A fault prediction method and system based on observable data

Generate business topology graphs through graph convolution networks, extract node features, identify fault nodes and fault types, and predict fault propagation paths, solving the problem of insufficient efficiency and accuracy of large-scale dynamic data in the existing technology, and achieving rapid fault response and efficient recovery.

CN118713991BActive Publication Date: 2024-11-05BEIJING ZHONGDIAN HUITONG TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411180471.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-27
Publication Date
2024-11-05
Estimated Expiration
2044-08-27

AI Technical Summary

Technical Problem

Existing fault diagnosis technology cannot efficiently process large-scale and dynamically changing data, especially in complex system dependencies, resulting in slow processing speed and insufficient diagnostic accuracy, and inability to respond to system failures in a timely manner, increasing system instability and operational risks. Moreover, the delay update and result analysis of fault data are not real-time, which limits the failure recovery efficiency.

Method used

Using a graph-based convolutional network method, the generation of business topology graphs, extract node characteristics, update node status, identify fault nodes and fault types, predict fault propagation paths and potential impacts, and combine deep learning algorithms to analyze the cause of faults to achieve rapid fault location and real-time monitoring.

Benefits of technology

Real-time monitoring of complex structures is realized, abnormal deviations are quickly identified, and fault nodes and their types are accurately positioned, which improves fault response speed, improves fault recovery efficiency and business continuity.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118713991B_ABST
    Figure CN118713991B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of fault diagnosis technology, specifically to a fault prediction method and system based on observable data, comprising the following steps: selecting and collecting real-time load data and topological structure data of a business platform, mapping service components to nodes of a graph, mapping dependencies to edges of the graph, generating a business topological graph, and extracting node features from the business topological graph using a graph convolutional network to obtain a node feature matrix. In the present invention, a business topological graph generated by integrating real-time load data and topological structure data is used, and a graph convolutional network is used to improve the ability to extract node features from complex structures. This method, which goes deep into the core of the network structure, can effectively monitor node status in real time, accurately identify abnormal deviations, and quickly locate faulty nodes and their types. By analyzing the equipment directly connected to the faulty node, it is expanded to the prediction of the propagation path and potential impact of the fault, thereby improving the speed of fault response.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a fault prediction method and system based on observable data. Background Art

[0002] The field of fault diagnosis technology mainly involves the use of various methods and systems to identify, analyze and predict potential or actual faults in equipment, machinery or software systems. The core of this technical field is to enhance the reliability and efficiency of the system and reduce downtime and maintenance costs by detecting problems early. Fault diagnosis can include technologies such as data acquisition, signal analysis, pattern recognition and machine learning to automatically detect and report anomalies in system behavior. With the popularization of the Internet of Things and smart devices, fault diagnosis technology has become an important tool for maintaining the normal operation of critical infrastructure such as industry, transportation, and information technology systems.

[0003] Among them, the failure prediction method based on observable data focuses on the development and application of algorithms or models to predict possible failures in business operation platforms (such as servers, network equipment, and application platforms). This type of method uses various observable data collected from business systems, such as performance indicators, log files, and user interaction data, to predict possible problems in the future by analyzing these data. Such predictive capabilities allow IT managers to take measures in advance to prevent or mitigate the impact of failures and ensure business continuity and service reliability. The purpose of this topic is mainly to enhance the stability of the business platform and user satisfaction, while reducing economic losses caused by system failures.

[0004] Existing fault diagnosis technologies are usually unable to efficiently process large-scale, dynamically changing data, especially when dealing with complex system dependencies, often showing problems of slow processing speed and insufficient diagnostic accuracy. In a rapidly changing business environment, this lag leads to an inability to respond to system failures in a timely manner, increasing system instability and potential operational risks. The lack of effective real-time data analysis tools and fault propagation prediction models means that traditional methods often fail to respond in a timely manner after a failure occurs, and are unable to accurately grasp the scope of the fault and the recovery time, which directly affects business continuity and customer satisfaction. In addition, the delayed update of fault data and the non-real-time nature of result analysis further limit the efficiency of fault recovery, resulting in high maintenance costs and waste of resources. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings in the prior art and to propose a fault prediction method and system based on observable data.

[0006] In order to achieve the above object, the present invention adopts the following technical solution: a fault prediction method based on observable data, comprising the following steps:

[0007] S1: Select and collect the real-time load data and topology data of the business platform, map the service components to the nodes of the graph, map the dependencies to the edges of the graph, generate a business topology graph, and use the graph convolutional network to extract node features from the business topology graph to obtain a node feature matrix;

[0008] S2: According to the node feature matrix, perform graph convolution operation, update node status, compare the current node status with the historical status, filter nodes based on status deviation, determine faulty nodes and fault types, and obtain preliminary fault location;

[0009] S3: Based on the preliminary fault location, select the network equipment and database directly connected to the fault node, identify the neighboring nodes affected by the fault node by analyzing the load and response time, generate an impact range analysis, and predict the fault propagation path and potential impact based on the key nodes in the impact range analysis to obtain the fault propagation analysis result;

[0010] S4: Input the fault propagation analysis result into an abnormality location algorithm, analyze the root cause of the fault point, compare the characteristics of the fault node with the known fault mode, and establish a fault cause matching result.

[0011] As a further solution of the present invention, the steps of obtaining the service topology map are specifically as follows:

[0012] S111: Select and integrate real-time load data and topology data from the business platform using the formula:

[0013] ,

[0014] Calculate the real-time load average value S;

[0015] in, represents the real-time load of the i-th service component, represents the priority factor of the i-th service component, c is a constant adjustment term, n represents the number of service components, is the smoothing parameter used to avoid division by zero errors;

[0016] S112: Map the real-time load average value S and the service components to the nodes of the graph one by one, and map the dependencies to the edges of the graph, using the formula:

[0017] ,

[0018] Calculate the total dependency weight E of each node;

[0019] in, represents the number of dependencies of the ith node, is the maximum number of dependencies for normalization, Represents the real-time load average of the corresponding node;

[0020] S113: Based on the dependency weight sum E and in combination with the mapping information of nodes and edges, a service topology graph is constructed using the formula:

[0021] ,

[0022] Generate a business topology graph G;

[0023] in, represents the dependency weight of the i-th node, and It is used to adjust the normalization of dependency weights to make the results more stable. n represents the total number of nodes.

[0024] As a further solution of the present invention, the step of obtaining the node feature matrix is ​​specifically as follows:

[0025] S121: Using the service topology graph G, extract the load value and dependency weight of each node, combine and generate an initial feature vector, using the formula:

[0026] ,

[0027] Calculate the initial eigenvalue X of the node;

[0028] in, represents the initial eigenvalue of the i-th node, represents the load value of the i-th node, represents the dependency weight of the i-th node, is the adjustment factor;

[0029] S122: Input the initial eigenvalue X of the node into the graph convolutional network, and combine the connection weights and degree information between the nodes, using the formula:

[0030] ,

[0031] Calculate the node feature matrix H after convolution;

[0032] in, represents the set of neighbor nodes of the ith node, is the activation function, and denote the degrees of the i-th and j-th nodes respectively, and is the weight matrix, b is the initial eigenvalue of the jth node of the feature matrix after bias convolution, is the location information of the jth node;

[0033] S123: Standardize the node feature matrix H after convolution using the formula:

[0034] ,

[0035] Establish the node feature matrix F;

[0036] in, Represents the average value of the node feature matrix after convolution, Represents the standard deviation of the node feature matrix after convolution.

[0037] As a further solution of the present invention, the preliminary fault location acquisition step is specifically:

[0038] S211: Based on the node feature matrix, perform graph convolution operation to update the node state, using the formula:

[0039] ,

[0040] Calculate the new state of each node , generate the node status update result U;

[0041] in, is the weight from node i to j, represents the eigenvalue of the jth neighbor node, represents the set of neighbor nodes of the ith node, is the number of neighbor nodes, is the smoothing parameter to avoid the denominator being zero;

[0042] S212: Based on the node status update result, compare the current node status with the historical status, using the formula:

[0043] ,

[0044] Establishing state deviation results ;

[0045] in, represents the normalized state deviation, is the current state, It is a historical state. is the fault judgment threshold, is the regulating factor, is the nonlinear influence of the historical state;

[0046] S213: Determine whether the state deviation result exceeds a preset threshold , using the logical judgment formula:

[0047] ,

[0048] List all faulty nodes i and obtain preliminary fault location.

[0049] As a further solution of the present invention, the acquisition step of the impact range analysis is specifically:

[0050] S311: Select the network device and database directly connected to the faulty node, using the formula:

[0051] ,

[0052] Calculate the average performance impact index , identify the affected direct neighboring nodes and generate the performance index V of the affected nodes;

[0053] in, represents the load of the jth device or database, Represents the response time, and are the weight factors for load and response time, respectively, and n is the number of directly connected devices or databases;

[0054] S312: Analyze the load and response time through the performance index of the affected node, using the formula:

[0055] ,

[0056] Calculate the influence range score of each node and generate the influence range analysis T;

[0057] in, represents the influence range score of the node, is the performance impact index of node k, is the connection strength between node i and node k, is the connection strength between node i and node j, and m is the number of all referenced nodes.

[0058] As a further solution of the present invention, the step of obtaining the fault propagation analysis result is specifically:

[0059] S321: According to the impact range analysis, predict the fault propagation path using the formula:

[0060] ,

[0061] Calculate the potential fault propagation path score of each key node and generate the fault propagation path P;

[0062] in, represents the fault propagation score of the kth key node, is the influence range score of node i, is the probability of failure node k affecting node i, is the set of all nodes identified in the impact range analysis;

[0063] S322: Analyze the fault propagation path and predict potential impacts using the formula:

[0064] ,

[0065] Determine the fault propagation path based on the possibility, calculate the maximum impact, and obtain the fault propagation analysis result Q;

[0066] in, is the fault propagation score caused by the kth critical node, and K is the set of all critical nodes.

[0067] As a further solution of the present invention, the step of obtaining the fault cause matching result is specifically as follows:

[0068] S411: Input the fault propagation analysis result into the abnormality location algorithm to analyze the root cause of the fault point, using the formula:

[0069] ,

[0070] Calculate the root cause score of fault point i and generate the root cause analysis result A;

[0071] in, represents the root cause score of fault point i, is the characteristic vector of the faulty node S and the fault mode j, is the weight, and is the adjustment parameter, n is the number of reference failure modes;

[0072] S412: Based on the root cause analysis result, the characteristics of the faulty node are compared with the known fault modes, using the formula:

[0073] ,

[0074] Calculate the highest matching fault mode score and generate a fault cause matching result M;

[0075] in, represents the highest matching fault mode score of faulty node i, is the root cause score of fault point i, is the matching strength parameter of fault mode j, is the smoothing parameter, is the set of all reference failure modes, and m is the total number of failure points in the analysis.

[0076] A fault prediction system based on observable data, the fault prediction system based on observable data is used to execute the above-mentioned fault prediction method based on observable data, the system comprising:

[0077] The data collection module selects and integrates the real-time load data and topology data of the business platform, creates a graph structure based on service components and dependencies, and generates a business topology graph;

[0078] The feature extraction module applies a graph convolutional network to extract key features of nodes based on the business topology graph, executes a deep learning algorithm to update the node status, and generates a node feature matrix;

[0079] The fault location module compares the current state of the node with the historical state based on the node feature matrix, identifies abnormal deviations, thereby determining potential fault nodes and fault types, and generating preliminary fault location;

[0080] The impact analysis module selects directly connected nodes to the faulty node based on the preliminary fault location, analyzes the load and response time of the directly connected nodes, identifies affected adjacent nodes, predicts the fault propagation path and potential impact, and generates a fault propagation analysis result;

[0081] The fault diagnosis module inputs the fault propagation analysis result into the abnormality location algorithm, compares it with the known fault mode, determines the cause of the fault, and establishes a fault cause matching result.

[0082] Compared with the prior art, the advantages and positive effects of the present invention are:

[0083] In the present invention, a business topology graph is generated by integrating real-time load data and topology data, and a graph convolutional network is used to improve the ability to extract node features from complex structures. This method, which goes deep into the core of the network structure, can effectively monitor node status in real time, accurately identify abnormal deviations, and quickly locate faulty nodes and their types. By analyzing the devices directly connected to the faulty node, it is expanded to predict the propagation path of the fault and the potential impact, thereby improving the speed of fault response. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 It is a schematic diagram of the workflow of the present invention;

[0085] Figure 2 A flowchart of the steps for obtaining a service topology map of the present invention;

[0086] Figure 3 Flow chart of the steps of obtaining the node feature matrix of the present invention;

[0087] Figure 4 A flowchart of the steps for obtaining the initial fault location of the present invention;

[0088] Figure 5 A flowchart of the steps for obtaining the impact range analysis of the present invention;

[0089] Figure 6 A flowchart of the steps for obtaining the fault propagation analysis results of the present invention;

[0090] Figure 7 This is a flow chart of the steps for obtaining the fault cause matching results of the present invention. DETAILED DESCRIPTION

[0091] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0092] In the description of the present invention, it should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, in the description of the present invention, "multiple" means two or more, unless otherwise clearly and specifically defined.

[0093] Embodiment 1:

[0094] See also Figure 1 The present invention provides a technical solution: a fault prediction method based on observable data, comprising the following steps:

[0095] S1: Select and aggregate the real-time load data and topology data of the business platform, map the service components to the nodes of the graph, map the dependencies to the edges of the graph, generate the business topology graph, and use the graph convolutional network to extract node features from the business topology graph to obtain the node feature matrix;

[0096] S2: Perform graph convolution operations based on the node feature matrix, update the node status, compare the current node status with the historical status, filter nodes based on status deviation, determine the faulty nodes and fault types, and obtain preliminary fault location;

[0097] S3: Based on the preliminary fault location, select the network devices and databases directly connected to the faulty node, identify the neighboring nodes affected by the faulty node by analyzing the load and response time, generate an impact range analysis, and predict the fault propagation path and potential impact based on the key nodes in the impact range analysis to obtain the fault propagation analysis results;

[0098] S4: Input the fault propagation analysis results into the anomaly location algorithm, analyze the root cause of the fault point, compare the characteristics of the fault node with the known fault mode, and establish the fault cause matching result.

[0099] The node feature matrix includes node processing capability, node communication frequency, and node dependency strength. The preliminary fault location specifically includes the fault occurrence node, fault category, and fault response priority. The fault propagation analysis results specifically refer to the fault affected nodes, fault transmission speed, and fault impact depth. The fault cause matching results include matching fault causes and fault impact scores.

[0100] See also Figure 2 , the specific steps for obtaining the business topology map are:

[0101] S111: Select and integrate real-time load data and topology data from the business platform using the formula:

[0102] ,

[0103] Calculate the real-time load average value S;

[0104] in, represents the real-time load of the i-th service component, represents the priority factor of the i-th service component, c is a constant adjustment term, n represents the number of service components, is the smoothing parameter used to avoid division by zero errors;

[0105] S112: Map the real-time load average value S and the service components to the nodes of the graph one by one, and map the dependencies to the edges of the graph, using the formula:

[0106] ,

[0107] Calculate the total dependency weight E of each node;

[0108] in, represents the number of dependencies of the ith node, is the maximum number of dependencies for normalization, Represents the real-time load average of the corresponding node;

[0109] S113: Based on the total dependency weight E and the mapping information of nodes and edges, a business topology graph is constructed using the formula:

[0110] ,

[0111] Generate a business topology graph G;

[0112] in, represents the dependency weight of the i-th node, and It is used to adjust the normalization of dependency weights to make the results more stable. n represents the total number of nodes.

[0113] formula:

[0114] ,

[0115] Assume there are three service components with real-time load and priority factor as follows:

[0116] ,

[0117] The constant adjustment term c=10, the smoothing parameter ;

[0118] Calculation process:

[0119] Calculate the adjusted real-time load for each service component: ,

[0120] ,

[0121] Calculate the weighted average of the real-time load: ,

[0122] The result S represents the weighted average real-time load, reflecting the average value adjusted by the comprehensive status of each component load.

[0123] formula:

[0124] ,

[0125] Assume that the number of dependencies of three nodes is As follows, and use :

[0126] ,

[0127] ,

[0128] Calculation process:

[0129] Normalize the dependency numbers and calculate the dependency weights: ,

[0130] ,

[0131] Calculate the sum of dependency weights: ,

[0132] The result E represents the sum of all node dependency weights, which quantifies the strength of the connections between nodes.

[0133] formula:

[0134] ,

[0135] Use again and calculation The average and standard deviation :

[0136] ,

[0137] ,

[0138] Calculation process:

[0139] Normalize and calculate the adjusted dependency weights for each node: ,

[0140] ,

[0141] Calculate the characteristic values ​​of the service topology graph: ,

[0142] The result G represents an overall characteristic measure of the business topology graph, reflecting the balance of dependencies between nodes and the adjustment of connection strength.

[0143] See also Figure 3 , the specific steps for obtaining the node feature matrix are:

[0144] S121: Using the service topology graph G, extract the load value and dependency weight of each node, combine and generate an initial feature vector, using the formula:

[0145] ,

[0146] Calculate the initial eigenvalue X of the node;

[0147] in, represents the initial eigenvalue of the i-th node, represents the load value of the i-th node, represents the dependency weight of the i-th node, is the adjustment factor;

[0148] S122: Input the initial eigenvalue X of the node into the graph convolutional network, and combine the connection weights and degree information between the nodes, using the formula:

[0149] ,

[0150] Calculate the node feature matrix H after convolution;

[0151] in, represents the set of neighbor nodes of the ith node, is the activation function, and denote the degrees of the i-th and j-th nodes respectively, and is the weight matrix, b is the initial eigenvalue of the jth node of the feature matrix after bias convolution, is the location information of the jth node;

[0152] S123: Standardize the node feature matrix H after convolution using the formula:

[0153] ,

[0154] Establish the node feature matrix F;

[0155] in, Represents the average value of the node feature matrix after convolution, Represents the standard deviation of the node feature matrix after convolution.

[0156] formula: ,

[0157] in:

[0158] represents the initial eigenvalue of the i-th node;

[0159] Represents the load value of the i-th node; (Assumed to be between 10 and 100, select according to actual situation)

[0160] Represents the dependency weight of the i-th node; (assuming the range is between 1 and 10, select according to the actual situation)

[0161] is the adjustment factor, which takes a value of 1.5;

[0162] Example:

[0163] Assume that the load value of the first node is , dependent weight , then the initial eigenvalue The calculation of is as follows:

[0164] ,

[0165] Generate the node initial feature matrix X, including the initial eigenvalues ​​of all nodes.

[0166] formula: ,

[0167] in:

[0168] H represents the node feature matrix after convolution;

[0169] is the activation function (ReLU);

[0170] and Represents the degree of the i-th and j-th nodes respectively; (the number of node connections)

[0171] and is the weight matrix;

[0172] b is the bias term;

[0173] is the initial eigenvalue of the jth node;

[0174] is the other eigenvalue of the jth node, such as the convolutional feature matrix of the node's location information. Each node has two neighboring nodes. , the neighbor node degrees are 3 and 4 respectively, and the weight matrix , , bias term b=0.1, and the initial eigenvalues ​​are and , and the other eigenvalues ​​are and , the activation function is ReLU.

[0175] Calculate the convolution eigenvalue of the first node:

[0176] ,

[0177] Assuming the activation function is ReLU, then ,

[0178] Generate the convolutional node feature matrix H.

[0179] formula: ,

[0180] in:

[0181] F represents the standardized node feature matrix;

[0182] Represents the average value of the node feature matrix after convolution; Represents the standard deviation of the node feature matrix after convolution;

[0183] Example:

[0184] Assume that the mean value of the node feature matrix H after convolution is 30 and the standard deviation is 5.

[0185] Calculate the feature matrix after the normalized section convolution

[0186] Generate a standardized node feature matrix F.

[0187] See also Figure 4 , the initial fault location acquisition steps are as follows:

[0188] S211: Based on the node feature matrix, perform graph convolution operation to update the node status using the formula:

[0189] ,

[0190] Calculate the new state of each node , generate the node status update result U;

[0191] in, is the weight from node i to j, represents the eigenvalue of the jth neighbor node, represents the set of neighbor nodes of the ith node, is the number of neighbor nodes, is the smoothing parameter to avoid the denominator being zero;

[0192] S212: Based on the node status update result, compare the current node status with the historical status, using the formula:

[0193] ,

[0194] Establishing state deviation results ;

[0195] in, represents the normalized state deviation, is the current state, It is a historical state. is the fault judgment threshold, is the regulating factor, is the nonlinear influence of the historical state;

[0196] S213: Determine whether the state deviation result exceeds a preset threshold , using the logical judgment formula:

[0197] ,

[0198] List all faulty nodes i and obtain preliminary fault location.

[0199] Assume the formula:

[0200] ,

[0201] Parameter explanation and examples:

[0202] : The updated state of node i; : The weight of node i to j, assuming that for all node connections, the default weight is 0.5; : The eigenvalue of node j, assuming that the eigenvalues ​​of nodes are all 1; : The set of neighbor nodes of node i, assuming that each node has 3 neighbors; : Smoothing parameter, set to 0.01 to avoid the situation where the denominator is zero.

[0203] Calculation process and examples:

[0204] Assume that node i has three neighbors, all of which have eigenvalues ​​of 1 and weights of 0.5:

[0205] ,

[0206] ,

[0207] Assume that the sigmoid function is used as the activation function:

[0208] ,

[0209] ,

[0210] This means that the new state of the node is 0.704, which represents the activity of the node after adjustment by the activation function.

[0211] Assume the formula:

[0212] ,

[0213] Parameter explanation and examples:

[0214] : Normalized state deviation; : The current state is 0.704; : Historical state, assumed to be 0.5; : Adjustment factor, set to 1; : The nonlinear influence of the historical state, assumed to be 2; : Fault judgment threshold, assumed to be 0.2;

[0215] Calculation process and examples:

[0216] ,

[0217] Assume the formula: ,

[0218] Parameter explanation and examples:

[0219] E: the set of faulty nodes; : The result is 0.1632; : Fault judgment threshold, set to 0.2;

[0220] Calculation process and examples:

[0221] because Less than threshold , node i is judged as a non-faulty node.

[0222] See also Figure 5 , the specific steps for obtaining the impact range analysis are:

[0223] S311: Select the network device and database directly connected to the faulty node, using the formula:

[0224] ,

[0225] Calculate the average performance impact index , identify the affected direct neighboring nodes and generate the performance index V of the affected nodes;

[0226] in, represents the load of the jth device or database, Represents the response time, and are the weight factors for load and response time, respectively, and n is the number of directly connected devices or databases;

[0227] S312: Analyze the load and response time through the performance index of the affected node, using the formula:

[0228] ,

[0229] Calculate the influence range score of each node and generate the influence range analysis T;

[0230] in, represents the influence range score of the node, is the performance impact index of node k, is the connection strength between node i and node k, is the connection strength between node i and node j, and m is the number of all referenced nodes.

[0231] formula: ,

[0232] in:

[0233] : The average performance index of nodes affected by the failure; : The load of the jth device or database; : The response time of the jth device or database; and : Weighting factor for load and response time; n: Number of directly connected devices or databases;

[0234] Assumptions:

[0235] n=3 (three devices),

[0236] , (Weight Factor)

[0237] (Load value)

[0238] (Response time, in milliseconds)

[0239] Calculation process:

[0240] For the first device: ,

[0241] For the second device: ,

[0242] For the third device: ,

[0243] Average performance impact index:

[0244] ,

[0245] formula:

[0246] ,

[0247] in: : The influence range score of node i; : Performance impact index of node k; : the connection strength between node i and node k; m: the number of all nodes considered;

[0248] Assumptions:

[0249] m=3 (including the original faulty node)

[0250] (Assuming connection strength)

[0251] (The performance impact index of each node is the same)

[0252] Calculation process:

[0253] Denominator ,

[0254] For the first node: ,

[0255] For the second node: ,

[0256] For the third node: ,

[0257] Impact range analysis results: ,

[0258] This value indicates the overall impact of the failure after considering the connection strength, multi-node load and response time.

[0259] See also Figure 6 ,The specific steps for obtaining the fault propagation analysis results are:

[0260] S321: Based on the impact range analysis, predict the fault propagation path using the formula:

[0261] ,

[0262] Calculate the potential fault propagation path score of each key node and generate the fault propagation path P;

[0263] in, represents the fault propagation score of the kth key node, is the influence range score of node i, is the probability of failure node k affecting node i, is the set of all nodes identified in the impact range analysis;

[0264] S322: Analyze the fault propagation path and predict the potential impact using the formula:

[0265] ,

[0266] Determine the fault propagation path based on the possibility, calculate the maximum impact, and obtain the fault propagation analysis result Q;

[0267] in, is the fault propagation score caused by the kth critical node, and K is the set of all critical nodes.

[0268] formula: ,

[0269] in: : The fault propagation score of the kth key node, : The influence range score of node i, : The probability of failure node k affecting node i, : The set of all nodes identified in the impact range analysis;

[0270] Assumptions:

[0271] set up There are 3 nodes in it, k is a key node

[0272] ,

[0273] ,

[0274] Derivation process:

[0275] Calculate the denominator: ,

[0276] Compute the numerator for each node:

[0277] Node 1: ,

[0278] Node 2: ,

[0279] Node 3: ,

[0280] Add up the numerator values ​​and divide by the denominator: ,

[0281] Result analysis: It represents the fault propagation score of key node k, indicating the potential influence of node k on fault propagation in the network.

[0282] formula: ,

[0283] Where: Q: the result of the fault propagation analysis, showing the maximum potential impact of fault propagation; : Fault propagation scores of multiple critical nodes;

[0284] Assumptions:

[0285] use The P values ​​of the other two nodes are 28 and 33 respectively.

[0286] Derivation process:

[0287] Calculate the square root of each node's score:

[0288] ,

[0289] Select the maximum value: ,

[0290] Result interpretation: Q = 6.28 represents the maximum predicted fault propagation impact in the network, which means that the fault is estimated to propagate through node k and cause the most serious impact.

[0291] See also Figure 7 , the specific steps for obtaining the fault cause matching result are:

[0292] S411: Input the fault propagation analysis result into the anomaly location algorithm to analyze the root cause of the fault point using the formula:

[0293] ,

[0294] Calculate the root cause score of fault point i and generate the root cause analysis result A;

[0295] in, represents the root cause score of fault point i, is the characteristic vector of the faulty node S and the fault mode j, is the weight, and is the adjustment parameter, n is the number of reference failure modes;

[0296] S412: Based on the root cause analysis results, the characteristics of the faulty node are compared with the known fault modes, using the formula:

[0297] ,

[0298] Calculate the highest matching fault mode score and generate the fault cause matching result M;

[0299] in, represents the highest matching fault mode score of faulty node i, is the root cause score of fault point i, is the matching strength parameter of fault mode j, is the smoothing parameter, is the set of all reference failure modes, and m is the total number of failure points in the analysis.

[0300] formula: ,

[0301] in: represents the root cause score of fault point i, is the characteristic vector of the faulty node S and the fault mode j, is the weight, and is the adjustment parameter, n is the number of reference failure modes;

[0302] Parameter calculation and examples:

[0303] Assume there are three failure modes in the system, the feature vector weight They are 0.3, 0.5, and 0.2 respectively, and the adjustment parameters is 1, is 0.1, assuming that the fault node gate and the characteristic vectors of these three modes , , They are 4, 5, and 6 respectively;

[0304] The calculation process is as follows:

[0305] ,

[0306] The root cause analysis result A is generated, indicating that the root cause score of the fault point is 16.264.

[0307] formula: ,

[0308] in: represents the highest matching fault mode score of faulty node i, is the root cause score of fault point i, is the matching strength parameter of fault mode j, is the smoothing parameter, is the set of all reference failure modes, m is the total number of failure points in the analysis;

[0309] Assume there are three faulty nodes with root cause scores of 16.624, 14.526, and 12.345 respectively, and the matching strength parameter are 0.8, 0.7 and 0.9 respectively, and the smoothing parameters is 0.1;

[0310] The calculation process is as follows:

[0311] ,

[0312] The fault cause matching result M is generated, indicating that the score of the fault mode with the highest matching value between the faulty node i and the faulty node is 0.338.

[0313] A fault prediction system based on observable data, the fault prediction system based on observable data is used to execute the above-mentioned fault prediction method based on observable data, the system comprises:

[0314] The data collection module selects and integrates the real-time load data and topology data of the business platform, creates a graph structure based on service components and dependencies, and generates a business topology graph;

[0315] The feature extraction module uses graph convolutional networks to extract key features of nodes based on the business topology graph, executes deep learning algorithms to update node status, and generates a node feature matrix;

[0316] The fault location module compares the current state of the node with the historical state based on the node feature matrix, identifies abnormal deviations, and thus determines the potential fault nodes and fault types, and generates preliminary fault location;

[0317] The impact analysis module selects directly connected nodes to the faulty node based on the preliminary fault location, analyzes the load and response time of the directly connected nodes, identifies the affected adjacent nodes, predicts the fault propagation path and potential impact, and generates fault propagation analysis results;

[0318] The fault diagnosis module inputs the fault propagation analysis results into the anomaly location algorithm, compares them with the known fault patterns, determines the cause of the fault, and establishes the fault cause matching results.

[0319] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A fault prediction method based on observable data, It is characterized in that The following steps are involved: Select and collect the real-time load data and topology data of the business platform, map the service components to the nodes of the graph, map the dependencies to the edges of the graph, generate a business topology graph, and use a graph convolutional network to extract node features from the business topology graph to obtain a node feature matrix; According to the node feature matrix, a graph convolution operation is performed to update the node state, compare the current node state with the historical state, filter the nodes based on the state deviation, determine the faulty nodes and fault types, and obtain preliminary fault location; Based on the preliminary fault location, select the network equipment and database directly connected to the fault node, identify the neighboring nodes affected by the fault node by analyzing the load and response time, generate an impact range analysis, predict the fault propagation path and potential impact based on the key nodes in the impact range analysis, and obtain the fault propagation analysis result; The fault propagation analysis result is input into the anomaly location algorithm to analyze the root cause of the fault point, compare the characteristics of the fault node with the known fault mode, and establish a fault cause matching result; The steps for obtaining the impact range analysis are specifically as follows: Select the network device and database directly connected to the faulty node and use the formula: , Calculate the average performance impact index , identify the affected direct neighboring nodes and generate the performance index V of the affected nodes; in, represents the load of the jth device or database, Represents the response time, and are the weight factors for load and response time, respectively, and n is the number of directly connected devices or databases; By analyzing the load and response time through the performance index of the affected node, the formula is used: , Calculate the influence range score of each node and generate the influence range analysis T; in, represents the influence range score of the node, is the performance impact index of node k, is the connection strength between node i and node k, is the connection strength between node i and node j, and m is the number of all referenced nodes; The steps for obtaining the fault propagation analysis result are specifically as follows: According to the impact range analysis, the fault propagation path is predicted using the formula: , Calculate the potential fault propagation path score of each key node and generate the fault propagation path P; in, represents the fault propagation score of the kth key node, is the influence range score of node i, is the probability of failure node k affecting node i, is the set of all nodes identified in the impact range analysis; Analyze the fault propagation path and predict the potential impact using the formula: , Determine the fault propagation path based on the possibility, calculate the maximum impact, and obtain the fault propagation analysis result Q; in, is the fault propagation score caused by the kth critical node, K is the set of all critical nodes; The steps for obtaining the fault cause matching result are specifically as follows: The fault propagation analysis results are input into the anomaly location algorithm to analyze the root cause of the fault point using the formula: , Calculate the root cause score of fault point i and generate the root cause analysis result A; in, represents the root cause score of fault point i, is the characteristic vector of the faulty node S and the fault mode j, is the weight, and is the adjustment parameter, n is the number of reference failure modes; Based on the root cause analysis results, the characteristics of the faulty node are compared with the known fault modes, using the formula: , Calculate the highest matching fault mode score and generate a fault cause matching result M; in, represents the highest matching fault mode score of faulty node i, is the root cause score of fault point i, is the matching strength parameter of fault mode j, is the smoothing parameter, is the set of all reference failure modes, and m is the total number of failure points in the analysis.

2. The fault prediction method based on observable data according to claim 1, It is characterized in that Select and aggregate real-time load data and topology data from the business platform using the formula: , Calculate the real-time load average value S; in, represents the real-time load of the i-th service component, represents the priority factor of the i-th service component, c is a constant adjustment term, n represents the number of service components, is the smoothing parameter used to avoid division by zero errors; The real-time load average value S and service components are mapped one by one to the nodes of the graph, and the dependencies are mapped to the edges of the graph, using the formula: , Calculate the total dependency weight E of each node; in, represents the number of dependencies of the ith node, is the maximum number of dependencies for normalization, Represents the real-time load average of the corresponding node; Based on the dependency weight sum E, combined with the mapping information of nodes and edges, a business topology graph is constructed using the formula: , Generate a service topology graph G; in, represents the dependency weight of the i-th node, and It is used to adjust the normalization of dependency weights to make the results more stable. n represents the total number of nodes.

3. The fault prediction method based on observable data according to claim 2, It is characterized in that The steps for obtaining the node feature matrix are specifically as follows: Using the service topology graph G, extract the load value and dependency weight of each node, combine them to generate the initial feature vector, using the formula: , Calculate the initial eigenvalue X of the node; in, represents the initial eigenvalue of the i-th node, represents the load value of the i-th node, represents the dependency weight of the i-th node, is the adjustment factor; The initial eigenvalue X of the node is input into the graph convolutional network, and the connection weight and degree information between nodes are combined, using the formula: , Calculate the node feature matrix H after convolution; in, represents the set of neighbor nodes of the ith node, is the activation function, and denote the degrees of the i-th and j-th nodes respectively, and is the weight matrix, b is the initial eigenvalue of the jth node of the feature matrix after bias convolution, is the location information of the jth node; The node feature matrix H after convolution is standardized using the formula: , Establish the node feature matrix F; in, Represents the average value of the node feature matrix after convolution, Represents the standard deviation of the node feature matrix after convolution.

4. The fault prediction method based on observable data according to claim 3, It is characterized in that The steps for obtaining the preliminary fault location are specifically as follows: Based on the node feature matrix, a graph convolution operation is performed to update the node state using the formula: , Calculate the new state of each node , generate the node status update result U; in, is the weight from node i to j, represents the eigenvalue of the jth neighbor node, represents the set of neighbor nodes of the ith node, is the number of neighbor nodes, is the smoothing parameter to avoid the denominator being zero; Based on the node status update result, the current node status is compared with the historical status, using the formula: , Establishing state deviation results ; in, represents the normalized state deviation, is the current state, It is a historical state. is the fault judgment threshold, is the regulating factor, is the nonlinear influence of the historical state; Determine whether the state deviation result exceeds the preset threshold , using the logical judgment formula: , List all faulty nodes i and obtain preliminary fault location.

5. A fault prediction system based on observable data, It is characterized in that According to the method for fault prediction based on observable data according to any one of claims 1 to 4, the system comprises: The data collection module selects and aggregates the real-time load data and topology data of the business platform, creates a graph structure based on service components and dependencies, and generates a business topology graph; The feature extraction module applies a graph convolutional network to extract key features of nodes based on the business topology graph, executes a deep learning algorithm to update the node status, and generates a node feature matrix; The fault location module compares the current state of the node with the historical state based on the node feature matrix, identifies abnormal deviations, thereby determining potential fault nodes and fault types, and generating preliminary fault location; The impact analysis module selects directly connected nodes to the faulty node based on the preliminary fault location, analyzes the load and response time of the directly connected nodes, identifies affected adjacent nodes, predicts the fault propagation path and potential impact, and generates a fault propagation analysis result; The fault diagnosis module inputs the fault propagation analysis result into the abnormality location algorithm, compares it with the known fault mode, determines the cause of the fault, and establishes a fault cause matching result.