Operation and maintenance decision-making method and system, terminal and medium

Abstract

The invention belongs to the technical field of operation and maintenance decision, and particularly discloses an operation and maintenance decision method and system, a terminal and a medium. Comprising the following steps: forming structured context representation by acquiring information of a target system; a plurality of natural language candidate operation and maintenance schemes are generated based on a large language model, semantic analysis is performed on scheme texts, action descriptions are mapped into atomic actions, and the atomic actions are arranged into an executable decision flow according to an action dependency relationship. State transition inference is carried out on decision streams through a world model, probability representation of future system states is constructed, comprehensive scoring is carried out on the decision streams based on a multi-target evaluation mechanism, and an executable target strategy is determined. The invention also provides a corresponding system, a terminal and a medium, which can support intelligent and autonomous operation and maintenance decisions in a complex system environment.

Description

Technical Field

[0001] This invention belongs to the field of operation and maintenance decision-making technology, and specifically relates to an operation and maintenance decision-making method, system, terminal and medium. Background Technology

[0002] With the continuous expansion of information infrastructure, cloud computing platforms, distributed microservice architectures, and multi-cluster heterogeneous resources have gradually become the mainstream production environments. The significantly increased complexity of system architecture has driven operations and maintenance activities to evolve from traditional manual script operations and static rule management towards automation and intelligence. To address frequent service anomalies, performance degradation, resource bottlenecks, and configuration drift, the industry has begun to widely adopt data-driven AIOps technology. This technology uses correlation analysis of multi-source operations and maintenance data, such as logs, metrics, topology, and events, to achieve functions such as fault detection, root cause analysis, configuration management, and capacity prediction.

[0003] In existing intelligent operation and maintenance (O&M) technologies, mainstream solutions typically rely on rule bases, templated scripts, or preset strategies to execute corresponding O&M operations upon detecting anomalies. Some technologies utilize machine learning models to predict fault types or resource trends and combine them with O&M script engines to automate processing; others introduce decision trees, reinforcement learning, or graph models to evaluate the effectiveness of multiple solutions before selecting the execution path. Furthermore, large language models are increasingly being used to generate natural language O&M suggestions.

[0004] Existing technologies still have significant shortcomings: First, most systems rely solely on static rules or preset scripts, failing to dynamically generate differentiated solutions based on fault contexts, making it difficult to adapt to changes in complex business environments; second, existing intelligent operation and maintenance technologies generally lack probabilistic extrapolation of system states after the execution of operation and maintenance actions, failing to systematically assess future risks, and lacking the ability to handle cognitive and accidental uncertainties. Summary of the Invention

[0005] This invention addresses the problems in existing technologies by providing an operation and maintenance decision-making method, system, terminal, and medium. It solves the problem that most systems in the aforementioned background technologies rely solely on static rules or preset scripts, failing to dynamically generate differentiated solutions based on fault contexts and thus struggling to adapt to changes in complex business environments. Simultaneously, it addresses the issue that existing intelligent operation and maintenance technologies generally lack probabilistic extrapolation of system states after the execution of operation and maintenance actions, are unable to systematically assess future risks, and lack the ability to handle cognitive and accidental uncertainties.

[0006] The technical solution adopted in this invention is as follows: Firstly, this application provides an operation and maintenance decision-making method, which includes the following steps: The system acquires operational metrics, log data, system topology information, and historical operation and maintenance data of the target system. After structuring the information, it obtains an input representation that includes system state characteristics, dependency characteristics, and contextual semantic characteristics. Perform large language model processing on the input representation to obtain at least one candidate operation and maintenance decision scheme expressed in natural language form; The candidate operation and maintenance decision schemes are processed through semantic parsing and action conversion, converting natural language descriptions into atomic actions, and then orchestrating the atomic actions to obtain an executable decision flow. The executable decision flow is processed by the world model, and uncertainty inference is performed on the state transition of the action sequence to obtain the probability distribution of the future system state. The decision evaluation process is performed based on the probability distribution, and the various decision flows are compared to obtain the target decision flow; The execution target decision flow monitors the system status during execution and updates the large language model and / or world model based on the execution results.

[0007] Furthermore, in the process of performing large language model processing on the input representation to obtain at least one candidate operation and maintenance decision scheme expressed in natural language, the input representation is constructed into prompt words and input into the pre-trained large language model. The large language model generates at least two natural language candidate operation and maintenance decision schemes based on the prompt word content. Each candidate scheme includes core actions, execution prerequisites and expected effects.

[0008] Furthermore, the natural language descriptions of candidate operation and maintenance decision schemes are semantically parsed to extract verb phrases, parameter values, and object identifiers related to operation and maintenance actions. The verb phrases are semantically matched with the action description fields in the predefined atomic action set. The corresponding atomic action is selected based on the matching degree, and the extracted parameter values ​​are filled into the input parameters of the atomic action. Based on the predefined action dependencies in the set of atomic actions, the mapped atomic actions are arranged in order, and an acyclic action execution topology is generated according to the dependency constraints. Each action node in the topology is bound to the corresponding pre-validation rules and rollback actions to construct an executable decision flow.

[0009] Furthermore, the predefined set of atomic actions is organized into an atomic action library. Each atomic action in the atomic action library is encapsulated into a standardized calling unit, which includes at least: an action identifier field for unique identification, an action description field for matching the semantic parsing result, an input parameter list, an output result description, a set of pre-validation rules, an execution logic definition, and a rollback action reference. The pre-validation rule set is used to check the resource quotas, dependent service status and configuration parameter validity of the target system, as well as the current business load status and operation and maintenance window constraints, before orchestrating atomic actions into the decision flow. Rollback operation references are used to call the reverse operation corresponding to the atomic action to restore the system to the state before the atomic action was executed when the corresponding atomic action fails or subsequent monitoring indicates that a rollback is needed.

[0010] Furthermore, the world model is trained based on historical operation and maintenance data and simulated environment data to learn the conditional state transition rules of the system state when performing operation and maintenance actions, and is expressed by a state transition probability function as follows:

[0011] Where S represents the current system state, and A represents the atomic actions already programmed in the decision flow. For the future system state; When performing uncertainty inference, the world model inputs the action sequence into the model, calculates the random uncertainty caused by system noise and the cognitive uncertainty caused by model cognitive bias, and applies both types of uncertainty to the state transition probability function to obtain the probability distribution of the future system state.

[0012] Furthermore, when jointly processing the calculation results of accidental uncertainty and cognitive uncertainty, based on the conditional state transition probability function output by the world model, the variance term obtained from system noise is used as the observation noise component, and the variance term obtained from model cognitive bias is used as the model confidence component. Uncertainty fusion is then performed on the predicted distribution of future states according to the following formula:

[0013] in, Variance represents random uncertainty. This represents the variance of cognitive uncertainty; The fused variance is used to construct the prediction confidence interval for the future system state, serving as a parameter of the probability distribution.

[0014] Furthermore, based on the probabilistic representation of the future system state obtained in step S4, a multi-objective evaluation function is constructed that includes business impact, recovery time, resource consumption cost, and risk confidence interval. The multi-objective evaluation function uses the weighted combination value of each indicator as the comprehensive score of the decision flow. During the scoring process, the state prediction confidence interval after incorporating uncertainty is used as a constraint. When the deviation between the prediction confidence interval corresponding to the decision flow and the historical statistical interval exceeds a preset threshold, the weight of the decision flow is reduced or its execution qualification is eliminated. Finally, the target decision flow is selected based on the comprehensive score.

[0015] Secondly, this application provides an operation and maintenance decision-making system for implementing the operation and maintenance decision-making method as described in the first aspect, the system comprising: The decision flow generation engine is used to generate natural language candidate operation and maintenance solutions based on root cause information, topology information and indicator data input from the fault context. After conceiving candidate solutions through a large language model, the decision flow compiler is called to perform semantic parsing, action transformation and dependency arrangement of the candidate solutions, forming an executable decision flow organized in a directed acyclic structure. The inference and quantification engine is used to input executable decision flows into the Bayesian world model, infer the uncertainty of future system states based on state transition probabilities, and perform multiple sampling inferences on action sequences based on Monte Carlo simulation to generate risk quantification reports for each candidate decision flow. The autonomous decision-making and closed-loop engine is used to comprehensively score the risk quantification reports of each candidate decision flow based on a multi-objective optimization function, select the preferred strategy and execute the strategy. During the execution process, the system status is monitored and compared with the prediction confidence interval. When the deviation exceeds the threshold, the automatic rollback logic is triggered, and the deviation information is recorded for updating the Bayesian world model and / or the large language model for generating candidate solutions.

[0016] Thirdly, this application provides a terminal, including: The memory is used to store operation and maintenance decision-making programs; A processor is used to implement the steps of the operation and maintenance decision-making method as described in the first aspect when executing the operation and maintenance decision-making device.

[0017] Fourthly, this application provides a computer-readable storage medium that stores computer instructions. When a computer reads the computer instructions in the storage medium, the computer executes the operation and maintenance decision-making method as described in the first aspect.

[0018] As can be seen from the above technical solutions, the advantages of the present invention are: By constructing a complete autonomous operation and maintenance decision-making process from data acquisition, solution generation, action compilation, risk inference to strategy execution and model update, the system can form an overall closed loop with understanding, inference and adaptive capabilities in complex operation and maintenance scenarios, thereby improving the automation and reliability of operation and maintenance decisions.

[0019] By leveraging large language models to generate natural language candidate operation and maintenance solutions, the system can automatically form diverse processing ideas based on context, significantly reducing reliance on human experience and static rules, and making strategy generation more flexible and closer to real-world scenarios.

[0020] By using semantic parsing and action mapping, natural language content is transformed into atomic actions, and an executable decision flow is constructed based on dependencies. This enables the system to automatically generate structured and schedulable operation chains from unstructured text, ensuring correct execution order and clear action boundaries, thus achieving automatic penetration from natural language to the execution level.

[0021] The structure of the atomic action library is refined by introducing action identifiers, description fields, input parameters, pre-validation and rollback references, etc., so that the action library has a high degree of standardization and composability, thereby maintaining stable availability, scalability and maintainability in large-scale complex operation and maintenance actions.

[0022] By introducing a world model trained on historical and simulated data and establishing state transition relationships, the system can make probabilistic predictions of the future impact of operational actions. At the same time, by integrating accidental uncertainty and cognitive uncertainty, the prediction results are closer to the real environment, providing a reliable risk quantification basis for subsequent strategy selection.

[0023] By integrating uncertainties, the variance and confidence interval of the future system state become more accurate, enabling the system to clearly define the risk boundaries of action sequences and improve prediction robustness, thereby giving strategy selection higher risk identification capabilities and reliability.

[0024] Based on a multi-objective evaluation mechanism that comprehensively considers recovery time, business impact, resource consumption, and risk confidence interval, the prediction uncertainty is used as a decision constraint to achieve scientific ranking of multiple candidate decision flows. The dynamic elimination mechanism of risk interval deviation ensures that the selected target strategy has higher stability and executability, making the overall autonomous operation and maintenance decision-making process more secure, controllable, and optimized. Attached Figure Description

[0025] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart illustrating the steps of the operation and maintenance decision-making method in the embodiment; Figure 2 This is a framework diagram of the operation and maintenance decision-making system in the embodiment. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Please see Figure 1 As shown, the present invention provides an operation and maintenance decision-making method, including the following steps: Step S1: Obtain the target system's operating metrics, log data, system topology information, and historical operation and maintenance data. After structuring the information, obtain an input representation that includes system state features, dependency features, and contextual semantic features. In some embodiments, operational metrics can be obtained through a monitoring system's self-collection plugin, periodically pulling real-time metrics such as CPU utilization, memory usage, container load, and interface latency. Log data can come from a centralized log service, extracting timestamps, exception levels, request identifiers, and error stack traces through key fields. System topology information can be provided by a service governance platform, including service dependencies, upstream call chains, and deployment node distribution. Historical operational data comes from a change log system, containing historical fault causes, handling steps, and handling results. These multi-source data are uniformly transformed into structured objects, and feature extraction algorithms are used to generate state feature vectors, dependency chain features, and semantic expressions describing the context of the current fault, enabling subsequent models to receive input in a unified format.

[0029] In one specific implementation, when a microservice reports an anomaly of excessively high response latency, the system retrieves the metric sequence of the most recent five minutes from the monitoring platform, extracts the log fragments showing the anomaly from the log system, and determines the upstream call pressure by combining this with topology information. Subsequently, key sentences in the logs are converted into semantic vectors, and metric trends and topology paths are encoded into feature vectors, ultimately forming a large language model input representation for subsequent steps.

[0030] Step S2: Perform large language model processing on the input representation to obtain at least one candidate operation and maintenance decision scheme expressed in natural language form; In some embodiments, the model uses the input representation obtained in step S1 as prompt words, which include fault phenomena, contextual indicator trends, changes in dependent services, and historical system processing experience. The large language model loads task-tuned weights, enabling it to generate structured operation and maintenance solution text based on the input context. The text content includes suggested processing approaches, descriptions of actions to be performed, and necessary risk warnings. The number of candidate solutions can be dynamically adjusted according to the model output length or confidence level.

[0031] In one embodiment, in the event of an exception where a database connection pool is exhausted, the natural language solution generated by the large language model may include suggestions such as "check the current concurrency of the connection pool", "clean up connections that have been occupied for a long time", "reload the database connection configuration", and "adjust the configuration parameters according to the business peak", forming a natural language solution text that can be parsed in the next step.

[0032] Step S3: After semantic parsing and action conversion processing of the candidate operation and maintenance decision schemes, the natural language description is converted into atomic actions, and the atomic actions are arranged to obtain an executable decision flow; In some embodiments, the semantic parsing process includes identifying verb phrases, parameter values, and operation objects in the candidate solution text, and mapping them to corresponding actions in a predefined set of atomic actions using a semantic matching algorithm. The set of atomic actions consists of standardized operation units, each containing an action identifier, input parameters, pre-validation content, and a corresponding rollback operation. The action orchestration process organizes actions into an executable directed acyclic structure based on the dependencies between actions, ensuring that the execution order conforms to the actual constraints of the system.

[0033] In one embodiment, when the natural language scheme includes "restart service A and clear the cache", the system first identifies the atomic action RestartService corresponding to "restart service" and binds "service A" as a parameter. Then, it identifies the ClearCache action corresponding to clear the cache. According to the dependencies defined in the atomic action library, the cache needs to be cleared before the service is restarted. Therefore, the final orchestration forms an ordered action flow for execution in subsequent steps.

[0034] Step S4: The executable decision flow is processed by the world model, and uncertainty inference is performed on the state transition of the action sequence to obtain the probability distribution of the future system state. In some embodiments, the world model is trained based on historical operational scenarios. The model simulates the current system state and pending actions, estimating possible future state changes through internal parameters. The model calculates fluctuations caused by system noise and uncertainties due to insufficient model knowledge, applying both types of uncertainty to the prediction process to generate a probabilistic representation of the future system state. The model can progressively simulate the impact of each action in the behavioral chain, ensuring that the generated probability distribution covers indicators such as recovery time, business continuity impact, and resource fluctuation range.

[0035] In one embodiment, if the decision flow includes two actions: "expanding instance capacity" and "restarting the service," the world model simulates the impact of the expansion action on system load and response latency, while also simulating the potential short-term unavailability risk caused by the restart action. Based on uncertainty inference, the model outputs the probability range of the system state within the next five minutes, enabling subsequent decisions to be evaluated based on the risk quantification results.

[0036] Step S5: Perform decision evaluation processing on the probability distribution, compare each decision flow, and obtain the target decision flow; In some embodiments, the system constructs a comprehensive evaluation function that includes multiple performance and risk dimensions. It takes the recovery time, business impact indicators, resource cost estimates, and risk confidence intervals from the probability distribution generated in step S4 as inputs to score each candidate decision flow. During the scoring process, the confidence intervals of the prediction results can be verified. When the prediction deviation of a decision flow is large, that decision flow will be downweighted or eliminated to ensure that the subsequent executed strategies have high stability.

[0037] In one embodiment, when two candidate decision flows offer two options, "restart service" and "expand instance", respectively, the scoring function evaluates the risk range and recovery time of each option. If "restart service" is predicted to have a large unavailability period, while "expand instance" has a lower risk, the system will select the expansion option as the target decision flow.

[0038] Step S6: Execute the target decision flow, monitor the system status during execution, and update the large language model and / or world model based on the execution results; In some embodiments, the system executes operations sequentially according to the order of actions in the decision flow, while simultaneously collecting execution feedback data in real time, including operation time, error messages, actual performance changes, and deviations from predicted values. When the deviation exceeds a preset threshold, the system triggers rollback logic and restores the current state to the checkpoint before the action execution. After execution, the system compares the collected feedback data with the predicted values ​​to form incremental samples for model training, thereby updating the large language model used for solution generation or the world model used for state inference, improving the accuracy of subsequent decisions.

[0039] In one embodiment, after the system performs the "expansion instance" action, if it finds that the actual resource usage change deviates significantly from the prediction, it records the deviation information into the sample set and uses it to adjust the parameters of the world model during the next model training, so that it more accurately reflects the system's true response to the expansion action.

[0040] In some embodiments, in step S2, the input representation is constructed into prompt words and input into a pre-trained large language model. The large language model generates at least two natural language candidate operation and maintenance decision schemes based on the prompt word content. Each candidate scheme includes a core action, execution prerequisites and expected effects.

[0041] In practical implementation, prompts can be constructed using templates to include the current fault phenomenon, historical information on similar faults, trend of indicator changes, and fragments of the topology call chain, enabling the large language model to understand the current operational context of the system. After loading the pre-trained weights, the model automatically generates multiple candidate solutions based on the prompts, each providing a different processing path for the same anomaly. For example, for the fault of "increased database response time," the model might generate solution A, which includes checking the database connection pool usage, cleaning up long-running queries, and indicating that database management privileges are required for execution; solution B might propose scaling up instances or reallocating resources, explaining the potential performance improvements from scaling up in the expected results. The generation of multiple solutions allows the system to filter and compare strategies in subsequent steps.

[0042] In some embodiments, step S3 includes: Semantic parsing is performed on the natural language descriptions of candidate operation and maintenance decision schemes to extract verb phrases, parameter values ​​and object identifiers related to operation and maintenance actions. The verb phrases are semantically matched with the action description fields in the predefined atomic action set. The corresponding atomic action is selected according to the matching degree, and the extracted parameter values ​​are filled into the input parameters of the atomic action. Based on the predefined action dependencies in the set of atomic actions, the mapped atomic actions are arranged in order, and an acyclic action execution topology is generated according to the dependency constraints. Each action node in the topology is bound to the corresponding pre-validation rules and rollback actions to construct an executable decision flow.

[0043] In actual operation, the system first performs part-of-speech analysis and semantic segmentation on the natural language scheme to identify verb phrases that indicate operational intentions, such as "restart service," "upgrade components," and "clear cache." Then, the system extracts parameter values ​​from the text, such as service names, resource quantities, or configuration item names. The semantic matching process generally uses similarity calculation, comparing the identified verb phrases with action descriptions in the atomic action set and selecting the closest action type. Each identified action is converted into a specific executable unit, and the action's input fields are completed based on the extracted parameter values.

[0044] During orchestration, the system automatically generates a loop-free execution topology based on dependencies defined in the atomic action library. For example, a configuration modification operation must be completed before a service restart, or resource cleanup needs to be performed before any scaling action. In a typical embodiment, the natural language description of "clear the cache and then restart service A" is parsed into two atomic actions, and an ordered action path of "clear the cache first, then restart service A" is generated based on dependencies. Each action node is bound to action-related checks, such as "check if service A is currently restartable," and rollback behavior in case of failure, thus forming a complete decision flow that can be run in subsequent steps.

[0045] In some embodiments, the set of atomic actions predefined in step S3 is organized into an atomic action library, and each atomic action in the atomic action library is encapsulated as a standardized calling unit, which includes at least: an action identifier field for unique identification, an action description field for matching the semantic parsing result, an input parameter list, an output result description, a set of pre-validation rules, an execution logic definition, and a rollback action reference; The pre-validation rule set is used to check the resource quotas, dependent service status and configuration parameter validity of the target system, as well as the current business load status and operation and maintenance window constraints, before orchestrating atomic actions into the decision flow. Rollback operation references are used to call the reverse operation corresponding to the atomic action to restore the system to the state before the atomic action was executed when the corresponding atomic action fails or subsequent monitoring indicates that a rollback is needed.

[0046] In its implementation, the atomic action library consists of multiple standardized action units, stored in a unified format for easy compiler access. The action description field assists in semantic matching, ensuring accurate correspondence between natural language text and entries in the action library. The input parameter list describes all parameters required to execute the action, such as service name, node name, or configuration item values. The output description indicates potential state changes after action execution, such as "service enters restart state" or "cache cleared."

[0047] The pre-verification rule set can be implemented through scripts or lightweight detection logic to verify whether the current system state meets the conditions for action execution before the action is performed. For example, before modifying the configuration, it is necessary to check whether the configuration file exists; before restarting the service, it is necessary to confirm that the service is not currently running in high-load mode. Business load status checks are used to avoid performing operations that would cause service interruptions during peak business periods. In a specific embodiment, when an operation requires additional resources (such as scaling up instances), the pre-verification rules will determine whether the current system resource quota is sufficient.

[0048] Rollback operations reference the corresponding reverse operation, such as performing a version rollback when an upgrade fails, or reclaiming requested resources after a resource expansion failure. In actual execution, if an atomic action returns to a failed state, or if the system detects a deviation from the expected state after execution, the system will automatically execute the rollback operation to restore the system to a safe state and ensure the stability of the overall operation and maintenance process.

[0049] In some embodiments, in step S4, the world model is trained based on historical operation and maintenance data and simulated environment data to learn the conditional state transition rules of the system state when performing operation and maintenance actions, and is expressed as a state transition probability function as follows:

[0050] Where S represents the current system state, and A represents the atomic actions already programmed in the decision flow. For the future system state; When performing uncertainty inference, the world model inputs the action sequence into the model, calculates the random uncertainty caused by system noise and the cognitive uncertainty caused by model cognitive bias, and applies both types of uncertainty to the state transition probability function to obtain the probability distribution of the future system state.

[0051] In practical deployments, world models are typically trained by collecting a large number of historical operational scenarios, including system state changes before and after different operational actions such as service restarts, configuration updates, instance scaling, and cache cleanup. Simultaneously, test data generated from simulated environments is used to supplement and cover extreme scenarios. During training, the model learns the patterns of how actions affect the system state, enabling it to predict future system performance changes and potential risks based on the current state and action sequences.

[0052] When performing uncertainty inference, the system inputs the decision flow generated in step S3 into the model item by item. The model provides a predicted distribution of the possible state changes that each action may produce in the current state. Random uncertainty arises from random noise during system operation, such as fluctuations in request volume and network jitter; cognitive uncertainty arises from the model's lack of experience with certain scenarios. The system estimates the impact of both types of uncertainty separately and superimposes them onto the prediction result, ensuring that the output future state distribution reflects the risk range in the real-world environment.

[0053] For example, when the decision flow includes the operation of "clearing the cache and restarting the service," the world model simulates the impact of cache clearing on metrics such as hit rate and latency, and infers the risk of short-term unavailability that may result from service restart. The final prediction is presented as a range, such as a latency interval or resource utilization interval for the next few minutes, thus providing a quantitative basis for the strategy scoring in the next step.

[0054] In some embodiments, in step S4, when jointly processing the calculation results of random uncertainty and cognitive uncertainty, the variance term obtained from system noise is used as the observation noise component, and the variance term obtained from model cognitive bias is used as the model confidence component, based on the conditional state transition probability function output by the world model. Uncertainty fusion is then performed on the predicted distribution of future states according to the following formula:

[0055] in, Variance represents random uncertainty. This represents the variance of cognitive uncertainty; The fused variance is used to construct the prediction confidence interval for the future system state, serving as a parameter of the probability distribution.

[0056] In its implementation, the system first calculates two types of variance terms for the predicted output of each action. The variance term generated by system noise mainly comes from the real-time index fluctuation model, while the variance term caused by model cognitive bias can be dynamically estimated by comparing the prediction results with historical sample deviations. After merging the two types of variance terms, the system can form a more reliable prediction confidence interval to reasonably reflect the potential fluctuation range after the action chain is executed.

[0057] In one embodiment, after predicting "restarting service A," the system may determine the range of latency increase. If the perceived uncertainty is high (e.g., the service rarely experiences similar scenarios), the system will automatically widen the confidence interval, indicating that the prediction is not certain enough; conversely, if the system noise is low, the confidence interval may be narrower. This interval then serves as an important basis for strategy selection, helping the system identify higher-risk options and reduce their priority in the next step.

[0058] In some embodiments, in step S5, a multi-objective evaluation function is constructed based on the probabilistic representation of the future system state obtained in step S4, which includes business impact, recovery time, resource consumption cost and risk confidence interval. The multi-objective evaluation function uses the weighted combination value of each indicator as the comprehensive score of the decision flow. During the scoring process, the state prediction confidence interval after incorporating uncertainty is used as a constraint. When the deviation between the prediction confidence interval corresponding to the decision flow and the historical statistical interval exceeds a preset threshold, the weight of the decision flow is reduced or its execution qualification is eliminated. Finally, the target decision flow is selected based on the comprehensive score.

[0059] The multi-objective evaluation function allows system administrators to preset weights to align with varying business priorities regarding recovery speed, reliability, and resource consumption. During scoring, the system inputs the predicted recovery time, potential business impact, resource usage, and risk confidence interval into the evaluation function, calculating a comprehensive score for each decision stream. The confidence interval serves as an additional safety factor; if the uncertainty of a predicted solution is too high, even if its average recovery performance is good, its weight will be reduced or it will be completely eliminated to avoid implementing high-risk strategies.

[0060] For example, if candidate solution A has a fast recovery speed but a wide prediction confidence interval, indicating that the execution result may deviate significantly from expectations, while solution B has a slightly slower recovery time but a more controllable risk range, the system will prioritize solution B as the target decision flow. In this way, the system can screen out an operation and maintenance strategy that balances efficiency and stability from multiple solutions.

[0061] In some embodiments, please refer to Figure 2 As shown, this application provides an operation and maintenance decision-making system for implementing the operation and maintenance decision-making method. The system includes: The decision flow generation engine is used to generate natural language candidate operation and maintenance solutions based on root cause information, topology information and indicator data input from the fault context. After conceiving candidate solutions through a large language model, the decision flow compiler is called to perform semantic parsing, action transformation and dependency arrangement of the candidate solutions, forming an executable decision flow organized in a directed acyclic structure. The inference and quantification engine is used to input executable decision flows into the Bayesian world model, infer the uncertainty of future system states based on state transition probabilities, and perform multiple sampling inferences on action sequences based on Monte Carlo simulation to generate risk quantification reports for each candidate decision flow. The autonomous decision-making and closed-loop engine is used to comprehensively score the risk quantification reports of each candidate decision flow based on a multi-objective optimization function, select the preferred strategy and execute the strategy. During the execution process, the system status is monitored and compared with the prediction confidence interval. When the deviation exceeds the threshold, the automatic rollback logic is triggered, and the deviation information is recorded for updating the Bayesian world model and / or the large language model for generating candidate solutions.

[0062] In some embodiments, this application provides a terminal, including: The memory is used to store operation and maintenance decision-making programs; A processor is used to execute the steps of the operation and maintenance decision-making system to implement the operation and maintenance decision-making method.

[0063] In some embodiments, this application provides a computer-readable storage medium that stores computer instructions. When a computer reads the computer instructions in the storage medium, the computer executes the operation and maintenance decision-making method.

[0064] The above description is merely a preferred embodiment of one or more embodiments of this specification and is not intended to limit the scope of one or more embodiments of this specification. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments of this specification should be included within the protection scope of one or more embodiments of this specification.

Claims

1. An operation and maintenance decision-making method, characterized in that, Includes the following steps: The system acquires operational metrics, log data, system topology information, and historical operation and maintenance data of the target system. After structuring the information, it obtains an input representation that includes system state characteristics, dependency characteristics, and contextual semantic characteristics. Perform large language model processing on the input representation to obtain at least one candidate operation and maintenance decision scheme expressed in natural language form; The candidate operation and maintenance decision schemes are processed through semantic parsing and action conversion, converting natural language descriptions into atomic actions, and then orchestrating the atomic actions to obtain an executable decision flow. The executable decision flow is processed by the world model, and uncertainty inference is performed on the state transition of the action sequence to obtain the probability distribution of the future system state. The decision evaluation process is performed based on the probability distribution, and the various decision flows are compared to obtain the target decision flow; The execution target decision flow monitors the system status during execution and updates the large language model and / or world model based on the execution results.

2. The operation and maintenance decision-making method according to claim 1, characterized in that, In the process of performing large language model processing on the input representation to obtain at least one candidate operation and maintenance decision scheme expressed in natural language, the input representation is constructed into prompt words and input into a pre-trained large language model. The large language model generates at least two natural language candidate operation and maintenance decision schemes based on the prompt word content. Each candidate scheme includes core actions, execution prerequisites and expected effects.

3. The operation and maintenance decision-making method according to claim 2, characterized in that, Semantic parsing is performed on the natural language descriptions of candidate operation and maintenance decision schemes to extract verb phrases, parameter values ​​and object identifiers related to operation and maintenance actions. The verb phrases are semantically matched with the action description fields in the predefined atomic action set. The corresponding atomic action is selected according to the matching degree, and the extracted parameter values ​​are filled into the input parameters of the atomic action. Based on the predefined action dependencies in the set of atomic actions, the mapped atomic actions are arranged in order, and an acyclic action execution topology is generated according to the dependency constraints. Each action node in the topology is bound to the corresponding pre-validation rules and rollback actions to construct an executable decision flow.

4. The operation and maintenance decision-making method according to claim 3, characterized in that, A predefined set of atomic actions is organized into an atomic action library. Each atomic action in the atomic action library is encapsulated as a standardized calling unit, which includes at least: an action identifier field for unique identification, an action description field for matching the semantic parsing result, an input parameter list, an output result description, a set of pre-validation rules, an execution logic definition, and a rollback reference. The pre-validation rule set is used to check the resource quotas, dependent service status and configuration parameter validity of the target system, as well as the current business load status and operation and maintenance window constraints, before orchestrating atomic actions into the decision flow. Rollback operation references are used to call the reverse operation corresponding to the atomic action to restore the system to the state before the atomic action was executed when the corresponding atomic action fails or subsequent monitoring indicates that a rollback is needed.

5. The operation and maintenance decision-making method according to any one of claims 1-4, characterized in that, The world model is trained based on historical operational data and simulated environment data to learn the conditional state transition rules of the system state when performing operational actions, and is expressed as a state transition probability function: Where S represents the current system state, and A represents the atomic actions already programmed in the decision flow. For the future system state; When performing uncertainty inference, the world model inputs the action sequence into the model, calculates the random uncertainty caused by system noise and the cognitive uncertainty caused by model cognitive bias, and applies both types of uncertainty to the state transition probability function to obtain the probability distribution of the future system state.

6. The operation and maintenance decision-making method according to claim 5, characterized in that, When jointly processing the calculation results of accidental uncertainty and cognitive uncertainty, the conditional state transition probability function output by the world model is used as the basis. The variance term obtained from system noise is used as the observation noise component, and the variance term obtained from model cognitive bias is used as the model confidence component. Uncertainty fusion is then performed on the predicted distribution of future states according to the following formula: in, Variance represents random uncertainty. This represents the variance of cognitive uncertainty; The fused variance is used to construct the prediction confidence interval for the future system state, serving as a parameter of the probability distribution.

7. The operation and maintenance decision-making method according to claim 6, characterized in that, Based on the probabilistic representation of the future system state obtained in step S4, a multi-objective evaluation function is constructed, which includes business impact, recovery time, resource consumption cost, and risk confidence interval. The multi-objective evaluation function uses the weighted combination value of each indicator as the comprehensive score of the decision flow. During the scoring process, the state prediction confidence interval after incorporating uncertainty is used as a constraint. When the deviation between the prediction confidence interval corresponding to the decision flow and the historical statistical interval exceeds a preset threshold, the weight of the decision flow is reduced or its execution qualification is eliminated. Finally, the target decision flow is selected based on the comprehensive score.

8. An operation and maintenance decision-making system for implementing the operation and maintenance decision-making method as described in claim 1, characterized in that, The system includes: The decision flow generation engine is used to generate natural language candidate operation and maintenance solutions based on root cause information, topology information and indicator data input from the fault context. After conceiving candidate solutions through a large language model, the decision flow compiler is called to perform semantic parsing, action transformation and dependency arrangement of the candidate solutions, forming an executable decision flow organized in a directed acyclic structure. The inference and quantification engine is used to input executable decision flows into the Bayesian world model, infer the uncertainty of future system states based on state transition probabilities, and perform multiple sampling inferences on action sequences based on Monte Carlo simulation to generate risk quantification reports for each candidate decision flow. The autonomous decision-making and closed-loop engine is used to comprehensively score the risk quantification reports of each candidate decision flow based on a multi-objective optimization function, select the preferred strategy and execute the strategy. During the execution process, the system status is monitored and compared with the prediction confidence interval. When the deviation exceeds the threshold, the automatic rollback logic is triggered, and the deviation information is recorded for updating the Bayesian world model and / or the large language model for generating candidate solutions.

9. A terminal, characterized in that, include: The memory is used to store operation and maintenance decision-making programs; A processor is used to implement the steps of the operation and maintenance decision-making method as described in claim 1 when executing the operation and maintenance decision-making device.

10. A computer-readable storage medium, characterized in that, The storage medium stores computer instructions. When the computer reads the computer instructions from the storage medium, the computer executes the operation and maintenance decision-making method as described in claim 1.

Images