Network fault prediction method
By constructing a cross-level semantic topology model and an anomaly propagation rule base, filtering key links, and using Page Ranking Algorithm to calculate fault propagation path weights, accurate prediction and optimization of network faults are achieved. This solves the problem of neglecting cross-level dependencies in existing technologies and improves the stability and resource utilization efficiency of network systems.
Patent Information
- Application Number
- CN202511231579.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-31
- Publication Date
- 2025-10-17
AI Technical Summary
Existing network fault prediction methods ignore the complex dependencies between the physical layer, protocol layer, service layer, and business layer, making it difficult to identify and predict the effects of cross-layer network faults.
By constructing a cross-level semantic topology model, filtering key links and extracting multi-dimensional features, building an anomaly propagation rule base by combining historical fault data and operation and maintenance experience, using a conflict resolution algorithm to determine the dominant factors, using Page ranking algorithm to calculate the fault propagation path weights, and using a depth-first search graph traversal algorithm for pre-simulation and corroboration verification, accurate fault prediction is achieved.
It can more accurately detect structural anomalies, identify high-risk fault propagation paths, provide precise fault handling strategies, improve system stability and reliability, avoid resource waste, and optimize the system through closed-loop learning.
Smart Images

Figure CN120811927A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of network fault engineering, in particular to a network fault prediction method. BACKGROUND
[0002] With the continuous expansion of network scale and the complexity of application scenarios, network systems have gradually evolved into a comprehensive operation system across layers, regions and protocols. In this environment, any single point failure can spread through complex dependencies, form a cascading effect, and ultimately affect the continuity and stability of large-scale services. Therefore, how to achieve accurate prediction and effective prevention before the occurrence of failure has become an important problem in the field of network operation and research.
[0003] Existing network fault prediction methods mostly rely on single-layer data monitoring and threshold judgment, relying on device logs or CPU utilization, memory occupancy hardware indicators to determine whether the device may have an abnormality. Since the existing method often only monitors and predicts a single layer, the complex dependencies between the physical layer, protocol layer, service layer and business layer are ignored, making it difficult to identify and predict the cross-layer network failure effect in network propagation. SUMMARY
[0004] In view of the defects of the prior art, the present application provides a network fault prediction method, which solves the problem of ignoring the complex dependencies between the physical layer, protocol layer, service layer and business layer.
[0005] To achieve the above purpose, the present application is realized by the following technical scheme: a network fault prediction method, comprising the following steps: Step S1: determine the layer to which the core component belongs, and construct a cross-layer semantic topology model based on the layer to which the core component belongs; Step S2: filter the link in the cross-layer semantic topology model by a node importance algorithm to obtain a key link, and perform feature extraction on the key link to obtain a structural fingerprint vector; collect a feature data vector of the current link, calculate the feature data vector of the current link and the structural fingerprint vector by a cosine similarity algorithm, and obtain a deviation degree; when the deviation degree is less than a preset threshold, it is determined that the current link is structurally abnormal; Step S3: collect historical fault data and corresponding operation and maintenance data, and construct an abnormal propagation rule library; when multiple structural abnormalities occur in the data center at the same time, the abnormal propagation rule library is used to judge the multiple structural abnormalities by a conflict resolution algorithm to obtain an abnormal dominant factor; Step S4: Collect potential signals related to the abnormal dominant factor from the monitoring points of the data center based on the abnormal dominant factor, determine the weight of the potential signals according to the abnormal propagation rule library, weight the potential signals to obtain a comprehensive potential signal, and construct a fault propagation path based on the comprehensive potential signal to obtain the weight of the fault propagation path by calculating based on the PageRank algorithm; Step S5: Based on the cross-level semantic topology model and the weight of the fault propagation path, the fault chain is preformed by the depth-first search graph traversal algorithm, and the fault prediction result is obtained by the collateral verification of the preformed fault chain through the preset inspection statistic.
[0006] Preferably, the core component belongs to the level is determined by analyzing the design drawings of the photovoltaic power station, including: First, the core components of the data center and the level to which they belong are sorted as follows: The physical layer includes servers, core switches, and storage array hardware devices. The virtualization layer includes virtual machines running on servers and virtual network logical entities. The service layer includes databases and load balancers callable services. The business layer includes applications directly used by users.
[0007] Preferably, based on the level to which the core component belongs, a cross-level semantic topology model is constructed, including: Topology graph model construction: In the data center, servers, virtual machines, and database components belong to different levels and have complex dependency relationships. These "invisible" dependencies are made explicit through structured modeling. A directed graph model is used To describe the hierarchical relationship, wherein: is a node set, and each node corresponds to a component. is an edge set, and each edge represents a dependency relationship. is a label function used to supplement semantic information: Node label ; Edge label .
[0008] Preferably, the key links in the cross-level semantic topology model are selected by a node importance algorithm, including: The node importance formula is calculated as follows: Imp(e)=bandwidth(e)× (e) where Imp(e) represents the importance of link e, and bandwidth(e) is the bandwidth of link e; (e) is the importance of the service carried by the link; Filtering the critical links, for each calculated link value, if , the link is marked as a critical link; all links marked as critical links are grouped into a set, denoted as , where represents the link.
[0009] Preferably, feature extraction is performed on the critical links to obtain a structural fingerprint vector, including: For the filtered critical link set S, a structural fingerprint is constructed by extracting multi-dimensional features, which helps to comprehensively describe the state of the critical link; Average traffic feature, the calculation formula is as follows:
[0010] where represents the average traffic feature of the link, which is a numerical value calculated by the traffic data in a period of time, used to describe the average traffic level of the link in a specific time range; is a sliding window, is the traffic value at time t; this can reflect the traffic situation of the link in a period of time; Delay fluctuation feature of the link, the calculation formula is as follows:
[0011] where represents the delay fluctuation feature of the link; this is a numerical value calculated by the delay data in a period of time, used to describe the delay fluctuation of the link in a specific time range; represents the delay value at time t, and the delay value usually refers to the time spent by data in the transmission of the link; represents the average value of the delay in the sliding window , i.e. ; the standard deviation of the delay is calculated, which reflects the stability of the link, because the smaller the delay fluctuation, the more stable the link; Redundant path number feature:
[0012] where represents the number of available backup paths; is a function for calculating the number of available backup paths of link The number of available backup paths, the more redundant paths mean that the link has more backup options when a failure occurs, and the higher the reliability; The final structure fingerprint vector is For example, the fingerprint of a certain link is Wherein, a, b, c represent the average flow characteristics, delay fluctuation characteristics, and redundant path number characteristics in normal state, respectively.
[0013] Preferably, the feature data vector of the current link is collected, the feature data vector of the current link is calculated with the structure fingerprint vector by the cosine similarity algorithm, and the deviation is obtained, and when the deviation is less than a preset threshold, it is determined that the current link is a structural anomaly. The feature data vector of the current link is collected in the key link set S, and the feature data vector of the current link is calculated with the structure fingerprint vector by the cosine similarity algorithm, and the formula is as follows:
[0014] Wherein, The feature data vector of the current link is collected; The feature data vector of the link in stable state is collected; Sim( , The cosine similarity calculated is used to measure the similarity between the current state and the stable state; after substituting the data, if (Preset threshold), it is determined as a structural anomaly.
[0015] Preferably, the historical fault data and the corresponding operation data are collected, and the anomaly propagation rule library is constructed, including: Historical fault data collection: Fault event record: extract all fault event records occurred in the past from the operation management system of the data center; Fault feature data: collect relevant feature data for each fault event; Operation data collection: collect all operation records of the operation personnel before and after the fault occurs; Construction of anomaly propagation rule library: Based on the historical fault data and the corresponding operation data, the causal relationship between the anomalies of each component in the data center is summarized; these causal relationships are stored in the form of rules, and each rule usually contains three elements: cause, effect and probability.
[0016] Preferably, the weight of the potential signal is determined according to the anomaly propagation rule library, the potential signal is weighted, and a comprehensive potential signal is obtained, including: The formula used for weighting is as follows:
[0017] wherein, is the integrated latent signal; is the weight of the th signal; is the intensity of the th signal; the summation symbol denotes summation from to wherein is the number of signals; Preferably, based on the integrated latent signal, a fault propagation path is constructed, the weight of the fault propagation path is calculated based on PageRank algorithm, including: The weight of the fault propagation path is calculated based on PageRank algorithm. The formula is as follows:
[0018] wherein, is the weight of the next round of fault propagation path, is the weight of the current fault propagation path, is the fault propagation path; is the transition matrix, which is used to describe the probability of transition from one state to another; is the damping coefficient; InDegree is the in-degree of the path , i.e. the number of edges pointing to the path P. In this way, the possibility and the degree of influence of the fault propagation on each path are evaluated.
[0019] Preferably, for each fault chain in the rehearsal, a side verification is performed by a preset inspection statistic, and a fault prediction result is obtained, including: Starting from the node in step S4 where the integrated latent signal exceeds the threshold value, traversal is performed along the edges in the topological graph and the fault propagation path with a higher weight Weight(P); The traversal terminates at two types of nodes as follows: Business layer nodes, i.e. the final components directly affecting the user; Dead-end nodes, i.e. nodes without downstream dependencies, or nodes with a path weight lower than 0.3.
[0020] The formula is expressed as:
[0021] wherein is the starting abnormal node, is the final node that may cause a business fault. The rehearsal process finds a series of fault chains that may be triggered starting from the starting abnormal node.
[0022] Then the fault chain circumstantial evidence verification is performed to obtain the fault prediction result: The basic logic of the test is as follows: Null hypothesis ,The rehearsed fault chain is not credible; Alternative hypothesis , the rehearsed failure chain is credible; Verify the goal, calculate the "test statistic" through actual detection data, and judge whether it can "overturn the "- If it can be overturned, then the path is credible; if not, then the path is not credible; Choose an appropriate test statistic, for example, for performance indicators on the path, you can use Statistics:
[0023] in, is the sample mean of the detection data, is the mean of the preview results, is the sample standard deviation, is the sample size. The statistic is calculated by actual detection data; The response time preview value of a node in the preview fault chain is ( ), the actual detection sample mean is ( ), calculate After the value, The value is input into the statistical software tool, and the output is the P value. P is the probability of observing the current sample result. The smaller the P value, the stronger the contradiction between the sample data and the original hypothesis. If the value is less than 0.05, the null hypothesis is rejected. , it is believed that if the actual response time differs significantly from the preview value, the fault chain is credible; otherwise, it is believed that the difference is not significant and the fault chain is not credible.
[0024] The present invention provides a network fault prediction method, which relates to fault prediction and health management technology and has the following beneficial effects: (1) This network fault prediction method can detect structural anomalies more accurately by building a detailed topology model, screening key links, and extracting multi-dimensional features. The abnormal propagation rule library built by combining historical data and operation and maintenance experience, as well as the use of conflict resolution algorithms to determine the main cause, helps to detect potential faults in advance and accurately determine the root cause of the fault.
[0025] (2) The network fault prediction method can identify high-risk fault propagation paths through weak signal fusion and fault path weight calculation, and then determine the reliable fault chain through pre-play and verification. Based on these results, the system is optimized, including updating the rule base and simplifying the topology structure, so that the fault handling strategy is more accurate and effective, avoiding unnecessary resource waste.
[0026] (3) The network fault prediction method forms a closed loop through the entire process, from topology modeling, through anomaly detection, propagation analysis, pre-play verification, to system optimization and disposal strategy matching, and continuously cycles and iterates. Through the feedback of fault prediction results to update the rule base and topology structure, the system can adaptively learn and optimize, improving the stability and reliability of long-term operation. BRIEF DESCRIPTION OF DRAWINGS
[0027] Figure 1 A flowchart of a network fault prediction method according to the present application.
[0028] Figure 2 A hierarchical diagram of structural anomalies obtained by a network fault prediction method according to the present application.
[0029] Figure 3 A hierarchical diagram of fault prediction results obtained by a network fault prediction method according to the present application. DETAILED DESCRIPTION
[0030] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.
[0031] Please refer to Figure 1 The present application provides a technical solution: a network fault prediction method. Specifically, the following network fault prediction method is provided, please refer to Figure 1 The method comprises the following steps: Step S1: Determine the level to which the core component belongs, and construct a cross-level semantic topology model based on the level to which the core component belongs.
[0032] First, the core components of the data center and their respective levels are as follows: The physical layer includes server (such as computing power node Server01), core switch (Switch01), and storage array (Storage01) hardware devices. The virtualization layer contains logical entities of virtual machines (VM001) running on servers, virtual networks (VLAN10); The service layer contains databases (MySQLCluster), load balancers (LB01) callable services; The business layer contains applications directly used by users (such as recommendation systems, data backup tasks).
[0033] Then the topology graph model is constructed: In the data center, servers, virtual machines, and database components belong to different levels and have complex dependency relationships (such as virtual machines depending on physical servers to run, and databases depending on storage arrays to read and write data). Through structured modeling, these "invisible" dependencies are made explicit. A directed graph model is used to describe the hierarchical relationship, where: is a set of nodes, each node corresponding to a component, for example represents a physical server; is a set of edges, each edge represents a dependency relationship, for example when , it means that the virtual machine calls the database; is a label function used to supplement semantic information: Node labels , for example , where "level = 2" corresponds to the virtualization layer; Edge labels , for example , indicating that the link has 2 backup paths.
[0034] A unique identifier is assigned to each node and edge to ensure that it can be traced back later when locating faults: The level coding is defined as: physical layer = 1, virtualization layer = 2, service layer = 3, and business layer = 4; The serial number of each component in each level starts from 001 and is arranged in order, for example, the first component of the physical layer is 1001, the second component is 1002, and so on. The first component of the virtualization layer is 2001, and the second component is 2002. The node identification rule is ; The edge identification rule is ; For example, "1001→2001@20250820" represents the connection between the physical server and the virtual machine established on August 20, 2025.
[0035] This step builds a dependency network of cross-level components in the data center through the directed graph model, labels the nodes and edges, including resource capacity of nodes, bandwidth and redundancy information of edges, and encodes the nodes by level. It provides a basic framework and data basis for subsequent key component screening. The component importance calculation and data sampling range determination in the subsequent steps depend on the topology structure and related parameters built in this step.
[0036] Step S2: Screen the links in the cross-level semantic topology model by a node importance algorithm to obtain key links, extract features of the key links to obtain a structure fingerprint vector, collect feature data vectors of the current links, calculate the feature data vectors of the current links and the structure fingerprint vector by a cosine similarity algorithm to obtain a deviation degree, and determine that the current link is structurally abnormal when the deviation degree is less than a preset threshold.
[0037] The screening of the key links is as follows: Due to the large number of components in the data center, the core objects that have a significant impact on the business (such as server links carrying AI training tasks) need to be prioritized.
[0038] The node importance formula is calculated as follows: Imp(e) = bandwidth(e) x (e) Where Imp(e) represents the importance of link e, bandwidth(e) is the link bandwidth (unit: Gbps), and (e) is the importance of the link carrying the business (0-1, core business takes 0.9-1.0). A comprehensive evaluation value calculated by the formula is used to measure the importance of the link in the entire data center architecture.
[0039] Screening key links, for each calculated link value, if , the link is marked as a key link.
[0040] For example, a link with a bandwidth of 10 Gbps carries an AI training task (importance 0.9), then Imp(e) = 10 x 0.9 = 9, which is significantly higher than the ordinary link bandwidth (usually < 3), and is marked as a key link. All links marked as key links are combined into a set, denoted as , where represents a link.
[0041] For the selected key link set S, a structure fingerprint is constructed by extracting multi-dimensional features, which helps to comprehensively describe the state of the key link.
[0042] The average flow feature is calculated by the following formula:
[0043] wherein, The average flow feature of a link is a numerical value calculated from the flow data within a certain period of time, which describes the average flow level of the link within a specific time range; is a sliding window (e.g., 1 hour), is The flow value at time t. This can reflect the flow situation of the link within a period of time.
[0044] The delay fluctuation feature of a link is calculated by the following formula:
[0045] wherein, The delay fluctuation feature of a link is a numerical value calculated from the delay data within a certain period of time, which describes the delay fluctuation of the link within a specific time range; The delay value at time t, which usually refers to the time spent by data in transmission in the link; represents the average value of the delay within the sliding window , i.e. The standard deviation of the delay is calculated, which reflects the stability of the link, as the smaller the delay fluctuation, the more stable the link.
[0046] Redundant path number feature:
[0047] wherein, represents the number of available backup paths. is a function for calculating the number of available backup paths of the link , and the larger the number of redundant paths, the more backup options the link has when a failure occurs, and the higher the reliability.
[0048] The final structure fingerprint vector is , for example, the fingerprint of a certain link is wherein, a, b, and c respectively represent the average flow feature, delay fluctuation feature, and redundant path number feature in normal state.
[0049] Calculate the normal fluctuation range of the feature through historical data: The historical mean calculation formula is:
[0050] wherein, represents the historical mean of the kth feature; k is the feature sequence index; is the number of historical samples, i.e., the number of data points used to calculate the mean value; represents the kth feature of the ith sample. is the mean value of the kth feature of the ith sample.
[0051] The standard deviation calculation formula is:
[0052] wherein, is the standard deviation value of the kth feature; is the number of samples; is the flow value of the ith sample; represents the kth feature historical mean value, and the formula is to calculate the square sum of the deviation between the sample feature value and the average feature value, and then take the square root of the average value. Based on the above historical mean and standard deviation,
[0053] the normal interval is defined as (covering 95% of normal samples). This determines a reasonable feature value fluctuation range, which is considered normal within this range.
[0054] The consistency calibration is as follows: Collect the current feature data vector of the link in the key link set S, and calculate the feature data vector of the current link and the structure fingerprint vector through the cosine similarity algorithm, as follows:
[0055] wherein, represents the feature data vector of the current link; is the structure fingerprint vector data, representing the feature data vector of the link in the stable state; Sim( , ) is the calculated cosine similarity, i.e., the deviation degree, used to measure the similarity between the current state and the stable state. After substituting the data, if (preset threshold), it is determined as a structural anomaly, which is transmitted to S3 for propagation analysis. This helps to discover abnormal changes in link state in a timely manner and take appropriate measures.
[0056] This step uses the node importance formula to screen the components of the data center to determine the key links. At the same time, multi-dimensional features are extracted from the screened key links, including the average flow feature f1, the delay fluctuation feature f2, and the number of redundant path features f3, to construct the structure fingerprint vector. The screened key components and their fingerprint features are the objects of subsequent anomaly detection, and these features will be used to compare with the normal interval to determine whether the link is abnormal, and then trigger the subsequent anomaly propagation analysis.
[0057] Step S3: Collect historical fault data and corresponding operation and maintenance data, and construct an abnormal propagation rule library; when multiple structural abnormalities occur in the data center at the same time, based on the abnormal propagation rule library, a conflict resolution algorithm is used to judge the multiple structural abnormalities, and the abnormal dominant factor is obtained.
[0058] When the link is determined to be an abnormal structure, the most likely dominant cause is mined through the following steps: Historical fault data collection: Fault event records: Extract all past fault event records from the operation and maintenance management system of the data center. These records should include the time, location (specific to the affected servers, network devices), and fault type (such as hardware failure, software failure, network failure) of the fault occurrence.
[0059] Fault feature data: For each fault event, collect relevant feature data such as device performance indicators (CPU usage, memory usage, network bandwidth utilization) before the fault occurs, system log information (including error logs, warning logs), and application running state data (such as database query delay, application response time).
[0060] Operation and maintenance data collection: Collect all operation records performed by operation and maintenance personnel before and after the fault occurs, including but not limited to device configuration changes, software updates, network topology adjustment operation times, operation contents, and operation performers.
[0061] Construction of abnormal propagation rule library: Based on historical fault data and corresponding operation and maintenance data, summarize the causal relationships between component abnormalities in the data center to form a rule set ; Store these causal relationships in the form of rules, and each rule Usually contains three elements: antecedent (event that triggers abnormality), consequence (abnormal event triggered), and probability (probability of antecedent triggering consequence).
[0062] For example, rule 1: (switch congestion, virtual machine delay, 0.8), indicating that switch congestion has an 80% probability of causing virtual machine delay; rule 2: (virtual machine overload, database timeout, 0.9), indicating that virtual machine overload has a 90% probability of causing database timeout.
[0063] Store the extracted rules in the rule library for subsequent query and use. The rule library can be implemented using a database structure to ensure efficient storage and retrieval of rules.
[0064] Collect other abnormal events that occur simultaneously with the current link abnormality from the monitoring system to form a set ; When multiple structural anomalies occur simultaneously in a data center, and these anomalies "may all lead to the same subsequent consequence, it is necessary to determine which anomaly is the primary cause.
[0065] For example, "virtual machine CPU utilization 85% (abnormal factor A)" and "core switch port packet loss rate 0.5% (abnormal factor B)" are detected simultaneously, both of which may lead to "database timeout" as the dominant factor; By using a conflict resolution algorithm to determine which of the multiple possible factors is the dominant factor, the formula is:
[0066] Among them, Score( ) represents the score of the rule , which is used to evaluate the importance of this rule in the current abnormal situation. is the rule to be evaluated. is the weight of the current observed anomaly, and the weight definition principle is that the weight of the anomaly that directly leads to business interruption or key component failure is higher; the weight of the anomaly that is strongly related to the link anomaly in time and space is second; the weight of the anomaly that occurs frequently but has no direct damage to the link is the lowest. For example, the weight of virtual machine high load ; the weight of switch packet loss is the matching degree of the rule and the current observed anomaly, and the value range is between 0 and 1, and the complete match is 1; is the current observed anomaly set. For example, in the above scenario, .
[0067] Take the above scenario as an example: for rule 1 (virtual machine high load → database timeout): (suppose the matching degree of virtual machine high load and rule 1 is 0.8, and the matching degree of switch packet loss and rule 1 is 0); For rule 2 (switch packet loss → database timeout): (suppose the matching degree of switch packet loss and rule 2 is 0.5, and the matching degree of virtual machine high load and rule 2 is 0).
[0068] The result is determined as: Compare the scores of each rule, and the factor corresponding to the rule with the highest score is determined as the primary cause of the current anomaly. In the above example, , so it is determined that "virtual machine overload" is the primary cause of the current anomaly.
[0069] The step constructs an anomaly propagation rule base, summarizes the causal relationship between each component anomaly in the data center based on historical fault data and operation experience summary data, and stores it in the form of rules, including antecedent, consequence and probability elements. When multiple anomalies occur, the conflict resolution algorithm is used to determine the dominant factor. The weight allocation basis is provided for subsequent potential signal fusion, that is, the weight of each signal is determined according to the causal relationship and probability in the anomaly propagation rule base, and the rule basis is provided for fault chain rehearsal.
[0070] Step S4: Based on the anomaly dominant factor, collect potential signals related to the anomaly dominant factor from the monitoring points of the data center, determine the weight of the potential signals according to the anomaly propagation rule base, weight the potential signals, and obtain the comprehensive potential signals; based on the comprehensive potential signals, construct a fault propagation path, and calculate the weight of the fault propagation path based on the PageRank algorithm.
[0071] In the data center, collect potential signals related to the anomaly dominant factor from each component and link. These signals can include CPU usage, memory usage, disk I / O of the server, bandwidth utilization and packet loss rate of the network device.
[0072] For example, the CPU usage of server A reaches 80% (signal S1), and the packet loss rate of network link B is 2% (signal S2).
[0073] According to the anomaly propagation rule base constructed in step S3, the weight of each signal is determined. In the anomaly propagation rule base, if a certain anomaly is more likely to cause system failure, the weight of the signal related to the anomaly is higher.
[0074] For example, in the rule base, server CPU overload is more likely to cause system problems, so the weight of the server CPU usage signal may be set to 0.7, and the weight of the network link packet loss rate signal is set to 0.3.
[0075] First, normalize multiple signals (such as CPU usage and packet loss rate) to convert them to the same dimension range, and then perform weighting using the following formula:
[0076] wherein, is the comprehensive potential signal; is the weight of the th signal; is the strength of the th signal; the summation symbol represents summation from to , wherein is the number of signals, and this formula is used to weight and combine multiple signals.
[0077] The fault propagation path weight is calculated as follows: Based on the abnormal propagation relationship determined in step S3 according to the rule base and the comprehensive potential signal, the fault propagation path is constructed. Specifically, first, the initial candidate path set of fault propagation is constructed using the comprehensive potential signal, and then the candidate path is screened and calibrated based on the abnormal propagation relationship (such as propagation direction, strength threshold, etc.) determined in step S3, and finally a complete fault propagation path that takes into account both signal characteristics and propagation rules is formed. For example, if there is a rule chain such as "server CPU overload → virtual machine delay → service response timeout" in the rule base, then a fault propagation path is formed.
[0078] The PageRank algorithm-based weight of these fault propagation paths is calculated. The formula is as follows:
[0079] wherein, is the weight of the next round of fault propagation paths, is the weight of the current fault propagation path, is the fault propagation path; is the transition matrix, which describes the probability of transitioning from one state to another; is the damping coefficient; InDegree is the in-degree of path , i.e. the number of edges pointing to path P. In this way, the possibility and impact of fault propagation on each path are evaluated.
[0080] This step collects signals related to potential anomalies from each monitoring point in the data center, determines the weight of each signal according to the abnormal propagation rule base in step S3, and performs weighted synthesis to obtain a comprehensive potential signal. At the same time, based on the abnormal propagation relationship in step S3, the fault propagation path is constructed, and the PageRank algorithm-based weight of these fault propagation paths is calculated. The fault chain rehearsal provides fault propagation path and its weight information, so that high-weight fault propagation paths can be prioritized during the rehearsal process, and the comprehensive potential signal provides a quantitative basis for subsequent anomaly judgment.
[0081] Step S5: Based on the cross-level semantic topology model and the weight of the fault propagation path, the fault chain is rehearsed by a depth-first search graph traversal algorithm. For each fault chain rehearsed, a collateral verification is performed by a pre-set verification statistic, and a fault prediction result is obtained.
[0082] The graph traversal-based rehearsal is as follows: Using the cross-level topology graph constructed in step S1 and the fault propagation path weights determined in step S4, possible fault chains are previewed through the depth-first search (DFS) graph traversal algorithm.
[0083] Input basis for rehearsal: Input 1: The cross-level semantic topology model (G = (V, E, L)) constructed in step S1. This model provides component dependencies (e.g., VM dependency on server, database dependency on storage) and is the traversable map. Input 2: The fault propagation path weights (Weight(P)) calculated in step S4. This provides path priority (high-weight paths, such as those with Weight(P) = 0.85, are traversed first, while low-weight paths, such as those with Weight(P) = 0.3, are traversed later or skipped). This prevents indiscriminate traversal and wastes resources.
[0084] Starting from the node where the comprehensive potential signal is detected to exceed the threshold in step S4 (such as a server with high CPU usage), traverse along the edges (dependencies) in the topology graph and the fault propagation paths with higher weights Weight(P).
[0085] The traversal terminates at two types of nodes as follows: Business layer nodes, i.e., the final components that directly affect users (such as "recommendation system" and "data backup task"), because the core goal of fault prediction is to "avoid business interruption"; Dead-end nodes are nodes with no downstream dependencies (such as independent backup storage arrays) or nodes with a path weight lower than 0.3 (low-risk paths do not need to be traversed further).
[0086] The formula is:
[0087] in is the starting abnormal node, This is the final node that may cause a business failure. The rehearsal process will identify a series of failure chains that may be triggered from the initial abnormal node.
[0088] Then the fault chain circumstantial evidence verification is performed to obtain the fault prediction result: The basic logic of the test is as follows: Null hypothesis ,The rehearsed fault chain is untrustworthy (i.e., the path will not actually fail); Alternative hypothesis ,The rehearsed fault chain is credible (i.e., the path will actually fail); Verify the goal, calculate the "test statistic" through actual detection data, and judge whether it can "overturn the If the hypothesis can be rejected, the path is confirmed as credible (i.e. the fault will occur along this path); if not, the path is not credible.
[0089] Select a suitable test statistic, for example, for the performance metric (such as response time, throughput) on the path, one can use statistic:
[0090] where, is the sample mean of the probe data, is the mean of the rehearsal result, is the sample standard deviation, is the sample size. The statistic is calculated from the actual probe data.
[0091] For example, the response time rehearsal value of a node in the fault chain of the rehearsal is ( ), the actual probe sample mean is ( ), after calculating the value, the value is input into the statistical software tool (SciPy library of Python), and the P value is output. P is the probability of observing the current sample result (or more extreme result). The smaller the P value, the stronger the contradiction between the sample data and the original hypothesis; if the obtained value is less than 0.05 (the usual significance level), the original hypothesis is rejected , considering that the actual response time is significantly different from the rehearsal value, the fault chain is credible, otherwise, considering that the difference is not significant, the fault chain is not credible.
[0092] Using the cross-layer topology graph constructed in step S1 and the fault propagation path weight determined in step S4, the possible fault chains are rehearsed by the depth-first search (DFS) or breadth-first search (BFS) graph traversal algorithm. For each fault chain rehearsed, set the original hypothesis and select a suitable test statistic for collateral verification. The results of fault chain rehearsal and verification will be used as feedback information for rule weight update and topology structure refinement in step S6, so as to realize the optimization of the system.
[0093] Step S6: updating the anomaly propagation rule library according to the fault prediction result to obtain an updated anomaly propagation rule library; calculating the dependency correlation between the core components and the key links based on the historical fault data and the fault prediction result, and refining the topology structure in the cross-layer semantic topology model to obtain a refined topology structure.
[0094] The rule weight update is as follows: The rule base is updated based on the results of the fault chain preview and circumstantial evidence verification in step S5. The Q-learning algorithm in reinforcement learning is used, and the Q value update formula is as follows:
[0095] in, is the Q value of rule R in the current state s (i.e. the weight or value of the rule), is the Q-value of rule R in the historical state s. A higher Q-value indicates a more reliable rule in the corresponding state and a higher probability of being selected in the future. α (learning rate) determines the step size of each update and controls the impact of new information on the Q-value. r is the reward value determined based on the consistency between the preview results and the actual verification. If the previewed fault chain is verified to be correct (prediction consistency), r>0 (positive reward); if the verification is incorrect (prediction inconsistency), r<0 (negative penalty). Function: Directly reflects the effectiveness of the rule in the current scenario and is the core signal driving Q-value updates. γ is the weight that balances immediate reward (current r) with future reward (Q-value of subsequent states). The closer γ is to 1, the more emphasis is placed on long-term benefits (considering the Q-value of subsequent states); when it is close to 0, only the immediate reward is considered. max_R'Q(R', s') selects the maximum Q-value among all possible next rules R' and next states s'. For example, if the fault chain previewed in step S5 is verified to be correct, the Q value of the rule related to the chain will increase, which means that this rule will be given more attention in future predictions, and the anomaly propagation rule base will be updated.
[0096] The topology is simplified as follows: Define the analysis object: Identify the components (such as servers, storage nodes) and links (such as network connections, data buses) to be evaluated, which are recorded as ; Extract quantitative features for each object (such as links, servers), define a set of quantifiable attributes as features, for example: links The eigenvectors of: ;in, Bandwidth utilization (%), Delay (ms), Number of failures (times / day), Number of redundant paths; For two objects and , first calculate the correlation of their corresponding features, and then synthesize the overall correlation of the object:
[0097] in, There are two objects and The correlation between is the correlation of the corresponding features; is an object No. Features, is an object No. features; v is the feature dimension (as in the above example ;; The molecule is The covariance of the features is the product of the standard deviations of the corresponding features (i.e., the correlation coefficient of a single feature); the final result is the average of all feature correlation coefficients, which measures the overall dependence strength of the two objects. , it is marked as weakly coupled and considered for simplification. For example, if a backup server is found to have low connectivity with other devices and has not played a critical role in multiple fault rehearsals, it can be considered for simplification from the core topology to reduce system complexity and computational burden.
[0098] Based on the results of the fault chain preview and circumstantial evidence verification in step S5, the rule weights in the rule base are updated using the Q-learning algorithm from reinforcement learning. Simultaneously, based on historical data and the fault prediction results from step S5, the dependencies between components and links in the data center are calculated to streamline the topology.
[0099] The updated rule base and streamlined topology will affect subsequent handling strategy matching and continuous optimization of the system, ensuring that the system can more accurately predict and respond to failures.
[0100] Step S7: Match the fault handling strategy according to the fault prediction result, the updated exception propagation rule base and the streamlined topology structure.
[0101] According to the high-confidence fault chain previewed and verified in step S5, as well as the updated rules and the streamlined topology, a corresponding handling strategy is matched.
[0102] The logic for generating a multi-dimensional matching strategy is as follows: Intelligently match the fault chain with the rule base, analyze the key nodes in the fault prediction results (such as "Server A is overloaded"), retrieve all high-trust chains containing this node in the exception propagation rule base (such as chain ID: FC-007), and extract the chain-associated disposal rule templates (such as "overload migration rule" and "current limiting and degradation rule").
[0103] Example: If database cluster node B is predicted to be overloaded, the rule base matches the chain "DB overload → query blocking → transaction timeout". The corresponding rule template is "IF node load > threshold AND storage node available THEN trigger read-write splitting policy".
[0104] Topology resource and rule condition dynamic verification as follows: Based on the simplified topology structure, real-time verification of resource constraint conditions in the rule template (such as "whether there is an available standby server" and "whether the virtual machine migration channel bandwidth is ≥200Mbps"). If the topology shows that the current load of standby server D has reached 80% (exceeding the migration load threshold), the rules of this node are automatically filtered, and the suboptimal rules (such as "start local cache expansion strategy") are preferentially matched.
[0105] Strategy priority dynamic sorting as follows: According to the fault impact range (such as core business fault priority higher than edge service) and resource scheduling cost (such as cross-machine room migration cost higher than same rack migration), the priority of the generated candidate strategy is sorted.
[0106] The output handling strategy is: containing strategy ID, handling action (such as "migrate virtual machine VM-102 to server D"), trigger condition (continuous overload time > 5 minutes), resource lock range (freeze new task access of server A).
[0107] According to the high-trust fault chain pre-acted and verified in step S5, and the updated rules and simplified topology structure in step S6, the corresponding handling strategy is matched. The timely handling of faults and the reasonable adjustment of resources are realized, and the rollback operation ensures that the system can recover to a reasonable running state after the fault is solved, guaranteeing the stability of the system and the effective use of resources.
[0108] Summary This technical solution sorts out the core components of the data center, determines its hierarchy, and constructs a cross-level semantic topology model. This model provides the infrastructure for subsequent analysis; uses the node importance algorithm to screen out key links, and extracts their features to obtain structural fingerprint vectors; collects the characteristic data vector of the current link, and compares it with the structural fingerprint vector through the cosine similarity algorithm to determine whether the current link has structural abnormalities; collects historical fault data and operation and maintenance experience to build an abnormality propagation rule base; when multiple structural abnormalities occur, uses the conflict resolution algorithm based on the rule base to determine the dominant factor of the abnormality; weak signal fusion and fault path construction; based on the dominant factor of the abnormality, collects signals related to potential abnormalities from the monitoring point, and determines the signal weight according to the abnormality propagation rule base, and obtains the weighted comprehensive potential signal; constructs a fault propagation rule base based on the comprehensive potential signal. Path, and calculate the weight of the fault propagation path through the PageRank algorithm; Fault chain preview and circumstantial verification; Based on the cross-level semantic topology model and the weight of the fault propagation path, the fault chain is previewed through the depth-first search graph traversal algorithm; For each previewed fault chain, circumstantial verification is performed through preset test statistics to obtain the fault prediction result; Update the anomaly propagation rule library according to the fault prediction result; Based on historical fault data and fault prediction results, calculate the dependency correlation between components and links, and streamline the topology structure in the cross-level semantic topology model; Match the fault handling strategy according to the fault prediction result, the updated anomaly propagation rule library and the streamlined topology structure.
[0109] It should be noted that, in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions. The sentence "including an element defined by..." does not exclude the presence of other identical elements in the process, method, article or device that includes the element."
[0110] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.
Claims
1. A network fault prediction method, characterized in that: The method comprises the following steps: Step S1: Determine the level to which the core components belong, and build a cross-level semantic topology model based on the level to which the core components belong; Step S2: Using a node importance algorithm, the links in the cross-level semantic topology model are screened to obtain key links, and features are extracted from the key links to obtain a structural fingerprint vector. The feature data vector of the current link is collected, and the feature data vector of the current link and the structural fingerprint vector are calculated using a cosine similarity algorithm to obtain a deviation. When the deviation is less than a preset threshold, the current link is determined to be structurally abnormal. Step S3: Collect historical fault data and corresponding operation and maintenance data to build an anomaly propagation rule base. When multiple structural anomalies occur simultaneously in the data center, a conflict resolution algorithm is used based on the anomaly propagation rule base to judge the multiple structural anomalies and determine the dominant factor of the anomaly. Step S4: Based on the abnormal dominant factor, potential signals related to the abnormal dominant factor are collected from monitoring points in the data center, weights of the potential signals are determined according to the abnormal propagation rule base, and the potential signals are weighted to obtain a comprehensive potential signal; based on the comprehensive potential signal, a fault propagation path is constructed, and the weight of the fault propagation path is calculated based on the Page ranking algorithm; Step S5: Based on the cross-level semantic topology model and the weight of the fault propagation path, the fault chain is previewed through a depth-first search graph traversal algorithm; for each previewed fault chain, circumstantial verification is performed using a preset test statistic to obtain a fault prediction result.
2. A network fault prediction method according to claim 1, characterized in that: By analyzing the design drawings of the photovoltaic power station, the core components are determined to which level they belong, including: First, let’s sort out the core components of the data center and their respective layers as follows: The physical layer includes servers, core switches, and storage array hardware devices; The virtualization layer includes virtual machines and virtual network logical entities running on the server; The service layer includes databases, load balancers, and callable services; The business layer contains applications that users use directly.
3. A network fault prediction method according to claim 2, characterized in that: Based on the layers to which the core components belong, a cross-layer semantic topology model is constructed, including: Construction of topology model: In a data center, servers, virtual machines, and database components belong to different layers and have complex dependencies. Through structured modeling, these "invisible" dependencies can be made explicit. Using directed graph model Describes a hierarchical relationship, where: is a set of nodes, each node corresponds to a component; is an edge set, each edge Indicates dependency relationships; is a label function used to supplement semantic information: Node Labels ; Edge Label .
4. A network fault prediction method according to claim 3, characterized in that: The node importance algorithm is used to filter the links in the cross-level semantic topology model to obtain key links, including: Calculated by the node importance formula: Imp(e)=bandwidth(e)× (e) Where Imp(e) represents the importance of link e, and bandwidth(e) is the link bandwidth; (e) is the importance of the link carrying services; Filter key links, for each calculated link Value, if , then mark the link as a critical link; all links marked as critical links are grouped into a set, recorded as ,in Indicates a link.
5. A network fault prediction method according to claim 4, characterized in that: Feature extraction is performed on the key link to obtain a structural fingerprint vector, including: For the selected critical link set S, a structural fingerprint is constructed by extracting multi-dimensional features, which helps to comprehensively describe the status of the critical links; Average flow characteristics, calculated as follows: ; in, The average traffic characteristic of a link is calculated based on traffic data over a period of time. It is used to describe the average traffic level of a link within a specific time range. is a sliding window, for Traffic value at a certain moment; this can reflect the traffic situation of the link over a period of time; The delay fluctuation characteristics of the link are calculated as follows: ; in, Represents the delay fluctuation characteristics of the link. This value is calculated by analyzing the delay data over a period of time and is used to describe the fluctuation of the link's delay within a specific time range. Represents the delay value at time t. The delay value usually refers to the time it takes for data to be transmitted in the link; Indicates that in the sliding window The average value of the internal delay, i.e. ; Calculate the standard deviation of the delay, which reflects the stability of the link, because the smaller the delay fluctuation, the more stable the link; Redundant path number characteristics: ; in, Indicates the number of available backup paths; Is a function used to calculate the link The number of available backup paths. More redundant paths means more backup options in case of link failure, which leads to higher reliability. The final structural fingerprint vector is , for example, the fingerprint of a link is , where a, b, and c represent the average traffic characteristics, delay fluctuation characteristics, and redundant path number characteristics under normal conditions, respectively.
6. A network fault prediction method according to claim 5, characterized in that: Collect the feature data vector of the current link, calculate the feature data vector of the current link and the structural fingerprint vector using the cosine similarity algorithm to obtain the deviation. When the deviation is less than a preset threshold, the current link is determined to have a structural abnormality, including: Collect the current feature data vectors of the links in the key link set S, and calculate the feature data vectors and structural fingerprint vectors of the current links using the cosine similarity algorithm. The formula is as follows: ; in, The characteristic data vector representing the current link; Represents the characteristic data vector of the link in a stable state; Sim( , ) is the calculated cosine similarity, which is used to measure the similarity between the current state and the stable state; after substituting the data, if The preset threshold is used to determine whether the structure is abnormal.
7. A network fault prediction method according to claim 6, characterized in that: Collect historical fault data and corresponding operation and maintenance data to build an exception propagation rule base, including: Historical fault data collection: Fault event records: Extract all past fault event records from the data center's operation and maintenance management system; Fault characteristic data: For each fault event, collect relevant characteristic data; Operation and maintenance data collection: Collect all operation records performed by operation and maintenance personnel before and after the failure occurs; Build an exception propagation rule base: Based on historical failure data and corresponding operation and maintenance data, the causal relationships between anomalies of various components in the data center are summarized; these causal relationships are stored in the form of rules. Each rule generally contains three elements: cause, consequence, and probability.
8. A network fault prediction method according to claim 7, characterized in that: Determining the weight of the potential signal according to the abnormal propagation rule base, and weighting the potential signal to obtain a comprehensive potential signal, including: The weighted formula is as follows: ; in, It is the comprehensive potential signal; It is The weight of the signal; It is The strength of a signal; the summation symbol Indicates from arrive Perform the summation, where is the number of signals.
9. A network fault prediction method according to claim 8, characterized in that: Based on the comprehensive potential signal, a fault propagation path is constructed, and the weight of the fault propagation path is calculated based on the Page ranking algorithm, including: The PageRank algorithm is used to calculate the weights of these fault propagation paths. The formula is as follows: ; in, is the weight of the fault propagation path in the next round, is the weight of the current fault propagation path, is the fault propagation path; is the transition matrix, which is used to describe the probability of transitioning from one state to another; is the damping coefficient; InDegree Is the path The in-degree is the number of edges pointing to the path P.
10. A network fault prediction method according to claim 9, characterized in that: For each rehearsed fault chain, circumstantial verification is performed using preset test statistics to obtain fault prediction results, including: Starting from the node where the integrated potential signal is detected to exceed the threshold in step S4, traverse along the edges in the topology graph and the fault propagation path with higher weight Weight(P); The traversal ends at two types of nodes as follows: Business layer nodes, that is, the final components that directly affect users, because the core goal of fault prediction is to "avoid business interruption"; Dead-end nodes are nodes with no downstream dependencies or nodes with path weights lower than 0.3; The formula is: ; in is the starting abnormal node, This is the final node that may cause a business failure. The rehearsal process will identify a series of failure chains that may be triggered starting from the initial abnormal node. Then the fault chain circumstantial evidence verification is performed to obtain the fault prediction result: The basic logic of the test is as follows: Null hypothesis ,The rehearsed fault chain is not credible; Alternative hypothesis , the rehearsed failure chain is credible; Verify the target, calculate the "test statistic" through actual detection data, and judge whether it can "overturn "- If it can be overturned, then the path is credible; if not, then the path is not credible; Choose an appropriate test statistic, for example, for performance indicators on the path, you can use Statistics: ; in, is the sample mean of the detection data, is the mean of the preview results, is the sample standard deviation, is the sample size, and the statistics are calculated through actual detection data; The response time preview value of a node in the preview fault chain is ( ), the actual detection sample mean is ( ), calculate After the value, The value is input into the statistical software tool, and the output is the P value. P is the probability of observing the current sample result. The smaller the P value, the stronger the contradiction between the sample data and the original hypothesis. If the value is less than 0.05, the null hypothesis is rejected. , it is believed that if the actual response time differs significantly from the preview value, the fault chain is credible; otherwise, it is believed that the difference is not significant and the fault chain is not credible.
Citation Information
Patent Citations
Fault early warning method and device of network node, medium and electronic equipment
CN114257493A
Fault root cause positioning method and device of service link, electronic equipment and medium
CN114710397A
Power distribution communication network fault path tracing method and system based on knowledge graph
CN120128467A
Cited By
Hospital security multi-source heterogeneous data fusion method and system
CN122310027A
