anomaly propagation information to improve the accuracy of subsequent link anomaly state inversion.

[0077] Step S4: Based on the calibrated fuzzy path-link correlation matrix, path anomaly residual vector, and link penalty weights, construct and solve the weighted sparse tomography optimization model to obtain the link anomaly state vector. Then, output the anomaly link location result and the link anomaly degree result based on the link anomaly state vector. The specific implementation process is as follows:

[0078] S401: Establish the tomographic inversion relationship between the path anomaly residual vector and the link anomaly state vector. ,in, This is the abnormal state vector of the link to be restored.

[0079] S402: The weighted sparse tomography optimization model is constructed from the objective function that combines the Huber loss term and the weighted sparse regularization term as follows:

[0080]

[0081]

[0082] in, This represents the Huber loss function. The regularization coefficient is . The loss threshold for Huber. For the first The penalty weight corresponding to each link, The abnormal state vector of the link to be restored The first in Each component. Non-negativity constraints are used to ensure that the estimated value of the link's abnormal state vector is consistent with the physical meaning of the link's deviation from the normal level.

[0083] The link anomaly state estimation vector is obtained by solving the objective optimization function. ,in Used to characterize the The estimated degree of abnormality of each link deviating from the normal level.

[0084] S403: Based on the link anomaly state estimation vector Output the anomaly level results for each link. When the estimated anomaly state of a link exceeds a preset threshold... When this happens, the link is determined to be an abnormal link, and the abnormal link location result is output. The preset threshold used for this determination... It can be set based on experience presets, validation set optimization results, or specific application scenario requirements.

[0085] Secondly, the present invention provides a bipartite graph message passing calibration network tomography diagnostic system for uncertain topologies.

[0086] like Figure 4 As shown, this invention provides a bipartite graph message-passing calibration network tomography diagnostic system for uncertain topologies, specifically comprising the following modules: a data acquisition and preprocessing module, a filtering and feature extraction module, a graph network calibration module, and an anomaly diagnosis module. The data acquisition and preprocessing module is used for topology and observation construction; the filtering and feature extraction module is used for background estimation and residual extraction; the graph network calibration module is used for bipartite graph message-passing calibration; and the anomaly diagnosis module is used for weighted sparse tomography solution. These modules are sequentially connected to output anomaly link location results and anomaly severity estimation results.

[0087] The data acquisition and preprocessing module mainly includes a path time series builder and a topology resolver, used to acquire the topology information, source-destination pair information, and end-to-end measurement data of the target network, determine the candidate transmission path set, and construct a fuzzy path-link correlation matrix. And collect path-level end-to-end measurement data to form a path time series observation matrix. ;

[0088] The filtering and feature extraction module mainly includes a sliding window value filter, a residual extractor, and a feature extractor, used to process the path time series observation matrix. Background estimation is performed to obtain the background estimation matrix. And extract the path anomaly residual vector. And the statistical characteristics of the current observation scenario;

[0089] The graph network calibration module mainly includes a bipartite graph builder and a message passing network, used to construct a bipartite graph composed of path nodes and link nodes, based on path anomaly residual vectors. Execute message passing to obtain the link suspicion vector. Generate link penalty weight vector It outputs the calibrated fuzzy path-link correlation matrix. ;

[0090] The anomaly diagnosis module mainly includes a weighted sparse solver and an anomaly detection unit, used to determine the anomaly based on the calibrated fuzzy path-link correlation matrix. Path anomaly residual vector and link penalty weight vector Performing a weighted sparse tomography solution yields the link anomaly state estimation vector. It outputs the abnormal link location results and the abnormality degree results.

[0091] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0092] To further illustrate the bipartite graph message passing calibration network tomography diagnostic method for uncertain topologies provided by this invention, this embodiment uses a backbone network scenario as the diagnostic object and selects the network structure abstracted from the ChinaNet topology as the implementation carrier: First, the network topology file corresponding to ChinaNet is read, and the node identifiers are normalized to determine the set of network nodes. Link set And several source-destination pairs; then, based on the source-destination pairs, a set of candidate transmission paths and their corresponding path-level observation scenarios are constructed. Then, using the bipartite graph message passing calibration network tomography diagnostic method for uncertain topology provided by the present invention, multiple consecutive steps are performed, including fuzzy path-link correlation matrix construction, path time series observation matrix formation, background estimation and residual extraction, statistical feature generation, bipartite graph message passing calibration, and weighted sparse tomography solution.

[0093] To verify the effectiveness of the Robust Topology-uncertainty-aware Network Diagnosis (RTND) method for bipartite graph message passing calibration provided in this embodiment, a comparative experiment was designed and compared with existing CLINK and MAP methods. The experiment mainly evaluated the performance of the method from three aspects: topology uncertainty, multi-link failure rate, and observation noise intensity.

[0094] The performance of the three methods (RTND, CLINK, MAP) under topological uncertainty is as follows: Figure 5 As shown, Figure 5The diagram contains two sub-figures, one above the other, illustrating the trends in accuracy and the F1 score (overall evaluation metric) of the three methods (RTND, CLINK, and MAP) as the topological uncertainty gradually increases from 0.0 to 0.6. The accuracy comparison shows that the CLINK method generally exhibits a low accuracy level; while the MAP method maintains an accuracy above 0.97, the method of this invention (RTND) consistently maintains the best accuracy across the entire uncertainty range. This indicates that even under drastic changes in network topology, the method of this invention (RTND) can still maintain an extremely high correct identification rate. The F1 score, as a comprehensive indicator of precision and recall, better reflects the true performance under imbalanced samples such as multi-link anomalies. Figure 5 The figure below shows that the F1 scores of the CLINK and MAP methods consistently hover around 0.55, indicating poor performance. In contrast, even under extreme disturbance scenarios with topological uncertainty as high as 0.6, the F1 score of the method of this invention (RTND) only shows a slow decrease and remains above 0.85, forming a significant lead over traditional methods.

[0095] The diagnostic accuracy of the three methods (RTND, CLINK, MAP) in scenarios with gradually increasing complexity of multi-link anomalies is compared as follows: Figure 6 As shown, Figure 6 The horizontal axis represents the proportion of links experiencing anomalies in the network, and the vertical axis represents the accuracy of anomaly link localization. As the proportion of links experiencing simultaneous anomalies in the network increases from approximately 2.5% to 20.0%, the coupling degree of path observation data caused by multiple link anomalies increases sharply, transforming the network tomography problem from a simple single anomaly identification to a highly challenging multivariate joint inversion problem. As a result, the accuracy of the three methods (RTND, CLINK, and MAP) all show varying degrees of decline. Throughout the entire anomaly proportion variation range, the accuracy curve of the method of this invention (RTND) consistently remains at the top, demonstrating optimal diagnostic performance. At a low anomaly proportion (2.5%), the accuracy of the method of this invention (RTND) approaches 0.98; even in extremely complex coupling scenarios with an anomaly proportion as high as 20.0%, the accuracy of the method of this invention (RTND) can still be maintained above 0.80.

[0096] The following table shows the comparison of the link anomaly estimation errors of the three methods (RTND, CLINK, MAP) under different observation noise levels (Noise Std).

[0097]

[0098] The data in the table above uses the mean. The error is presented in the form of "standard deviation," with a smaller value indicating a smaller deviation between the estimated result and the actual state, and higher diagnostic accuracy. As the noise standard deviation in the network measurement data gradually increases from 0.1 to 2.0, the errors of all three methods (RTND, CLINK, and MAP) inevitably show an upward trend, reflecting the interference of noise such as sudden congestion and measurement jitter on network tomography inference. At low noise levels, the proposed method (RTND) and CLINK method both exhibit extremely low estimation errors, significantly outperforming the MAP method. With a significant increase in noise intensity, the performance of the traditional CLINK and MAP methods begins to decline significantly, especially under strong noise interference with a Noise Std of 2.0, where the average error of the CLINK method surges to 0.2566, and the error of the MAP method reaches 0.2346. In contrast, the proposed method (RTND) maintains the lowest average error at all high-noise nodes (only 0.2025 at a Noise Std of 2.0), and its error growth is the most gradual.

[0099] In summary, the method of the present invention (RTND) can achieve relatively effective link anomaly localization and anomaly degree estimation under conditions of uncertain topology, noise interference, and multiple link anomalies.

[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. However, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for tomographic diagnosis of bipartite graph message-passing calibration networks with uncertain topologies, characterized in that, Includes the following steps: Step S1: Construct the fuzzy path-link correlation matrix and path time series observation matrix; Step S2: Perform background estimation on the path time series observation matrix and extract the path anomaly residual vector; The statistical features of the current observation scene are extracted using the fuzzy path-link correlation matrix and path anomaly residual vector, which are used to characterize the residual fluctuation level, noise level and anomaly intensity of the current observation scene; Step S3: Construct a bipartite graph consisting of path nodes and link nodes, perform message passing on the bipartite graph to obtain the link suspicion vector and generate the link penalty weight vector, and perform posterior reweighting calibration on the fuzzy path-link association matrix. Step S4: Construct the objective optimization function based on the path anomaly residual vector, link penalty weight, and calibrated fuzzy path-link correlation matrix; establish and solve the weighted sparse tomography optimization model to obtain the link anomaly state vector; output the abnormal link location result and the link anomaly degree result based on the link anomaly state vector.

2. The bipartite graph message passing calibration network tomography diagnostic method for uncertain topology as described in claim 1, characterized in that, In step S1, the topology information and source-destination pair information of the target network are first obtained, then the candidate transmission path set corresponding to each source-destination pair is determined, a fuzzy path-link correlation matrix is ​​constructed based on the candidate transmission path set, and path-level end-to-end measurement data is collected to form a path time series observation matrix.

3. The bipartite graph message passing calibration network tomography diagnostic method for uncertain topologies according to claim 1, characterized in that, In step S1, the mathematical expression for the fuzzy path-link association matrix is: in, The total number of links. The total number of paths, Indicates the first The path and the first The fuzzy association weight or expected occupancy strength between links is used to characterize the degree of association between the path and the link under routing uncertainty.

4. The bipartite graph message passing calibration network tomography diagnostic method for uncertain topologies according to claim 1, characterized in that, In step S1, the mathematical expression for the path time series observation matrix is: in, The total number of paths, The length of the observation period. Indicates the first The path at time The end-to-end delay measurement value, express OK The space of real matrices in the column; path time series observation matrix This can be represented as the superposition of the background component matrix and the anomalous perturbation component matrix as follows: ;in, This represents the actual path background delay component under normal network conditions. This refers to the abnormal disturbance component formed at the path layer after the link anomaly, congestion, or failure factors propagate through the path.

5. The bipartite graph message passing calibration network tomography diagnostic method for uncertain topology according to claim 1, characterized in that, In step S2, the path time series observation matrix is... A path-by-path robust temporal filtering process is performed along the time dimension to obtain the background estimation matrix. , used to characterize normal background behavior estimated from observation data; Calculate the residual matrix This is used to represent the degree of deviation of the time series observation matrix of each path from the normal background behavior; based on the residual matrix... Extracting path anomaly residual vectors , For the time to be diagnosed, The total number of paths is given; statistical features reflecting the current observation scenario are extracted based on the path anomaly residual vector and the fuzzy path-link correlation matrix.

6. The bipartite graph message passing calibration network tomography diagnostic method for uncertain topology according to claim 1, characterized in that, In step S2, the statistical features include: residual interquartile range and maximum outlier peak value; based on the statistical features, the parameter configuration of the weighted sparse tomography solver is adaptively determined according to preset rules: the maximum outlier peak value is... With preset peak threshold Compare the residual interquartile ranges. With preset noise threshold When comparing, When using weak sparsity penalty and wide loss boundary parameter configuration; when In the case of strong sparsity penalty and narrow loss boundary, parameter configuration is adopted; otherwise, balanced parameter configuration is adopted.

7. The bipartite graph message passing calibration network tomography diagnostic method for uncertain topology according to claim 1, characterized in that, In step S3, the mathematical expression for the link suspicion vector is: ; in, This is a non-negative constraint mapping used to suppress the negative reward of negative residuals on link suspicion, so that link suspicion can be used to characterize the degree of suspicion of each link for the current path anomaly: This is a fuzzy path-link correlation matrix; This represents the path anomaly residual vector. Total number of links; symbol express A non-negative real vector space.

8. The bipartite graph message passing calibration network tomography diagnostic method for uncertain topologies according to claim 1, characterized in that, In step S3, the link suspicion vector is bounded and normalized to generate the link penalty weight vector; based on the link suspicion, the original non-zero association items in the fuzzy path-link association matrix are reweighted and calibrated posteriorly to obtain the posterior matrix while maintaining the overall association quality of each path as basically stable. The posterior matrix and the original fuzzy path-link correlation matrix are fused and updated according to a preset mixing ratio to obtain the calibrated fuzzy path-link correlation matrix.

9. The bipartite graph message-passing calibration network tomography diagnostic method for uncertain topologies according to claim 1, characterized in that, In step S4, the tomographic inversion relationship between the path anomaly residual vector and the link anomaly state vector is established as follows: ; The calibrated fuzzy path-link correlation matrix, This is the abnormal state vector of the link to be restored.

10. The bipartite graph message passing calibration network tomography diagnostic method for uncertain topology according to claim 1, characterized in that, In step S4, the objective optimization function combines the Huber loss term and the weighted sparse regularization term to construct a weighted sparse tomographic optimization model, the mathematical expression of which is as follows: ， ， in, This represents the Huber loss function. The regularization coefficient is . The Huber loss threshold For the first The penalty weight corresponding to each link, The abnormal state vector of the link to be restored The first in One portion, Path anomaly residual vector The Middle One portion, The calibrated fuzzy path-link correlation matrix, This is the abnormal state vector of the link to be restored.

Images