Fault root cause dynamic prediction method and system based on probabilistic propagation reasoning

Abstract

The disclosure provides a fault root cause dynamic prediction method and system based on probability propagation reasoning, relates to the technical field of fault positioning prediction, and comprises the following steps: constructing a directed graph structure, associating a dynamic fault probability with each node, and constructing a fault state expression system; based on the fault state expression system, calculating and dynamically updating the edge weight to construct a dynamic propagation graph model; generating the fault occurrence probability of the monitoring data, and fusing the fault occurrence probability to obtain aligned unified probability data; based on the dynamic propagation graph model, performing graph probability propagation reasoning on the unified probability data to determine the final root cause, thereby realizing rapid fault positioning; introducing an intelligent self-healing and strategy optimization algorithm based on reinforcement learning to automatically generate and execute a repair strategy, and continuously optimizing the decision through online learning to form a self-evolving safe operation and maintenance system. The disclosure realizes accurate and rapid prediction and positioning of system faults.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of fault location prediction, in particular to a fault root cause dynamic prediction method and system based on probabilistic propagation reasoning. BACKGROUND

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute the prior art.

[0003] With the deepening of the digitization process in the government field, the internal logic of the information system supporting business operation is becoming increasingly complex, and the dependency relationship between components presents a highly dynamic and meshed characteristic. In such an environment, the abnormality of any component state can quickly spread through the dependency chain, resulting in a large area of business function being damaged. The existing mainstream fault analysis scheme usually relies on centralized analysis of multi-dimensional monitoring indicators and log data. However, these methods have the following inherent limitations in the face of complex and dynamic system environments from the perspective of data analysis:

[0004] (1) Consistency and credibility of multi-source data. The state data used for analysis comes from multiple independent collection processes or computing engines, and these data have inherent differences in generation time, collection period and calculation logic. When directly aggregated and analyzed, the data itself has internal contradictions due to time offset, non-uniform dimension and unknown confidence, resulting in a distorted system state view based on contradictory data.

[0005] (2) Insufficient root cause quantization accuracy. Generally, FTA (Fault Tree Analysis) algorithm is used for cause positioning, but this fault analysis algorithm has defects. On the one hand, it does not verify the authenticity of the fault basic event in combination with multi-source data, which is prone to one-sided information judgment. On the other hand, it lacks a dynamic adjustment mechanism, making it difficult to cope with the impact of system dynamic changes (such as resource scheduling and business iteration) on fault correlation, resulting in insufficient accuracy of root cause judgment.

[0006] (3) Insufficient explainability and quantization accuracy of root cause inference process. Traditional analysis methods often rely on pre-set rules or simple probability statistics when selecting the root cause from a large number of alarm events, lacking a comprehensive model that can quantitatively evaluate the importance of each potential cause. This results in an "black box" inference process, which cannot provide clear and quantitative priority ranking among multiple possible root causes, thereby affecting the efficiency of fault decision-making. SUMMARY

[0007] In order to solve the above problems, the present disclosure proposes a fault root cause dynamic prediction method and system based on probabilistic propagation reasoning, which constructs a dynamic propagation graph model that can perceive system state changes, fuses multi-source heterogeneous data, and performs quantifiable probabilistic propagation reasoning, ultimately achieving accurate and dynamic positioning and prediction of root causes.

[0008] According to some embodiments, the present disclosure adopts the technical solutions as follows:

[0009] A fault root cause dynamic prediction method based on probability propagation reasoning comprises:

[0010] A globally unique identifier is established for each component and its dependency relationship is obtained;

[0011] Based on the dependency relationship, a node type is defined and a directed graph structure is constructed, each node is associated with a dynamic fault probability, and a fault state expression system is constructed;

[0012] Based on the fault state expression system, a dynamic factor reflecting the real-time state of the system is introduced, the edge weight is calculated and dynamically updated, and a dynamic propagation graph model is constructed;

[0013] Multi-dimensional monitoring data is obtained, the original data is converted into standardized fault probability through a pre-defined index-probability mapping function, and multi-source data is processed using a fusion strategy to obtain aligned unified probability data;

[0014] Based on the dynamic propagation graph model, the unified probability data is used for graph probability propagation calculation, the fault probability of each component is deduced, and a root cause candidate set is extracted; a key importance weighted model is introduced to multi-dimensionally quantify and sort the root cause candidate set, and the final root cause is predicted and located according to the sorting result;

[0015] An intelligent self-healing and strategy optimization algorithm based on reinforcement learning is introduced, a repair strategy is automatically generated and executed based on the final root cause, and the decision is continuously optimized through online learning to form a self-evolving operation and maintenance closed loop.

[0016] According to some embodiments, the present disclosure adopts the technical solutions as follows:

[0017] A fault root cause dynamic prediction system based on probability propagation reasoning comprises:

[0018] An initialization module is configured to establish a globally unique identifier for each component and obtain its dependency relationship;

[0019] A fault dependency modeling module is configured to define a node type and construct a directed graph structure based on the dependency relationship, associate each node with a dynamic fault probability, and construct a fault state expression system;

[0020] Based on the fault state expression system, a dynamic factor reflecting the real-time state of the system is introduced, the edge weight is calculated and dynamically updated, and a dynamic propagation graph model is constructed;

[0021] A multi-source data alignment module is configured to acquire multi-dimensional monitoring data, convert the original data into standardized failure probabilities through a predefined index-probability mapping function, and obtain aligned unified probability data by processing the multi-source data using a fusion strategy.

[0022] A reasoning and root cause positioning module is configured to perform graph probability propagation calculation based on the dynamic propagation graph model using the unified probability data, deduce failure probabilities of each component, and extract a root cause candidate set; a key importance weighting model is introduced to perform multi-dimensional quantitative sorting on the root cause candidate set, and the final root cause is predicted and located according to the sorting result.

[0023] A reinforcement learning optimization module is configured to introduce an intelligent self-healing and strategy optimization algorithm based on reinforcement learning, automatically generate a repair strategy based on the final root cause and execute the repair strategy, and continuously optimize the decision through online learning to form a self-evolving operation and maintenance closed loop.

[0024] According to some embodiments, the present disclosure adopts the following technical solutions:

[0025] A non-transitory computer-readable storage medium is configured to store computer instructions, which are executed by a processor to implement the fault root cause dynamic prediction method based on probability propagation reasoning.

[0026] According to some embodiments, the present disclosure adopts the following technical solutions:

[0027] An electronic device includes a processor, a memory, and a computer program; the processor is connected to the memory, and the computer program is stored in the memory; when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the fault root cause dynamic prediction method based on probability propagation reasoning.

[0028] Compared with the prior art, the present disclosure has the following beneficial effects:

[0029] The fault root cause dynamic prediction method based on probability propagation reasoning of the present disclosure converts multi-source heterogeneous data into unified probabilistic input by constructing a dynamic propagation graph model that fuses real-time state factors, significantly improving the accuracy and reliability of fault root cause positioning in a complex dependency environment; further introducing a key importance quantitative sorting and an intelligent self-healing mechanism driven by reinforcement learning, not only realizes accurate and rapid positioning of faults, but also forms an autonomous closed loop from diagnosis to repair, effectively reducing the mean time to recovery (MTTR) of the system and improving the continuity and overall operation and maintenance resilience of the government business system.

[0030] The fault root cause dynamic prediction method based on probabilistic propagation reasoning of the present disclosure upgrades the system state from a static snapshot to a dynamic image, so that the simulation of the fault propagation path is more in line with the real running conditions of the complex system, and effectively overcomes the problem of insufficient root cause judgment accuracy caused by the inability of the traditional static model to perceive the internal dynamic changes of the system.

[0031] The fault root cause dynamic prediction method based on probabilistic propagation reasoning of the present disclosure adopts a multi-source data alignment strategy based on a probabilistic time window and a confidence decay mechanism, which can effectively reconcile the differences in time and reliability from different data sources, solve the analysis distortion caused by the internal contradictions of the data, and provide high-quality and consistent probabilistic inputs for subsequent reasoning, thereby guaranteeing the reliability of the fault determination results from the source.

[0032] The fault root cause dynamic prediction method based on probabilistic propagation reasoning of the present disclosure performs probabilistic propagation reasoning on the multi-alignment fault probability data on the dynamic propagation graph, which can quantitatively evaluate the accumulation and diffusion effect of the fault on the complex dependency link, thereby systematically revealing the root cause hidden behind the local phenomenon, rather than isolatedly looking at the alarms of a single node, and greatly improving the ability to predict the real fault source in the complex components.

[0033] The fault root cause dynamic prediction method based on probabilistic propagation reasoning of the present disclosure introduces a key importance weighting model to perform multi-dimensional quantitative sorting of the root cause candidate set by fusing cut set probability, node influence range and data reliability, thereby providing clear and quantitative decision basis for the operation and maintenance personnel, avoiding the one-sidedness of relying on probability level or experience for judgment, and thereby realizing accurate screening and effective differentiation of the priority of the root cause.

[0034] The fault root cause dynamic prediction method based on probabilistic propagation reasoning of the present disclosure introduces an intelligent self-healing and strategy optimization algorithm based on reinforcement learning, and constructs a complete autonomous closed loop from fault diagnosis to repair execution, so that the system can automatically select and execute the optimal recovery strategy based on historical experience, and continuously optimize the decision through online learning, thereby forming a self-evolving safe operation and maintenance system, significantly reducing the dependence on manual intervention, and improving the system resilience and business continuity. BRIEF DESCRIPTION OF DRAWINGS

[0035] The accompanying drawings, which form a part of the present disclosure, are used to provide further understanding of the present disclosure, and the schematic embodiments of the present disclosure and the description thereof are used to explain the present disclosure, and do not constitute an improper limitation on the present disclosure.

[0036] Figure 1 A flowchart of the fault root cause dynamic prediction method based on probabilistic propagation reasoning of the present disclosure embodiment. DETAILED DESCRIPTION

[0037] The present disclosure will be further described with reference to the drawings and examples.

[0038] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the present disclosure. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs.

[0039] It should be noted that the terms used herein are only intended to describe specific embodiments and are not intended to limit exemplary embodiments according to the present disclosure. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it should also be understood that when the terms "comprise" and / or "include" are used in the specification, they indicate the presence of the features, steps, operations, devices, components and / or combinations thereof.

[0040] Example 1

[0041] In an embodiment of the present disclosure, a fault root cause dynamic prediction method based on probabilistic propagation reasoning is provided, and the method steps include:

[0042] Step 1: Establish a globally unique identifier for each component and obtain its dependency relationship;

[0043] Step 2: Based on the dependency relationship, define component types and build a directed graph structure, associate a dynamic failure probability with each node in the graph, and build a fault state expression system;

[0044] Step 3: Based on the fault state expression system, introduce a dynamic factor reflecting the real-time state of the system, calculate and continuously update the weight of the edges in the graph, and build a dynamic propagation graph model;

[0045] Step 4: Obtain monitoring data, generate fault occurrence probability of monitoring data through a set of index-probability mapping functions, and process multi-source data using a fusion strategy to obtain aligned unified probability data;

[0046] Step 5: Based on the dynamic propagation graph model, use the unified probability data to perform graph probability propagation calculation, deduce the failure probability of all components in the system, extract the root cause candidate set; introduce a key importance weighted model to multi-dimensionally quantify and sort each root cause candidate set, and predict and locate the final root cause according to the sorting result;

[0047] Step 6: Introduce an intelligent self-healing and strategy optimization algorithm based on reinforcement learning, automatically generate a repair strategy based on the final root cause and execute it, and continuously optimize the decision through online learning to form a self-evolving operation and maintenance closed loop.

[0048] As an embodiment, the probability propagation reasoning-based fault root cause dynamic prediction method of the present disclosure combines a multi-source data cross-validation method to solve the inherent contradictions of multi-source data, and further realizes accurate and rapid prediction and positioning of complex system faults through the construction of a dynamic fault propagation graph model, the calculation of a minimum cut set, the quantification of key importance, and a dynamic updating mechanism. The specific implementation process is as follows:

[0049] Step 1: Establish a globally unique identifier for each component and obtain its dependency relationship;

[0050] Specifically, using configuration management technology, a globally unique identifier is established for each logical component in the system and its dependency relationship is obtained;

[0051] Among them, the configuration management library, as the core carrier of component dependency relationship, allocates a unique asset number for all monitoring objects and records their dependency relationship, providing accurate node, edge and topology data support for subsequent dynamic propagation graph model construction. The process includes:

[0052] The monitoring objects cover the business system full-stack logical components, including application services, data middleware, computing instances, etc. A globally unique asset number is assigned to each monitoring object, and this number is used as the unique identifier in the fault reasoning process. The dynamic dependency relationship between components is obtained and recorded, and finally a continuously updated configuration item library containing unique asset numbers and dependency relationships is formed, as follows:

[0053] Step 11: Asset uniform coding and registration.

[0054] Covering business system full-stack resources, including application services (such as order service), middleware (such as Redis cluster), host instances (such as DC01-SVR-001), etc. A globally unique asset number (such as BE001, NET002) is assigned to each monitoring object (component), which is used as its unique identifier (event_id) in the subsequent fault reasoning process.

[0055] Step 12: Dependency relationship storage and graphing.

[0056] In the configuration management library, not only the static properties of the components are stored, but more importantly, the dynamic dependency relationships between the components (such as "order service depends on Redis cluster", "data processing task runs on Spark executor") are recorded. These dependency relationships form a directed graph, where the nodes are components and the edges represent dependency directions (such as "depends on" and "runs on"). This graph is the structural basis for building the fault propagation graph.

[0057] Step 13: Data health and real-time synchronization mechanism.

[0058] To ensure the reliability of the configuration management library data, a set of health status monitoring and real-time synchronization mechanism is established, including:

[0059] (1) Health monitoring: heart beat detection of configuration management library service every minute, if 3 consecutive failures, trigger high level alarm, and enable the cache of the latest dependency snapshot in each data receiving module, to ensure uninterrupted reasoning process.

[0060] (2) Automatic synchronization: integration with infrastructure as code (IaC) tools (such as Terraform) and service discovery system (such as Consul). Any infrastructure or application deployment changes will be automatically synchronized to the configuration management library within 5 minutes, ensuring the timeliness of the dependency graph. All change records are time stamped, used to distinguish data loss and normal maintenance in step 3.

[0061] Finally, a continuously updated configuration item library containing unique asset number and dependency relationship is obtained, providing complete node set (V), edge set (E) and topology structure for subsequent dynamic propagation graph model construction.

[0062] Step 2: Based on the dependency relationship, define node type and build directed graph structure, associate each node with a dynamic failure probability, build failure state expression system; based on the failure state expression system, introduce dynamic factor reflecting real-time state of the system, calculate and dynamically update edge weight, build dynamic propagation graph model, including:

[0063] Step 21: graph structure construction and node definition.

[0064] Based on the asset dependency relationship and business logic registered in the configuration management library in step 1, a directed graph G = (V, E) is constructed. Through weekly loop detection (Tarjan algorithm is used to identify strongly connected components SCC), if loop dependency is found, SCC is merged into a super node or virtual node is inserted to break the cycle, to ensure the convergence of reasoning process. Among them, node set V represents all possible fault events in the system, edge set E represents the propagation relationship between faults.

[0065] Further, nodes are divided into three types: root nodes correspond to basic components in the configuration management library, whose state is directly driven by monitoring data; propagation nodes represent services, middleware and other logical entities, with fault receiving and forwarding capability; sink nodes represent system level business faults, which are the end point of fault propagation. Each node is associated with a dynamic failure probability , which changes over time and is affected by multi-source monitoring data, forming a complete failure state expression system.

[0066] Step 22: dynamic edge weight and dynamic propagation graph model.

[0067] edge representing the fault from node to the weight of the path through which the fault propagates, which is a dynamic variable determined by the structural weight learned based on historical fault data and real-time load, call or processing delay, and the weight calculation formula is:

[0068]

[0069] wherein, is the structural weight learned based on historical fault data, reflecting the inherent tendency of fault propagation; is the load rate of node at the current time; is the load upper limit of node ; is the network delay from the current slave node to ; is the delay sensitivity coefficient, usually taking a value of 50~100ms; is the Sigmoid function, used to map the load rate to an influence factor between 0 and 1.

[0070] Through dynamic monitoring of the weight , the influence of load and delay in the fault propagation process can be captured in real time, the edge weight is calculated and dynamically updated within a certain time period, so as to update the propagation graph model, and a dynamic propagation graph model can be obtained, overcoming the shortcomings of traditional static models.

[0071] Step 23: online learning and dynamic adjustment.

[0072] To adapt to changes in the system environment, a perfect online learning mechanism is established. The edge weight is calculated and dynamically updated within a certain time period.

[0073] As an embodiment, the edge weight is updated every 5 minutes to reflect the latest network status and load situation; the structural weight is updated based on the fault data of the past 24 hours every morning using the maximum likelihood estimation method; and the data source credibility weight is re-evaluated every week to ensure the accuracy of multi-source data fusion.

[0074] Step 3: Obtain multi-dimensional monitoring data, convert the original data into standardized fault probability through a pre-defined index-probability mapping function, and process multi-source data using a fusion strategy to obtain aligned unified probability data;

[0075] ​Specifically, under the complex system architecture of the present disclosure, the monitoring data collected by each data source has time deviation and confidence difference. To achieve high-precision fault reasoning, the present disclosure proposes a multi-source data alignment method based on a probabilistic time window, which provides standardized input for a dynamic propagation graph model. This method realizes the unified processing of multi-source monitoring data through time window alignment, probability fusion calculation, and confidence decay mechanism, as follows:

[0076] Step 31: Calculate the fault occurrence probability in the monitoring data based on the monitoring indicators, which is generated by a pre-defined indicator-probability mapping function.

[0077] The monitoring data is stored in the multi_source_data table, which includes the following fields: event_id (asset number), source_node (source node), timestamp (timestamp), status (status), raw_data (raw data), probability (fault probability), and confidence (confidence). The probability field represents the fault occurrence probability calculated based on the monitoring indicators, which is generated by a pre-defined indicator-probability mapping function.

[0078] (1) For continuous monitoring indicators (such as CPU usage, memory usage, network delay, etc.), a piecewise function is used for mapping:

[0079]

[0080] where, is the monitoring indicator value, is the lower limit of the normal range of the indicator, is the upper limit of the abnormal threshold of the indicator, is the intermediate threshold, is the curve steepness coefficient, which controls the rate of probability growth.

[0081] (2) For discrete monitoring indicators (such as service status, port status, etc.), a direct mapping method is used:

[0082]

[0083] Step 32: Fuse the fault occurrence probability to obtain unified probability data.

[0084] Specifically, the time alignment process uses a dynamic sliding window mechanism. With the earliest collection time as the reference, a time window (default 5 seconds) is set, and the window size is dynamically adjusted according to the network delay. For monitoring data with the same asset number, if its timestamp satisfies If the time difference is within the same processing window, the data is included in the same window. During the window period, the failure probability value reported by each data source is fused and calculated. During the window period, the failure probability value reported by each independent data source is fused, and the following formula is used to calculate the comprehensive failure probability:

[0085]

[0086] Wherein, is the fused failure probability, is the total number of data sources that have reported valid data in the window period, is the original failure probability calculated by the k th data source according to its monitoring indicators; the weight is the confidence weight of the k th data source, with a value range of [0, 1], and the initial value is set based on the historical accuracy of the data source.

[0087] Further, the data missing processing adopts an exponential decay model, and when a data source continuously fails to report data, its confidence weight decays exponentially:

[0088]

[0089] Wherein, is the weight at the current time, is the initial weight, is the decay coefficient (default value 0.1 / minute), is the continuous missing duration, which is obtained by querying the configuration change record of CMDB to exclude the data missing situation caused by system maintenance.

[0090] Further, the final output data format includes the following fields: event_id (asset number), timestamp (aligned timestamp), probability (fused probability), confidence (overall confidence), data_source_count (number of valid data sources). The overall confidence is calculated as follows, which reflects the overall confidence of the multi-source data fusion result:

[0091]

[0092] Wherein, represents the overall confidence, reflecting the confidence of the multi-source data fusion result.

[0093] Step 4: Based on the dynamic propagation graph model, perform graph-based probability propagation calculations using unified probability data to deduce the failure probability of each component and extract a root cause candidate set; introduce a key importance weighted model to perform multi-dimensional quantitative ranking of the root cause candidate set, and predict and locate the final root cause based on the ranking results, as follows:

[0094] Step 41: Propagation reasoning of fault probability on the graph.

[0095] Based on the constructed dynamic propagation graph model, the unified probability data output in step 32 is used as the initial failure probability input for the root node (basic component) in the graph. Subsequently, probability propagation calculations are performed on the graph to deduce the real-time failure probabilities of all components (all nodes in the graph) in the system. The propagation process employs a hybrid strategy combining backward reasoning and forward propagation. Backward reasoning starts from the business node (convergence node) where anomalies have been observed, tracing back along dependency edges to identify possible root causes of failures; forward propagation starts from the root node identified as having a high probability of failure, simulating the path and impact of the failure spreading downstream along the dependency chain. This hybrid strategy simultaneously considers the bottom-up propagation path of the failure and the top-down impact range of business dependencies, ensuring the comprehensiveness of the reasoning.

[0096] Each node Failure probability Through all its upstream nodes Fault probability and edge weight The calculation shows that:

[0097]

[0098] in, Represents a node The set of all upstream nodes, The dynamic edge weights are calculated in step 22. upstream node The formula represents the probability of failure. It reflects the cumulative effect of failure propagation in the system and can accurately reflect the failure state under complex dependencies.

[0099] Step 42: Extract the root cause candidate set.

[0100] Based on the dynamic propagation graph model constructed in step 2, the convergence node of interest (i.e., system-level business failure) is regarded as the top event of the fault tree, and the propagation path from the convergence node to all its reachable root nodes is logically regarded as a fault tree. Based on this fault tree view, combined with the real-time failure probability of each node calculated in step 41, all minimum cut sets corresponding to the current top event are extracted as root cause candidate sets.

[0101] The minimal cut set refers to the smallest set of basic events that can lead to the top event; each cut set represents a possible combination of failures. The specific process is as follows:

[0102] First, starting from the top event, the fault tree is traversed layer by layer downwards to identify all basic event combinations that led to the top event. Second, a dynamic update mechanism is used to ensure that the cut set is consistent with the current system state. For each cut set... Calculate their joint failure probability:

[0103]

[0104] in, For cut set The basic events included (i.e., the bottom events in the fault tree, corresponding to the root node in the dynamic propagation graph model). Basic events The current failure probability (this probability value has been calculated and updated during the full-graph probability propagation in step 41), this joint probability This reflects the possibility that the top event is caused by the combination of these basic events.

[0105] Step 43: Introduce a key importance weighted model to perform multi-dimensional quantitative ranking of each root cause candidate set.

[0106] To improve the accuracy of root cause localization, a key importance weighted model is introduced to perform multi-dimensional quantitative ranking of each root cause candidate set.

[0107] Key importance Taking into account factors such as cut set probability, node influence factor, and data source credibility, the calculation formula is as follows:

[0108]

[0109] in, For cut set The combined failure probability; The number of basic events in the cut set; For nodes The impact factor represents the scope of the system's impact from the node's failure. For nodes The confidence level of multi-source data fusion reflects the reliability of the node's failure probability.

[0110] This formula comprehensively evaluates the root cause probability of each cut set from three dimensions: probability, scope of influence, and data credibility, thereby achieving accurate ranking of root causes.

[0111] Step 44: Final root cause prediction and output.

[0112] According to the key importance ranking result, the top-ranked cut sets are selected as the root cause candidate set K K The number of cut sets is configurable, and the default is 3, and the final confirmation is combined with the cross-verification mechanism of multiple sources. The verification rules are as follows:

[0113] If all basic events in a cut set are simultaneously determined as fault states by multiple data sources, the cut set is confirmed as the root cause;

[0114] If multiple cut sets meet the condition, the cut set with the highest key importance is selected as the final root cause;

[0115] The output result includes the root cause cut set list, the key importance of each cut set, and the multi-source data identification supporting the judgment. The final output format is shown in Table 1.

[0116] Table 1 Final output format

[0117]

[0118] Step 5: Introduce an intelligent self-healing and strategy optimization algorithm based on reinforcement learning, automatically generate repair strategies based on the final root cause and execute them, and continuously optimize decisions through online learning to form a self-evolving operation and maintenance closed loop.

[0119] To realize the closed loop from "fault rapid positioning" to "fault autonomous repair" and improve the overall resilience of complex systems, an intelligent self-healing and strategy optimization module based on reinforcement learning is introduced. The module receives the root cause cut set output by step 4, automatically generates and executes repair strategies, and continuously optimizes decisions through online learning to form a self-evolving safety operation and maintenance system. Specifically as follows:

[0120] Step 51: Self-healing strategy knowledge base and action generation.

[0121] A strategy knowledge base for fault self-healing is constructed, where each strategy contains: applicable root cause mode (such as MCS001: database master node failure), execution action (such as "restart database service", "switch to standby node"), pre-execution condition (such as "standby node resource sufficient"), expected recovery time, historical success rate, etc.

[0122] When the intelligent decision module confirms the root cause cut set, strategy retrieval is automatically triggered: according to the root cause cut set ID (such as MCS001) and the current system state (such as load, resource utilization), the Top-N most suitable repair strategies are matched from the knowledge base. If there is no complete match in the knowledge base, a strategy recommendation based on graph similarity is used: calculate the similarity between the current fault propagation graph and the historical case graph, and recommend historical strategies with a similarity higher than a threshold (such as 0.85).

[0123] ​Step 52: reinforcement learning driven policy decision and execution.

[0124] (1) First, construct a reinforcement learning model framework:

[0125] State (S): State vector composed of the output root cause cut set list, node failure probability, system real-time load, network quality indicators, etc.

[0126] Action (A): A set of executable repair strategies in the knowledge base (such as action a1: restart service, action a2: traffic switching).

[0127] Reward (R): Design a multi-dimensional reward function:

[0128] to encourage rapid recovery;

[0129] to encourage low-cost operation;

[0130] to encourage low-risk operation;

[0131] .

[0132] (2) Online decision and learning:

[0133] Use the PPO (Proximal Policy Optimization) algorithm to train the RL agent so that it can select the optimal action A under the current state S.

[0134] Calculate the reward R according to the final recovery effect (success, time consumption, resource consumption), and update the agent's policy network accordingly to achieve experience accumulation and policy optimization.

[0135] Finally, output a fault self-healing execution dashboard, as shown in Table 2.

[0136] Table 2: Example of fault self-healing execution dashboard

[0137]

[0138] Example 2

[0139] In an embodiment of the present disclosure, a fault root cause dynamic prediction system based on probability propagation reasoning is provided, comprising:

[0140] An initialization module for establishing a globally unique identifier for each component and obtaining its dependency relationship;

[0141] The fault dependency modeling module is configured to define node types and build a directed graph structure based on the dependency relationships, associate each node with a dynamic fault probability, and build a fault state expression system; based on the fault state expression system, a dynamic factor reflecting the real-time state of the system is introduced, edge weights are calculated and dynamically updated, and a dynamic propagation graph model is built;

[0142] The multi-source data alignment module is configured to obtain multi-dimensional monitoring data, convert the original data into standardized fault probabilities through a pre-defined index-probability mapping function, and obtain aligned unified probability data by processing the multi-source data using a fusion strategy.

[0143] The reasoning and root cause positioning module is configured to perform graph probability propagation calculation based on the dynamic propagation graph model using the unified probability data, deduce the fault probability of each component, and extract a root cause candidate set; a key importance weighting model is introduced to multi-dimensionally quantify and sort the root cause candidate set, and the final root cause is predicted and located according to the sorting result.

[0144] The reinforcement learning optimization module is configured to introduce an intelligent self-healing and strategy optimization algorithm based on reinforcement learning, automatically generate a repair strategy based on the final root cause and execute it, and continuously optimize the decision through online learning to form a self-evolving operation and maintenance closed loop.

[0145] Embodiment 3

[0146] In an embodiment of the present disclosure, a non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the fault root cause dynamic prediction method based on probability propagation reasoning.

[0147] Embodiment 4

[0148] In an embodiment of the present disclosure, an electronic device is provided, which includes a processor, a memory, and a computer program; the processor is connected to the memory, and the computer program is stored in the memory; when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the fault root cause dynamic prediction method based on probability propagation reasoning.

[0149] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks. Figure 1 one or more flow or blocks.

[0150] These computer program instructions can also be loaded into a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks. Figure 1 one or more flow or blocks.

[0151] Although the present disclosure has been described with reference to specific implementations, it will be apparent to those skilled in the art that various modifications and changes can be made to the implementations without departing from the scope of the present disclosure as set forth in the claims below. Accordingly, the disclosure is not to be limited by what has been presented in the description above, but is to be understood as including all variations falling within the scope of the present disclosure.

Claims

1. A dynamic prediction method for root causes of failures based on probability propagation reasoning, characterized in that, include: Establish globally unique identifiers for each component and obtain its dependencies; Based on the aforementioned dependencies, node types are defined and a directed graph structure is constructed, with each node associated with a dynamic fault probability, thus building a fault state representation system. Based on the fault state representation system, a dynamic factor reflecting the real-time state of the system is introduced, edge weights are calculated and dynamically updated, and a dynamic propagation graph model is constructed, including: Edges are represented as paths through which a fault propagates from one node to other nodes; The weights of the edges are dynamic variables, determined by the structural weights learned from historical fault data and the real-time load, call or processing latency. Calculate and dynamically update edge weights within a set time period to construct a dynamic propagation graph model; Acquire multi-dimensional monitoring data, transform the raw data into standardized fault probabilities using a predefined indicator-probability mapping function, and employ a fusion strategy to process the multi-source data, obtaining aligned unified probability data, including: A multi-source data alignment method based on probabilistic time windows is adopted, which realizes unified processing of multi-source monitoring data through time window alignment, probability fusion calculation, and confidence decay mechanism. The monitoring data is stored in the multi_source_data table, and the structure of the multi_source_data table includes the following fields: asset number event_id, source node, timestamp, status, raw data raw_data, probability of failure, and confidence. Among them, the probability field represents the probability of failure occurrence in the monitoring data calculated based on monitoring indicators, which is generated by a predefined indicator-probability mapping function. Based on the dynamic propagation graph model, the unified probability data is used to perform graph-based probability propagation calculations to deduce the failure probability of each component and extract a root cause candidate set; a key importance weighting model is introduced to perform multi-dimensional quantitative ranking of the root cause candidate set, and the final root cause is predicted and located based on the ranking results. The introduction of a key importance weighted model involves multi-dimensional quantitative ranking of each root cause candidate set, and determining the final root cause based on the quantitative ranking results, including: Introduce a key importance weighted model; By comprehensively considering factors such as cut set probability, node influence factor, and data source credibility, the root cause probability of each cut set is comprehensively evaluated, thereby achieving accurate ranking of root causes. Based on the ranking of critical importance, the top K cut sets are selected as the final root cause candidate set, and a verification mechanism is used to confirm them. We introduce an intelligent self-healing and strategy optimization algorithm based on reinforcement learning. Based on the root cause, we automatically generate and execute repair strategies. At the same time, we continuously optimize decisions through online learning, forming a self-evolving operation and maintenance closed loop.

2. The fault root cause dynamic prediction method based on probability propagation reasoning as described in claim 1, characterized in that, The process of establishing globally unique identifiers for each component and obtaining its dependencies includes: The monitored objects cover all logical components of the business system, including application services, data middleware, and computing instances; Each monitored object is assigned a globally unique asset number, which serves as a unique identifier during the fault reasoning process. The system acquires and records the dynamic dependencies between components, forming a continuously updated configuration item library containing unique asset IDs and dependencies.

3. The fault root cause dynamic prediction method based on probability propagation reasoning as described in claim 1, characterized in that, Based on the aforementioned dependencies, node types are defined and a directed graph structure is constructed. Each node is associated with a dynamic fault probability, thus building a fault state representation system, including: Based on a configuration item library with global asset dependencies, a directed graph structure is constructed, where the nodes of the directed graph structure are assets. Nodes are divided into three types: root nodes correspond to infrastructure layer assets in the configuration item library, and their status is directly driven by monitoring data; propagation nodes represent service and middleware logical entities, and have the ability to receive and forward faults; and aggregation nodes represent system-level business faults and are the endpoints of fault propagation. Each node is associated with a dynamic failure probability, which changes over time and is affected by multi-source monitoring data, forming a complete failure state expression system.

4. The fault root cause dynamic prediction method based on probability propagation reasoning as described in claim 1, characterized in that, Based on the dynamic propagation graph model, probability propagation calculations are performed on the graph using unified probability data to extrapolate the failure probability of each component and extract a root cause candidate set, including: Perform graph-based probability propagation reasoning to calculate the real-time failure probability of each node. Graph-based probability propagation reasoning adopts a combination of backward reasoning and forward propagation, taking into account both the bottom-up propagation path of the failure and the top-down impact range of business dependencies. Based on the real-time failure probability of each node, all minimum cut sets corresponding to the current top event are extracted as root cause candidate sets. The minimum cut set refers to the smallest set of basic events that can lead to the occurrence of the top event, and each cut set represents a possible combination of failures.

5. A dynamic prediction system for root causes of failures based on probabilistic propagation reasoning, characterized in that, Specifically, the fault root cause dynamic prediction method based on probability propagation reasoning as described in any one of claims 1-4 includes: The initialization module is used to establish globally unique identifiers for each component and obtain its dependencies. The fault dependency modeling module is used to define node types and construct a directed graph structure based on the dependency relationship, associate each node with a dynamic fault probability, and construct a fault state expression system. Based on the fault state representation system, a dynamic factor reflecting the real-time state of the system is introduced, the edge weights are calculated and dynamically updated, and a dynamic propagation graph model is constructed. The multi-source data alignment module is used to acquire multi-dimensional monitoring data, convert the raw data into standardized fault probabilities through a predefined indicator-probability mapping function, and process the multi-source data using a fusion strategy to obtain aligned unified probability data. The reasoning and root cause localization module is used to perform graph probability propagation calculations based on the dynamic propagation graph model and the unified probability data, to infer the failure probability of each component and extract the root cause candidate set; a key importance weighting model is introduced to perform multi-dimensional quantitative ranking of the root cause candidate set, and the final root cause is predicted and located based on the ranking results. The reinforcement learning optimization module introduces intelligent self-healing and strategy optimization algorithms based on reinforcement learning. It automatically generates and executes repair strategies based on the root cause, and continuously optimizes decisions through online learning, forming a self-evolving operation and maintenance closed loop.

6. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the fault root cause dynamic prediction method based on probability propagation reasoning as described in any one of claims 1-4.

7. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform the fault root cause dynamic prediction method based on probability propagation reasoning as described in any one of claims 1-4.

Images