Microservice Root Cause Analysis Method Based on Evidence Extraction and Large Language Model Inference
By constraining the reasoning process of large language models through declarative adapters and restricted tools, the instability and system specificity of existing microservice root cause analysis methods in heterogeneous systems are solved, and stable localization and attribution capabilities across systems are achieved.
Patent Information
- Application Number
- CN202611318876.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-28
- Publication Date
- 2026-11-17
AI Technical Summary
Existing microservice root cause analysis methods lack structured constraints on fault propagation effects, suffer from severe system-specific coupling, mismatch between data processing methods and model characteristics, and unstable exploration paths due to task coupling, making it difficult to achieve zero-code migration and cross-system generalization in heterogeneous systems.
By isolating system-specific knowledge through declarative adapters, constraining the reasoning process of large language models with structured evidence cards and restricted read-only tools, separating deterministic anomaly detection and candidate localization, and combining morphological anomaly detection and topological roll-up, evidence cards are generated and reasoning is performed using a restricted toolset.
It improves the stability and portability of microservice root cause analysis, enabling adaptation between different microservice systems, enhancing the ability to locate and attribute causes in complex fault propagation scenarios, and reducing issues such as field misreading and unstable running paths.
Smart Images

Figure CN123114822A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent operation and maintenance and microservice system fault diagnosis technology, specifically involving a microservice root cause analysis method based on evidence extraction and large language model reasoning. Background Technology
[0002] Microservice systems typically consist of numerous services, instances, containers, or physical nodes, with complex call dependencies and deployment topologies between different components. When a system experiences failures such as decreased throughput, slowed response, or increased error rates, operations personnel need to use root cause analysis (RCA) to pinpoint the time, component, and cause of the failure. Microservice root cause analysis heavily relies on multimodal observability data, including resource metrics, call chains, logs, and business health signals. Due to the large scale, significant format differences, and varying sampling granularities of this data, and because failure signals propagate along call chains and deployment topologies, root cause analysis usually needs to simultaneously address three issues: anomaly detection, candidate localization, and semantic attribution.
[0003] Traditional root cause analysis methods, such as those in the literature [Li Z, Chen J, Jiao R, et al. Practical root cause localization for microservice systems via trace analysis[C]. ACM 29th International Symposium on Quality of Service. IEEE, 2021: 1-10], typically rely on pre-defined topologies, system-specific models, manual rules, or training data. When migrating to new microservice systems, they require remodeling or parameter tuning, and lack cross-system generalization capabilities. In recent years, with the development of Large Language Models (LLMs), methods have emerged that allow LLMs to directly access raw telemetry data as agents, freely exploring metrics, call chains, and logs by generating and executing code, ultimately outputting root causes, such as those in the literature [Xu J, Zhang Q, Zhong Z, et al. OpenRCA: Can large language models locate the root cause of software failures[C]. International Conference on Learning Representations. 2025: 84562-84590]. However, these methods based on free exploration using large language models face the following significant technical drawbacks in practical applications: 1. Lack of structured constraints on the propagation effect of faults: Microservice faults have complex cascading propagation effects. The strongest abnormal signals often appear in downstream victim components (such as avalanche-like timeout errors) rather than the root cause component. Existing agent methods lack topological constraints during free exploration and are easily misled by high-amplitude anomalies in the downstream, misjudging the victim component as the root cause.
[0004] 2. Severe system-specific coupling and lack of generalization ability: The root cause analysis logic or prompt words of existing methods are highly coupled with the topology and indicator naming of specific systems. When facing heterogeneous enterprise-level microservice systems (such as some systems lacking log modalities or different component hierarchical structures), it is necessary to rewrite the code or redesign the reasoning logic, lacking zero-code migration and cross-system generalization ability.
[0005] 3. Mismatch between data processing methods and model characteristics: Raw telemetry data are mostly high-frequency numerical signals and machine-generated text, lacking semantic abstraction. Directly handing them over to large language models for processing or having them generate code for processing can easily lead to context bloat, field misreading, illusion generation, and extremely high computational costs.
[0006] 4. Task coupling leads to unstable exploration paths: Large language models simultaneously undertake evidence search, anomaly detection, and semantic attribution, resulting in unstable exploration paths and making it difficult to distinguish between "failure to find evidence" and "incorrect reasoning in the face of existing evidence," thus reducing the accuracy and reproducibility of diagnosis. Summary of the Invention
[0007] In view of the above, the present invention provides a microservice root cause analysis method based on evidence extraction and large language model reasoning. This method separates the anomaly detection and candidate localization processes, which are suitable for deterministic processing, from the large language model, isolates system-specific knowledge through a declarative adapter, and constrains the reasoning process of the large language model using structured evidence cards and restricted read-only tools, thereby improving the stability, transferability, and interpretability of microservice root cause analysis.
[0008] A microservice root cause analysis method based on evidence extraction and large language model reasoning includes the following steps: (1) Build a declarative adapter for the microservice system to be diagnosed, and perform structural consistency verification on it after construction; the declarative adapter is a set of structured configurations of the microservice system; (2) Receive diagnostic queries input by users and read telemetry data from the microservice system to calculate the daily robust statistical baseline; (3) Based on the daily robust statistical baseline, the daily time series of each component KPI (key performance indicator) is divided into four forms, and form-aware multimodal anomaly detection is performed; (4) Convert the detected multimodal anomaly results into a unified evidence card, and use the deployment topology to perform roll-up correction operation on the evidence card; (5) Encapsulate the evidence cards and auxiliary summary information into a predefined read-only toolset and dynamically expose it to the large language model according to the modality declared by the adapter; (6) The large language model receives diagnostic queries and read-only toolsets, and infers and outputs root cause triples based on restricted tools and diagnostic rules.
[0009] Furthermore, the microservice system in step (1) is a software system composed of multiple services, instances, and nodes that collaboratively complete business functions through network calls; the structured configuration set corresponding to the declarative adapter is denoted as... A and ,in This is a telemetry data specification used to describe the file location, field names, and time units of metrics, call chains, logs, and business health signals; This is a candidate set used to describe the component hierarchy, candidate components, and candidate failure causes. This is a mapping relationship used to describe the normalization of the original component identifier, the mapping from KPI name to resource type, and the mapping from resource type to coarse-grained reason candidates; For deployment topology, it is used to describe the ownership or co-location relationships between instances, services, and nodes.
[0010] Further, the diagnostic query in step (2) is represented as ,in Indicates the time window to be analyzed. This represents the set of target fields that need to be answered. This indicates the expected number of failures; then, based on the specification in the declarative adapter, it reads the microservice system's performance within the time window. The telemetry data within the system refers to observable data generated during the operation of the microservice system, including indicator data, call chain data, log data, and business health signals.
[0011] Furthermore, in step (2), for any component c (Service, instance, or node), read the component c The daily time series of each KPI, and the calculation of the robust statistical baseline for the whole day based on this, are expressed as follows: in: For components c The k A robust daily statistical baseline for each KPI. These correspond to the mean, standard deviation, minimum extreme value, maximum extreme value, and quantile of a daily time series, respectively.
[0012] Furthermore, the specific implementation of step (3) is as follows: like If so, the daily time series of KPIs is determined to be a dead series and discarded. This is a preset minimum threshold; If the difference between the high quantile and the low quantile of the sequence is less than a preset threshold If the daily time series of the KPI is determined to be a saturated series, it will be marked as an indicator blind zone, and anomaly detection will be forced to rely on logs or call chain modalities, i.e., the time windows will be counted separately. Sudden increases in log frequency and call chain latency escalation events within the system; like If the daily time series of KPIs is deemed to be a low baseline series, then anomaly detection requires that the relative deviation threshold be met simultaneously. With absolute amplitude threshold That is, any observation in the sequence relative to The deviation is greater than And this observed value is consistent with The absolute difference is greater than In such cases, the sequence is determined to be abnormal. This is a preset minimum threshold; All sequences other than the three types mentioned above are classified as normal sequences. Robust statistical banding based on interquartile range (IQR) or median absolute deviation (MAD) is used to detect abnormal segments in the sequence that deviate from the baseline by more than a preset multiple.
[0013] Furthermore, the evidence card in step (4) is a structured record generated during the deterministic evidence extraction stage, used to describe an abnormal combination of component and resource types. The evidence card is recorded as follows: ,in Represents a component. Indicates component hierarchy, Indicates the exception type. Indicates the start time of the exception. This indicates the supporting evidence set. Indicates the modality of evidence source. This represents the coarse-grained candidate set of causes.
[0014] Furthermore, the roll-up correction operation on the evidence card in step (4) includes the following three cases: Vertical Co-location Roll-Up: Setting up physical nodes Its managed collection of containers If there is a co-location relationship, At least one ( , (For a preset threshold number of containers) in time difference If an anomaly of the same resource type occurs within the same physical node, then that physical node will be generated. Level of evidence cards The original container-level card will be retained as supporting evidence. Horizontal Services Volume 1: Setting up Microservices and the collection of instances it contains There is a relationship of ownership, if If multiple instances of the same resource type experience anomalies within a similar timeframe, a service-level evidence card will be generated by aggregation. ; Subthreshold fallback: For scores that do not meet the strict outlier threshold but have soft deviations Higher sequences generate weak cue sets , and according to Arranged in descending order; when strictly using a set of evidence cards When empty, As a fallback input in the reasoning stage.
[0015] Furthermore, the read-only toolset in step (5) includes an evidence list tool, a component viewing tool, a KPI sequence viewing tool, and an answer submission tool. When the microservice system has the corresponding modality, it also includes a call graph tool, a log summary tool, and a business health signal tool.
[0016] Furthermore, in step (6), the large language model infers based on restricted tools and diagnostic rules, and strictly follows the following restricted diagnostic rules to determine the root cause: Relative amplitude priority rule: When comparing multiple candidate components, the relative deviation amplitude of the anomaly is obtained through the component viewing tool. The component with the largest specific deviation amplitude is identified as the root cause to distinguish the root cause from cascaded victim components. Call chain reverse tracing rules: When multiple components are abnormal, the call topology is obtained through the call graph tool. Starting from the downstream victim node with the highest latency or failure rate, the call chain is traced back along the call edge to lock the upstream component with evidence of underlying resource abnormality as the root cause. Golden signal time alignment rule: Obtain the start time of system-level faults through business health signal tools. Using this as a time anchor, filter out events that occurred at a time that are not related to the time of occurrence. Isolated evidence cards that do not match and cannot be explained by propagation delay.
[0017] Furthermore, in step (6), after the large language model completes the root cause identification, it outputs the root cause triples through the answer submission tool. ,in Indicates the time when the fault occurred. Indicates a faulty component. Indicate the cause of the fault; verification and It will then output the validated root cause triplet in a formatted manner.
[0018] A computer device includes a memory and a processor, wherein the memory stores a computer program and the processor executes the computer program to implement the aforementioned microservice root cause analysis method based on evidence extraction and large language model reasoning.
[0019] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned microservice root cause analysis method based on evidence extraction and large language model reasoning.
[0020] Based on the above technical solution, the present invention has the following beneficial technical effects: 1. This invention improves the stability and reproducibility of large language models when participating in microservice root cause analysis. By entrusting the reading of raw telemetry data, anomaly detection, and candidate localization to the deterministic evidence extraction module, and enabling the large language model to reason only through structured evidence cards and read-only tools, it reduces the problems of field misreading, evidence omission, and unstable running path caused by freely exploring raw telemetry data.
[0021] 2. This invention improves the adaptability of root cause analysis methods across different microservice systems. By describing telemetry files, fields, time units, component levels, candidate components, candidate causes, resource type mappings, and deployment topologies through declarative adapters, system differences are centralized in the adapters. The same evidence extraction and reasoning process can be applied to microservice systems with different hierarchical structures and different telemetry modalities.
[0022] 3. This invention improves the ability to locate and attribute causes in complex fault propagation scenarios. By extracting indicators, call chains, logs and business health signals into evidence cards, and combining them with morphological anomaly detection, topology roll-up, weak clue fallback and restricted tool reasoning, it can help distinguish the root cause component and the victim component when multiple components are abnormal at the same time, and further distinguish among multiple coarse-grained cause candidates. Attached Figure Description
[0023] Figure 1 This is a schematic diagram of the overall system architecture of the microservice root cause analysis method of the present invention.
[0024] Figure 2 This is a flowchart illustrating the microservice root cause analysis method based on evidence extraction and large language model reasoning of the present invention. Detailed Implementation
[0025] To describe the present invention in more detail, the technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
[0026] This embodiment provides a microservice root cause analysis method based on evidence extraction and large language model reasoning. The method takes diagnostic queries and microservice telemetry data as input, and outputs the fault occurrence time, faulty component, and fault cause. It mainly includes three stages: declarative system adaptation, deterministic multimodal evidence extraction, and restricted large language model reasoning based on structured evidence. The system implementation framework of this method is as follows: Figure 1 As shown, the system includes an observable data source, a declarative adapter configuration module, a deterministic evidence extraction module, a restricted tool interface module, and a large language model inference engine. The observable data source is responsible for continuously collecting multimodal telemetry data from the microservice system. The declarative adapter configuration module is responsible for maintaining the structural configuration of the target system. The deterministic evidence extraction module is responsible for performing anomaly detection and evidence card generation. The restricted tool interface module is responsible for encapsulating structured evidence into a read-only API. The large language model inference engine is responsible for root cause determination based on the evidence returned by the tool.
[0027] In this embodiment, it is assumed that the target microservice system is an online transaction system with a three-layer topology consisting of a node layer, a service layer, and an instance layer (container layer). During system operation, resource metrics, call chains, logs, and business health signals are continuously generated. The operations and maintenance personnel input a diagnostic query. The query contains the fault time window to be analyzed. Target fields (failure time, failed component, cause of failure) and expected number of failures The diagnostic engine performs as follows Figure 2 The steps shown are as follows: Step 1. Build a declarative adapter: Build a declarative adapter for this online trading system. ,adapter The system's telemetry data specifications (such as metrics, call chains, log file paths and field mappings) and candidate set (containing a set of nodes) were declared. Service Collection Instance collection and candidate cause set Mapping relationships (such as KPI to resource type mapping, resource type to coarse-grained cause mapping) and deployment topology (e.g., co-address mapping from instance to node, affiliation mapping from instance to service); the adapter only describes the objective structure of the system and does not contain parameters for fitting fault answers.
[0028] After the system is built, it will work with the adapter. Perform structural consistency checks; in this way, the structural differences between different microservice systems are confined to the adapter, and the main process of the root cause analysis method does not need to be rewritten for specific systems.
[0029] Step 2. Receive diagnostic queries and read telemetry data: The system receives diagnostic queries. Determine the time window According to the adapter The specification in the text reads the window. Full telemetry data for the given date. For any component. And any KPI, calculate the robust statistical baseline for the whole day. (Includes mean, standard deviation, and multiple quantiles); anomaly detection within the fault window is based on the corresponding component's own daily baseline. For reference.
[0030] Step 3. Perform morphology-aware multimodal anomaly detection: The system iterates through the KPI sequences of each component. If the variance of an instance's CPU utilization sequence is extremely small, it is identified as a dead sequence and discarded. If a node's memory utilization sequence remains consistently high (low quantiles are close to the maximum value), it is identified as a saturated sequence and marked as a blind spot. If the mean of a service's request volume sequence is close to zero, an absolute amplitude threshold is applied for low baseline detection to suppress false alarms caused by minor noise. For normal sequences, a robust statistical band detection window based on interquartile range is used. The system continuously deviates from its target range. Simultaneously, it detects sudden increases in log frequency and escalating call chain latency.
[0031] Step 4. Generate and refine structured evidence cards: The system converts detected anomalies into a set of evidence cards. For example, if a certain instance is detected... If both CPU utilization and trace latency are abnormal, then an instance-level evidence card is generated. .
[0032] To eliminate the problem of "downstream symptoms masking upstream root causes" caused by the propagation effect of microservice failures, the system subsequently utilized topology Perform volume-up correction: If an instance is found The physical node Above, there are a total of indivual( If co-located instances experience trace delay inflation at similar times, then node-level evidence cards are generated. and will Downgraded to supporting evidence; if multiple instances under the same service exhibit the same anomaly, aggregated evidence to generate service-level evidence cards. For sequences that do not meet strict thresholds but have high deviation, generate a set of weak clues. As a safety net.
[0033] Step 5. Provide evidence to the large language model through restricted read-only tools: The system encapsulates evidence cards, downsampled KPI sequences, call graph summaries, etc., into a read-only toolset. (such as evidence list tools, component viewing tools, call graph tools, etc.), and based on the adapter The declared available modalities are dynamically exposed to the large language model inference engine, and the large language model can only read the structured summary through tools and cannot directly access the original telemetry files or execute code.
[0034] Step 6. Reasoning and outputting root cause triples based on constrained tools and diagnostic rules: The large language model receives the query. Candidate set and toolset The explanation states that strict evidence cards or weak clues are obtained by calling the evidence list tool. The model follows diagnostic rules for reasoning: for example, obtaining the system-level fault start time by calling the business health signal tool. The system filters out early transient anomaly cards with mismatched occurrence times; it obtains the call topology using a call graph tool, tracing back from the downstream victim node with the highest latency to the upstream component with evidence of underlying resource anomalies; and it compares the health status of co-located neighbors using a component viewing tool to distinguish between node-level and instance-level failures. Finally, the model outputs root cause triples via an answer submission tool. System verification and Check whether it belongs to the candidate space of the closed set declared by the adapter, and output it in a formatted manner after the verification is passed.
[0035] In other embodiments, if the target system is a system lacking a logging modality, then a declarative adapter is used. It does not include logging specifications and has a limited toolset. The documentation does not provide log-related tools; the evidence extraction phase relies solely on metrics and call chains to generate evidence cards. If the target system is a microservice system with a flat architecture, the adapter only declares a single-level candidate component, and cross-level topology roll-up is not performed during the evidence extraction phase. All of these variations can utilize the same deterministic extraction process and restricted inference interface to complete root cause analysis.
[0036] The above description of the embodiments is provided to enable those skilled in the art to understand and apply the present invention. Those skilled in the art can readily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without creative effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made to the present invention by those skilled in the art based on the disclosure thereof should be within the scope of protection of the present invention.
Claims
1. A microservice root cause analysis method based on evidence extraction and large language model reasoning, characterized in that, Includes the following steps: (1) Build a declarative adapter for the microservice system to be diagnosed, and perform structural consistency verification on it after construction; the declarative adapter is a set of structured configurations of the microservice system; (2) Receive diagnostic queries input by users and read telemetry data from the microservice system to calculate the daily robust statistical baseline; (3) Based on the daily robust statistical baseline, the daily time series of each component KPI is divided into four forms, and form-aware multimodal anomaly detection is performed; (4) Convert the detected multimodal anomaly results into a unified evidence card, and use the deployment topology to perform roll-up correction operation on the evidence card; (5) Encapsulate the evidence cards and auxiliary summary information into a predefined read-only toolset and dynamically expose it to the large language model according to the modality declared by the adapter; (6) The large language model receives diagnostic queries and read-only toolsets, and infers and outputs root cause triples based on restricted tools and diagnostic rules.
2. The microservice root cause analysis method based on evidence extraction and large language model reasoning as described in claim 1, characterized in that: The microservice system in step (1) is a software system composed of multiple services, instances, and nodes that collaborate to complete business functions through network calls; the structured configuration set corresponding to the declarative adapter is denoted as... A and ,in This is a telemetry data specification used to describe the file location, field names, and time units of metrics, call chains, logs, and business health signals; This is a candidate set used to describe the component hierarchy, candidate components, and candidate failure causes. This is a mapping relationship used to describe the normalization of the original component identifier, the mapping from KPI name to resource type, and the mapping from resource type to coarse-grained reason candidates; For deployment topology, it is used to describe the ownership or co-location relationships between instances, services, and nodes.
3. The microservice root cause analysis method based on evidence extraction and large language model reasoning according to claim 1, characterized in that: The diagnostic query in step (2) is represented as follows: ,in Indicates the time window to be analyzed. This represents the set of target fields that need to be answered. This indicates the expected number of failures; then, based on the specification in the declarative adapter, it reads the microservice system's performance within the time window. The telemetry data within the system refers to observable data generated during the operation of the microservice system, including indicator data, call chain data, log data, and business health signals.
4. The microservice root cause analysis method based on evidence extraction and large language model reasoning according to claim 1, characterized in that: In step (2), for any component c Read the component c The daily time series of each KPI, and the calculation of the robust statistical baseline for the whole day based on this, are expressed as follows: in: For components c The k A robust daily statistical baseline for each KPI. These correspond to the mean, standard deviation, minimum extreme value, maximum extreme value, and quantile of a daily time series, respectively.
5. The microservice root cause analysis method based on evidence extraction and large language model reasoning according to claim 4, characterized in that, The specific implementation method of step (3) is as follows: like If so, the daily time series of KPIs is determined to be a dead series and discarded. This is a preset minimum threshold; If the difference between the high quantile and the low quantile of the sequence is less than a preset threshold If the daily time series of the KPI is determined to be a saturated series, it will be marked as an indicator blind zone, and anomaly detection will be forced to rely on logs or call chain modalities, i.e., the time windows will be counted separately. Sudden increases in log frequency and call chain latency escalation events within the system; like If the daily time series of KPIs is deemed to be a low baseline series, then anomaly detection requires that the relative deviation threshold be met simultaneously. With absolute amplitude threshold That is, any observation in the sequence relative to The deviation is greater than And this observed value is consistent with The absolute difference is greater than In such cases, the sequence is determined to be abnormal. This is a preset minimum threshold; All sequences other than the three types mentioned above are classified as normal sequences, and robust statistical bands based on interquartile range or median absolute deviation are used to detect abnormal segments in the sequence that deviate from the baseline by more than a preset multiple.
6. The microservice root cause analysis method based on evidence extraction and large language model reasoning according to claim 2, characterized in that: The evidence card in step (4) is a structured record generated during the deterministic evidence extraction stage, used to describe an abnormal combination of component and resource types. The evidence card is denoted as... ,in Represents a component. Indicates component hierarchy, Indicates the exception type. Indicates the start time of the exception. This indicates the supporting evidence set. Indicates the modality of evidence source. This represents the coarse-grained set of candidate causes.
7. The microservice root cause analysis method based on evidence extraction and large language model reasoning as described in claim 1, characterized in that, The step (4) of performing the roll-up correction operation on the evidence card includes the following three cases: Vertical Co-location Roll-Up: Setting up physical nodes Its managed collection of containers If there is a co-location relationship, At least one The containers are in time difference If an anomaly of the same resource type occurs within the same physical node, then that physical node will be generated. Level of evidence cards The original container-level card will be retained as supporting evidence. Horizontal Services Volume 1: Setting up Microservices and the collection of instances it contains There is a relationship of ownership, if If multiple instances of the same resource type experience anomalies within a similar timeframe, a service-level evidence card will be generated by aggregation. ; Subthreshold fallback: For scores that do not meet the strict outlier threshold but have soft deviations... Higher sequences generate weak cue sets , and according to Arranged in descending order; when strictly using a set of evidence cards When empty, As a fallback input in the reasoning stage.
8. The microservice root cause analysis method based on evidence extraction and large language model reasoning according to claim 1, characterized in that: The read-only toolset in step (5) includes an evidence list tool, a component viewing tool, a KPI sequence viewing tool, and an answer submission tool. When the microservice system has the corresponding modality, it also includes a call graph tool, a log summary tool, and a business health signal tool.
9. The microservice root cause analysis method based on evidence extraction and large language model reasoning according to claim 1, characterized in that, In step (6), the large language model infers based on restricted tools and diagnostic rules, and strictly follows the following restricted diagnostic rules for root cause identification: Relative amplitude priority rule: When comparing multiple candidate components, the relative deviation amplitude of the anomaly is obtained through the component viewing tool. The component with the largest specific deviation amplitude is identified as the root cause to distinguish the root cause from cascaded victim components. Call chain reverse tracing rules: When multiple components are abnormal, the call topology is obtained through the call graph tool. Starting from the downstream victim node with the highest latency or failure rate, the call chain is traced back along the call edge to lock the upstream component with evidence of underlying resource abnormality as the root cause. Golden signal time alignment rule: Obtain the start time of system-level faults through business health signal tools. Using this as a time anchor, filter out events that occurred at a time that are not related to the time of occurrence. Isolated evidence cards that do not match and cannot be explained by propagation delay.
10. The microservice root cause analysis method based on evidence extraction and large language model reasoning according to claim 6, characterized in that: In step (6), after the large language model completes the root cause identification, it outputs the root cause triples through the answer submission tool. ,in Indicates the time when the fault occurred. Indicates a faulty component. Indicate the cause of the fault; verification and The validated root cause triplet will be output in a formatted manner.
