Fault root cause positioning method and device based on large model and medium
By using knowledge-driven and data-driven large-scale models for collaborative reasoning, combined with semantic consistency analysis and cross-questioning mechanisms of expert agents, the problem of low accuracy and reliability of single large-scale models in fault root cause localization is solved, and more accurate root cause localization results are achieved.
Patent Information
- Application Number
- CN202511682889.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-17
- Publication Date
- 2026-02-13
AI Technical Summary
In existing technologies, a single large model has high task complexity in fault root cause localization, which can easily lead to inaccurate reasoning, and the accuracy and reliability of the generated results are low, and are affected by multiple factors.
It employs knowledge-driven and data-driven large-scale model collaborative reasoning, combined with expert intelligent agents for semantic consistency analysis, and triggers a cross-questioning mechanism when semantic inconsistencies occur, correcting the root cause localization results through reasonable questioning of the text.
By employing multi-perspective and differentiated root cause reasoning, the accuracy and reliability of fault root cause localization results are improved, the reasoning process is dynamically optimized, and more accurate root cause localization results are generated.
Smart Images
Figure CN121525866A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of large model technology, and in particular to a method, device and medium for fault root cause localization based on large models. Background Technology
[0002] Large models refer to artificial intelligence models based on neural networks with massive parameters. In recent years, they have made significant progress in fields such as natural language processing and visual understanding. Due to their powerful semantic understanding and contextual reasoning capabilities, large models are widely used in fault root cause analysis tasks in operation and maintenance. Current technologies mainly use a single large model, combined with methods such as prompting engineering, retrieval enhancement generation, or fine-tuning, to understand and reason about multi-source fault information such as logs, indicators, and alarms, and then generate and output root cause location results.
[0003] However, the above method also has the following technical problems: In the above methods, large models need to complete data understanding, association analysis, causal inference, and natural language generation in a single inference, which is too complex and can easily lead to inaccurate reasoning or "illusions". Furthermore, the generation results of large models are affected by multiple factors such as prompt design, retrieval quality, and the uncertainty of the model itself, resulting in insufficient stability. Therefore, the accuracy and reliability of root cause localization results determined by relying solely on the content generated by a single large model are low. Summary of the Invention
[0004] To address the aforementioned technical problems, the technical solution adopted by this invention is as follows: According to a first aspect of the present invention, a fault root cause localization method based on a large model is provided, the method comprising the following steps: S1. Input the fault information into the knowledge-driven large model and the data-driven large model respectively to obtain the first root cause localization result A and the second root cause localization result B corresponding to the fault information; where A is the root cause localization result output by the knowledge-driven large model; B is the root cause localization result output by the data-driven large model; the knowledge-driven large model is connected to the knowledge database.
[0005] S2. Based on the semantic consistency analysis of A and B by the expert intelligent agent, a semantic consistency judgment result is generated.
[0006] S3. If the semantic consistency determination result indicates semantic consistency, then A or B will be used as the target root cause localization result corresponding to the fault information, and the process will end; otherwise, the fault information and B will be input into the knowledge-driven large model to obtain the reasonableness questioning text corresponding to B; the fault information and A will be input into the data-driven large model to obtain the reasonableness questioning text corresponding to A.
[0007] S4. Input the fault information and the text questioning the reasonableness of A into the knowledge-driven large model to obtain the updated first root cause localization result A. 0 The fault information and the text questioning the reasonableness of B are input into the data-driven large model to obtain the updated second root cause localization result B. 0 .
[0008] S5, A 0 As A, B 0 As B, proceed to step S2.
[0009] According to a second aspect of the present invention, a non-transitory computer-readable storage medium is provided, wherein a computer program is stored in the storage medium, and the computer program is loaded and executed by a processor to implement the aforementioned method.
[0010] According to a third aspect of the present invention, an electronic device is provided, comprising: a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned method.
[0011] The present invention has at least the following beneficial effects: This invention provides a fault root cause localization method, device, and medium based on a large model. In the method, fault information is input into a knowledge-driven large model and a data-driven large model respectively to obtain a first root cause localization result and a second root cause localization result corresponding to the fault information. A semantic consistency analysis is performed on the first and second root cause localization results by an expert agent to generate a semantic consistency judgment result. If the semantic consistency judgment result indicates semantic consistency, then the first or second root cause localization result is used as the target root cause localization result corresponding to the fault information; otherwise, the fault information and the second root cause localization result are input into the knowledge-driven large model to obtain... Extract the text of the reasonableness challenge corresponding to the second root cause location result; input the fault information and the first root cause location result into the data-driven large model to obtain the text of the reasonableness challenge corresponding to the first root cause location result; input the fault information and the text of the reasonableness challenge corresponding to the first root cause location result into the knowledge-driven large model to obtain the updated first root cause location result; input the text of the reasonableness challenge corresponding to the fault information and the second root cause location result into the data-driven large model to obtain the updated second root cause location result; perform semantic consistency analysis again on the updated first root cause location result and the updated second root cause location result to obtain the target root cause location result corresponding to the fault information. As can be seen, this invention achieves multi-perspective and differentiated root cause reasoning through collaborative reasoning using a knowledge-driven large model and a data-driven large model. Furthermore, by combining semantic consistency analysis with an expert intelligent agent, a cross-questioning mechanism is triggered when semantic inconsistencies occur. One large model generates a reasonableness questioning text for the root cause localization result generated by another large model, allowing all models to revise the generated root cause localization result based on the reasonableness questioning text, generating an updated root cause localization result. This enables the large model to dynamically optimize the reasoning process. Compared to using a single large model to obtain the target root cause localization result, this invention effectively improves the accuracy and credibility of the target root cause localization result. Attached Figure Description
[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments 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.
[0013] Figure 1 This is a flowchart of a fault root cause localization method based on a large model, provided as an embodiment of the present invention. Detailed Implementation
[0014] 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.
[0015] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar tasks and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that comprises a series of steps or units is not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.
[0016] Embodiments of the present invention provide a fault root cause localization method based on a large model, the method comprising the following steps, such as... Figure 1 As shown: S1. Input the fault information into the knowledge-driven large model and the data-driven large model respectively to obtain the first root cause localization result A and the second root cause localization result B corresponding to the fault information; wherein, A is the root cause localization result output by the knowledge-driven large model; B is the root cause localization result output by the data-driven large model; the knowledge-driven large model is connected to the knowledge database.
[0017] Specifically, fault information includes at least a fault description text, and may also include alarm information, log data, call chain information, and system change records.
[0018] Specifically, a knowledge-driven large model can be understood as a large language model that has a connection with an external knowledge base. It retrieves knowledge fragments related to the input of the large language model from the external knowledge base through a retrieval mechanism, and inputs the knowledge fragments as contextual information into the large language model, so that the large language model generates responses based on real and traceable knowledge during the generation process.
[0019] Specifically, a data-driven large model can be understood as a large language model that is trained with large-scale corpus data, implicitly encodes language rules, factual knowledge and task logic into the model parameters, and completes the generation task based on the parameterized knowledge.
[0020] Specifically, while inputting the fault information into the knowledge-driven and data-driven large-scale models respectively, the preset prompts corresponding to the root cause localization task are also input into the knowledge-driven and data-driven large-scale models respectively. The preset prompts corresponding to the root cause localization task are used to guide the large-scale models to perform root cause analysis based on the input content and generate the corresponding root cause localization results. For example: Please analyze the possible root causes based on the following fault description and return the results in JSON format.
[0021] Specifically, the root cause analysis results are the analytical conclusions about the root cause of the failure, and these conclusions are presented in text form.
[0022] Specifically, the knowledge database includes a keyword index database and a vector database. The keyword index database includes several keywords and at least one text corresponding to each keyword; the vector database includes several semantic vectors and the text corresponding to each semantic vector.
[0023] Through the above steps, the knowledge-driven large model connects to the knowledge database, performing logical reasoning based on the knowledge in the database, emphasizing causality and interpretability. The data-driven large model does not connect to any database and does not rely on external knowledge bases; it directly mines statistical patterns and anomaly correlations from massive amounts of multi-source fault information such as logs, monitoring indicators, and alarms, possessing strong pattern recognition and context awareness capabilities. Fault information is input into both the knowledge-driven and data-driven large models to obtain the first and second root cause location results corresponding to the fault information. The target root cause location result is determined based on the first and second root cause location results. This dual-model integration approach leverages both knowledge reasoning and data mining. By leveraging the complementary advantages of knowledge-driven models and the divergent associative capabilities ("illusion" characteristics) of data-driven models, differentiated root cause hypotheses are generated. This provides input for subsequent consistency verification and cross-examination, triggering the large model to revise the generated root cause localization results in conjunction with reasonable questioning texts, generating updated root cause localization results. This allows the large model to dynamically optimize the reasoning process, thereby driving the root cause localization results to iteratively converge towards a more accurate and credible direction. Furthermore, semantic consistency analysis is performed again on the updated first and second root cause localization results to obtain the target root cause localization results corresponding to the fault information, which helps improve the accuracy and credibility of the target root cause localization results.
[0024] Specifically, the method further includes the following steps S01-S04: constructing a knowledge database: S01. For each document to be added to the database, obtain the document type corresponding to the document to be added to the database.
[0025] Specifically, the documents to be added to the database should include at least historical fault reports, and may also include technical documents reflecting fault handling experience, such as operation and maintenance manuals.
[0026] Furthermore, historical fault reports can be understood as fault reports that have been received up to the current point in time.
[0027] Specifically, a fault report should include at least the fault information, the root cause localization result corresponding to the fault information, and the fault handling method corresponding to the fault information.
[0028] Specifically, the document types corresponding to the documents to be entered into the database include semi-structured documents and unstructured documents. Those skilled in the art will know that any method in the prior art for determining the document type corresponding to the document to be entered into the database is within the protection scope of this invention, such as manual annotation, classification based on machine learning or deep learning, which will not be elaborated here.
[0029] S02. If the document to be imported is a semi-structured document, then the document is segmented according to a preset keyword dictionary to obtain several texts corresponding to the document and proceed to step S03. If the document to be imported is an unstructured document, then the document is segmented according to a preset text segmentation method to obtain several texts corresponding to the document and proceed to step S04. The preset text segmentation method is a text segmentation method pre-set by those skilled in the art according to actual needs, such as segmenting according to paragraph structure and font size. For long paragraphs, punctuation marks are combined with a sliding window method for segmentation. This will not be elaborated further here.
[0030] In one specific embodiment, the document type corresponding to the document to be added to the database also includes structured documents.
[0031] Furthermore, in step S2, if the document to be added to the database is a structured document, then the document to be added to the database is segmented into text according to a preset keyword dictionary to obtain several texts corresponding to the document to be added to the database and proceed to step S03.
[0032] Specifically, before performing text segmentation on the documents to be imported, document preprocessing is performed on the documents to be imported, which includes at least data cleaning.
[0033] Specifically, the preset keyword dictionary includes several preset keyword types and several preset keywords corresponding to each preset keyword type. The preset keyword types are keyword types pre-set by those skilled in the art according to actual needs. For example, the preset keyword type is fault phenomenon, and its corresponding preset keywords include: error, exception, failure, timeout, and inaccessibility; the preset keyword type is root cause of fault, and its corresponding preset keywords include: configuration error, resource exhaustion, and code defect; these will not be elaborated further here.
[0034] Specifically, in step S02, the document to be added to the database is segmented according to a preset keyword dictionary to obtain several texts corresponding to the document, including the following sub-steps: S021. Based on the preset keyword dictionary and the document to be added to the database, determine several initial keyword information corresponding to the document to be added to the database. The initial keyword information includes the initial keyword, the position of the initial keyword in the document to be added to the database, and the preset keyword type corresponding to the initial keyword; wherein, the initial keyword is the preset keyword in the preset keyword dictionary.
[0035] S022. Based on the preset keyword segmentation method and all initial keyword information corresponding to the document to be added to the database, determine the text segmentation point corresponding to the document to be added to the database; wherein, the preset keyword segmentation method is a segmentation method pre-set by those skilled in the art according to actual needs, such as: consecutive initial keywords of the same preset keyword type are not segmented repeatedly, initial keywords of different preset keyword types are regarded as segmentation points, and the text length between two adjacent segmentation points is not less than the preset length threshold, etc., which will not be elaborated here.
[0036] S023. Segment the document to be added to the database according to the text segmentation points corresponding to the document to be added to the database, so as to obtain several texts corresponding to the document to be added to the database.
[0037] Through the above steps, based on the preset keyword dictionary and the document to be added to the database, several initial keyword information corresponding to the document to be added to the database is determined; based on the preset keyword segmentation method and all the initial keyword information corresponding to the document to be added to the database, the text segmentation points corresponding to the document to be added to the database are determined; the document to be added to the database is segmented according to the text segmentation points corresponding to the document to be added to the database to obtain several texts corresponding to the document to be added to the database; this avoids incomplete semantics of the text obtained after segmentation, which helps to improve the accuracy and rationality of text segmentation.
[0038] S03. Obtain the keywords of each text corresponding to the document to be added to the database and store each keyword and its corresponding text in the keyword index database; those skilled in the art will know that any method of obtaining keywords of text in the prior art is within the protection scope of this invention, such as: filtering keywords based on word frequency and document frequency, filtering keywords based on graph sorting algorithm, using a pre-trained model to extract keywords, and matching based on a preset keyword dictionary to obtain keywords, which will not be elaborated here.
[0039] Specifically, a keyword index database can be understood as an inverted index database.
[0040] S04. Obtain the semantic vector corresponding to each text of the document to be added to the database and store each semantic vector and its corresponding text in the vector database. Those skilled in the art will know that any method of obtaining the semantic vector corresponding to the text in the prior art is within the protection scope of this invention. For example, the semantic vector corresponding to the text is obtained by encoding the text through the embedding model, which will not be described in detail here.
[0041] Through the above steps, different document segmentation strategies are adopted for texts of different document types to be added to the database. For semi-structured texts, the text is segmented according to a preset keyword dictionary, and keywords corresponding to the segmented text are extracted to build a keyword index database, thereby preserving key fields and structured semantics, supporting efficient and accurate matching and rule-based queries. For unstructured texts, the text is segmented according to a preset text segmentation method, and semantic vectors corresponding to the segmented text are obtained to build a vector database, supporting fuzzy retrieval and contextual understanding based on semantic similarity. The hybrid storage strategy of keyword index and vector index achieves fine-grained management and efficient retrieval support for documents to be added to the database. It not only preserves the accuracy of structured information but also leverages the generalization ability of semantic vectors, laying a high-quality, multimodal knowledge foundation for subsequent multi-path retrieval and improving retrieval efficiency.
[0042] Specifically, in step S1, while inputting the fault information into the knowledge-driven large model, several target recall texts corresponding to the fault information are also input into the knowledge-driven large model, and the target recall texts are obtained from the knowledge database.
[0043] Specifically, the method further includes the following steps S10-S40: obtaining the target recall text corresponding to the fault information: S10. Based on the fault keywords, perform keyword recall in the keyword index database to obtain several first candidate texts; the fault keywords are extracted from the fault information.
[0044] Specifically, the same method used in step S03 to obtain the keywords of each text corresponding to the document to be added to the database is used to extract fault keywords from the fault information, which will not be elaborated here.
[0045] Furthermore, there is at least one fault keyword.
[0046] Specifically, step S10 also includes the following sub-steps: S11. Traverse the keyword index database. When a keyword in the keyword index database is the same as any faulty keyword, take the text corresponding to the keyword in the keyword index database as the first candidate text.
[0047] S12. According to the preset matching score algorithm, obtain the matching score of each first candidate text and sort all the first candidate texts in descending order of matching score.
[0048] Specifically, the preset matching score algorithm is BM25 or TF-IDF.
[0049] S13. Select the top k first candidate texts from all sorted first candidate texts as k first candidate texts, where k is the preset number of first candidate texts. The preset number of first candidate texts can be set by those skilled in the art according to actual needs, for example, 5, 10, 15, which will not be elaborated here.
[0050] S20. Based on the semantic vector corresponding to the fault information, perform vector recall in the vector database to obtain several second candidate texts.
[0051] Specifically, the semantic vector corresponding to the fault information is obtained in the same way as in step S04, which is used to obtain the semantic vector corresponding to each text of the document to be added to the database. This will not be elaborated further here.
[0052] Specifically, step S20 includes the following sub-steps: S21. Traverse the vector database and obtain the vector similarity between the semantic vectors in the vector database and the semantic vectors corresponding to the fault information. The greater the vector similarity, the more similar the two semantic vectors are.
[0053] Specifically, the vector similarity is no greater than 1 and no less than 0.
[0054] In one specific embodiment, the vector distance between the semantic vector in the vector database and the semantic vector corresponding to the fault information is converted into a value between 0 and 1, and this value is used as the vector similarity between the semantic vector in the vector database and the semantic vector corresponding to the fault information. The larger the vector distance between the semantic vector in the vector database and the semantic vector corresponding to the fault information, the closer the value between 0 and 1 obtained by the vector distance conversion is to 0; the smaller the vector distance between the semantic vector in the vector database and the semantic vector corresponding to the fault information, the closer the value between 0 and 1 obtained by the vector distance conversion is to 1. Those skilled in the art will understand that any existing method for converting vector distance into a value between 0 and 1, where a larger vector distance results in a value closer to 0 and a smaller vector distance results in a value closer to 1, falls within the protection scope of this invention and will not be elaborated further here.
[0055] S22. Sort all semantic vectors in the vector database according to the order of vector similarity between the semantic vectors in the vector database and the semantic vectors corresponding to the fault information from largest to smallest, and select the top k semantic vectors from all sorted semantic vectors as k key semantic vectors.
[0056] S23. Use the text corresponding to each key semantic vector in the vector database as the second candidate text.
[0057] S30. The first and second candidate texts are used as the third candidate texts, and all third candidate texts are sorted by the re-ranking model. The re-ranking model scores each third candidate text and sorts all third candidate texts in descending order of the scores.
[0058] Specifically, the re-ranking model is the rerank model; the rerank model scores the relevance of each third candidate text and ranks all third candidate texts in descending order of the scores.
[0059] S40. Select the top n third candidate texts from all sorted third candidate texts as n target recall texts, where n is the preset number of target recall texts. The preset number of target recall texts can be set by those skilled in the art according to actual needs, for example, 3, 5, 7, 10, which will not be elaborated here.
[0060] Through the above steps, based on the fault keywords extracted from the fault information, keyword recall is performed in the keyword index database to obtain several first candidate texts; based on the semantic vectors corresponding to the fault information, vector recall is performed in the vector database to obtain several second candidate texts; the first and second candidate texts are used as third candidate texts, and all third candidate texts are sorted through a re-ranking model, and the top n third candidate texts after sorting are obtained as target recall texts; the use of multi-path recall to obtain first and second candidate texts achieves the complementary advantages of precise matching and semantic understanding, expanding the screening range of target recall texts; all third candidate texts are uniformly fine-ranked through a re-ranking model, and finally the top n third candidate texts after sorting are used as target recall texts, ensuring that the target recall texts achieve optimal relevance, accuracy, and completeness.
[0061] S2. Based on the semantic consistency analysis of A and B by the expert intelligent agent, a semantic consistency judgment result is generated.
[0062] Specifically, in step S2, A, B, and the preset prompt words corresponding to the semantic consistency analysis task are input together to the evaluation expert agent, which analyzes whether the semantics of A and B are consistent and generates a semantic consistency judgment result. The preset prompt words corresponding to the semantic consistency analysis task are used to guide the large model to judge whether the semantics of the two root cause localization results are consistent and generate a semantic consistency judgment result.
[0063] Specifically, the semantic consistency determination result includes consistency and inconsistency. If the semantic consistency determination result is consistent, it indicates semantic consistency; if the semantic consistency determination result is inconsistent, it indicates semantic inconsistency.
[0064] Specifically, the expert evaluator is a software module that integrates a large language model, several external API interfaces, and several callable functions.
[0065] S3. If the semantic consistency determination result indicates semantic consistency, then A or B is taken as the target root cause localization result corresponding to the fault information, and the process ends; otherwise, the fault information and B are input into the knowledge-driven large model to obtain the reasonableness questioning text corresponding to B; the fault information and A are input into the data-driven large model to obtain the reasonableness questioning text corresponding to A.
[0066] Specifically, while inputting the fault information and B into the knowledge-driven large model, several target recall texts corresponding to the fault information and preset prompt words corresponding to the rationality questioning task are also sent to the knowledge-driven large model; the preset prompt words corresponding to the rationality questioning task are used to guide the large model to verify the rationality of the root cause localization results in the input and generate rationality questioning text.
[0067] Specifically, while inputting the fault information and A into the data-driven large model, the preset prompt words corresponding to the rationality questioning task are also sent to the data-driven large model.
[0068] In one specific embodiment, the text challenging the reasonableness of the case may include pointing out logical contradictions, insufficient evidence, or conflicts with historical experience, in order to assist in the final decision-making process.
[0069] Through the above steps, when the semantic consistency determination result indicates semantic consistency, it means that the root cause of the fault expressed by the first root cause location result and the second root cause location result is the same. Therefore, the first root cause location result or the second root cause location result can be directly used as the target root cause location result. Otherwise, it means that the root cause of the fault expressed by the first root cause location result and the second root cause location result is different. At this time, the correct root cause location result cannot be determined. Therefore, it is necessary to question and verify the rationality of the first root cause location result and the second root cause location result. Based on the rationality questioning and verification result, that is, the rationality questioning text, the updated first root cause location result and the updated second root cause location result are further obtained. Through iterative processing, that is, semantic consistency analysis is performed again on the updated first root cause location result and the updated second root cause location result to obtain the target root cause location result corresponding to the fault information, which is conducive to improving the accuracy and credibility of the target root cause location result.
[0070] S4. Input the fault information and the reasonableness questioning text corresponding to A into the knowledge-driven large model to obtain the updated first root cause localization result A. 0 The fault information and the text questioning the reasonableness of B are input into the data-driven large model to obtain the updated second root cause localization result B. 0 .
[0071] Specifically, in step S4, while inputting the fault information and the reasonableness questioning text corresponding to A into the knowledge-driven large model, several target recall texts corresponding to the fault information are also input into the knowledge-driven large model.
[0072] Specifically, in step S4, while inputting the fault information and the text of reasonableness questioning corresponding to A into the knowledge-driven big model, the preset prompt words corresponding to the root cause localization task are also input into the knowledge-driven big model; while inputting the fault information and the text of reasonableness questioning corresponding to B into the data-driven big model, the preset prompt words corresponding to the root cause localization task are also input into the data-driven big model.
[0073] S5, A 0 As A, B 0 As B, proceed to step S2.
[0074] Through the above steps, the fault information is input into the knowledge-driven large-scale model and the data-driven large-scale model respectively to obtain the first root cause localization result and the second root cause localization result corresponding to the fault information. The expert agent performs semantic consistency analysis on the first and second root cause localization results to generate a semantic consistency judgment result. If the semantic consistency judgment result indicates semantic consistency, then the first or second root cause localization result is used as the target root cause localization result corresponding to the fault information; otherwise, the fault information and the second root cause localization result are input into the knowledge-driven large-scale model to obtain the rationality corresponding to the second root cause localization result. Questioning text; inputting fault information and the first root cause location result into a data-driven large model to obtain the reasonableness questioning text corresponding to the first root cause location result; inputting fault information and the reasonableness questioning text corresponding to the first root cause location result into a knowledge-driven large model to obtain the updated first root cause location result; inputting fault information and the reasonableness questioning text corresponding to the second root cause location result into a data-driven large model to obtain the updated second root cause location result; performing semantic consistency analysis again on the updated first root cause location result and the updated second root cause location result to obtain the target root cause location result corresponding to the fault information. By employing collaborative reasoning through knowledge-driven and data-driven large-scale models, multi-perspective and differentiated root cause reasoning is achieved. Furthermore, semantic consistency analysis is performed in conjunction with expert intelligent agents. When semantic inconsistencies arise, a cross-questioning mechanism is triggered, where one large model generates a reasonableness challenge text for the root cause localization results generated by another large model. This allows all models to revise their generated root cause localization results based on the reasonableness challenge text, resulting in updated root cause localization results. This enables the large models to dynamically optimize the reasoning process. Compared to using a single large model to obtain the target root cause localization results, this approach effectively improves the accuracy and credibility of the target root cause localization results.
[0075] Specifically, after step S3, steps S31-S32 are also included: S31. Based on the target root cause localization result corresponding to the fault information, execute the fault handling method corresponding to the target root cause localization result to obtain the fault handling result.
[0076] In one specific embodiment, the fault handling method corresponding to the target root cause localization result is determined manually based on the target root cause localization result.
[0077] In one specific embodiment, the fault handling results include information such as service recovery status, execution logs, and time consumption information for subsequent review and effect evaluation.
[0078] S32. Based on the fault information, the target root cause localization result corresponding to the fault information, and the user's feedback score on the target root cause localization result, a training sample is constructed and added to the training sample set; wherein, the training sample includes the fault information and the target root cause localization result corresponding to the fault information, and the user's feedback score on the target root cause localization result corresponding to the fault information is greater than the preset score; the training sample set is used to fine-tune the data-driven large model.
[0079] Specifically, after obtaining the fault handling results, the user's feedback rating on the root cause localization results is obtained.
[0080] Specifically, the feedback score should be no less than 0.1 and no greater than 1.
[0081] Specifically, the preset score is 0.6.
[0082] In one specific embodiment, when the user's feedback score for the root cause localization result is not greater than the preset score, detailed correction instructions are manually added and the root cause localization result is corrected to obtain an updated root cause localization result. The updated root cause localization result is then evaluated for reasonableness using a data-driven large model to obtain a reasonableness evaluation result. If the reasonableness evaluation result indicates that the updated root cause localization result is reasonable, a training sample is constructed based on the fault information and the updated root cause localization result, and this training sample is added to the training sample set.
[0083] Through the above steps, based on the root cause localization result corresponding to the fault information, the fault handling method corresponding to the root cause localization result is executed to obtain the fault handling result; after obtaining the fault handling result, the user's feedback score on the root cause localization result is obtained; based on the fault information, the root cause localization result corresponding to the fault information, and the user's feedback score on the root cause localization result, a training sample is constructed and added to the training sample set; the training sample set is used to fine-tune the data-driven large model; the training sample is constructed based on the root cause localization result with high feedback score and its corresponding fault information, making the training sample a high-quality sample. Fine-tuning the data-driven large model based on the high-quality training sample enables the data-driven large model to continuously learn and optimize on the basis of real and reliable experience, which can improve the robustness and long-term availability of the data-driven large model.
[0084] Specifically, after step S32, the following steps S33-S34 are also included: S33. When a new training sample is added to the training sample set, let count = count + 1, where count is the cumulative number of updates and count is initially 0.
[0085] S34. When count matches the preset update count, fine-tune the data-driven large model based on the current training sample set and set count=0. The preset update count is the number of updates that a person skilled in the art sets in advance according to actual needs, such as 1000, 2000, 3000, which will not be elaborated here.
[0086] Through the above steps, the training sample set is updated every time a new training sample is added to the training sample set. When the cumulative number of updates reaches the preset number of updates, the data-driven large model is fine-tuned based on the current training sample set, and the cumulative number of updates is reset to 0. This ensures that the data-driven large model is fine-tuned only after a cumulative number of new training samples have been added. This avoids the waste of resources caused by frequent training and ensures that the data-driven large model can regularly absorb high-quality experience and continuously optimize.
[0087] In one specific embodiment, the following steps are included after step S4: S41. Let Z = Z + 1; where Z is the cumulative number of iterations corresponding to the target root cause localization result, and the initial cumulative number of iterations corresponding to the target root cause localization result is 0.
[0088] S42, When Z≤Z 0 At that time, A 0 As A, B 0 As B, proceed to step S2; when Z > Z 0 If the current knowledge-driven large model's model label is a trustworthy large model label, then A will be... 0 As the target root cause localization result corresponding to the fault information; if the model label of the current data-driven large model is a trustworthy large model label, then B will be... 0 As the target root cause localization result corresponding to the fault information; where Z 0 The maximum number of iterations is preset; the preset maximum number of iterations can be set by those skilled in the art according to actual needs, for example: 5, 10, 15, which will not be elaborated here.
[0089] Specifically, model tags include trustworthy large model tags and untrustworthy large model tags. Trustworthy large model tags indicate that the generated content of the large model has high credibility, while untrustworthy large model tags indicate that the generated content of the large model has low credibility.
[0090] Through the above steps, each time the updated first root cause localization result and the updated second root cause localization result are obtained, the cumulative iteration count corresponding to the target root cause localization result is incremented by 1. If the cumulative iteration count corresponding to the target root cause localization result is not greater than the preset maximum iteration count, then semantic consistency analysis is performed again on the updated first root cause localization result and the updated second root cause localization result to obtain the target root cause localization result corresponding to the fault information. Otherwise, when the model label of the current knowledge-driven large model is a trustworthy large model label, then the updated first root cause localization result is taken as the target root cause localization result corresponding to the fault information; when the current data-driven... When the model label of the dynamic large model is a credible large model label, the updated second root cause localization result is used as the target root cause localization result corresponding to the fault information. By setting the cumulative iteration count and the preset maximum iteration count, the number of repetitions of semantic consistency analysis and cross-questioning process is limited. Even if the semantics of the root cause localization results generated by the knowledge-driven large model and the root cause localization results generated by the data-driven large model cannot be consistent, the target root cause localization result can be determined based on the model label of the current knowledge-driven large model and the model label of the current data-driven large model when the cumulative iteration count reaches the preset maximum iteration count, thus avoiding getting stuck in infinite iteration.
[0091] Specifically, after step S3, the method further includes: when Z=0 and the semantic consistency determination result indicates semantic inconsistency, then the current A is taken as the first initial root cause localization result corresponding to the fault information, the current B is taken as the second initial root cause localization result corresponding to the fault information, and the fault information is marked as specific fault information.
[0092] Specifically, the model labels for the current knowledge-driven large model and the current data-driven large model are obtained through the following steps: S100. Sort the target root cause localization results corresponding to all specific fault information in order from latest to earliest according to the generation time of the target root cause localization results.
[0093] S200. From the target root cause localization results corresponding to all sorted specific fault information, select the specific fault information corresponding to the first m target root cause localization results as the specified fault information corresponding to the current time point, and obtain the specified fault information set C={C1, C2, ..., C...} corresponding to the current time point. j , ..., C m}, C j This is the j-th specified fault information corresponding to the current time point, 1≤j≤m; m is the preset number of specified fault information. The preset number of specified fault information can be set by those skilled in the art according to actual needs, for example: 5000, 10000, 20000, which will not be elaborated here.
[0094] S300, if C m The corresponding root cause localization results and C m If the semantic similarity between the corresponding second initial root cause localization results is not less than the preset semantic similarity, then C will be... m Important fault information is marked as corresponding to the current time point.
[0095] Specifically, C m The semantic vector corresponding to the target root cause localization result and C m The vector similarity between the semantic vectors corresponding to the second initial root cause localization results is used as C. m The corresponding root cause localization results and C m The semantic similarity between the corresponding second initial root cause localization results.
[0096] Specifically, C m The corresponding root cause localization results and C m The semantic similarity between the corresponding second initial root cause localization results is not less than 0 and not greater than 1, and the greater the semantic similarity, the better C. m The corresponding root cause localization results and C m The more similar the corresponding second initial root cause localization results are.
[0097] Specifically, the preset semantic similarity is a value that is no greater than 1 and no less than 0.6, set by those skilled in the art according to actual needs, such as 0.8 or 0.9, which will not be elaborated here.
[0098] S400, if E 0 / m≥E 1 If the condition is met, then the model label for the current data-driven large model is determined to be a trustworthy large model label, and the model label for the current knowledge-driven large model is determined to be an untrustworthy large model label; otherwise, the model label for the current knowledge-driven large model is determined to be a trustworthy large model label, and the model label for the current data-driven large model is determined to be an untrustworthy large model label. 0 E represents the number of critical fault messages corresponding to the current time point. 1 This is a preset percentage threshold.
[0099] Specifically, 0.6<E 1 <1.
[0100] Through the above steps, for fault information, if the semantics of the first root cause localization result generated by the knowledge-driven large model and the second root cause localization result generated by the data-driven large model are inconsistent, then the fault information is treated as specific fault information; the second root cause localization result is treated as the second initial root cause localization result; the target root cause localization results corresponding to all specific fault information are sorted according to the generation time of the target root cause localization results from latest to earliest; from the sorted target root cause localization results corresponding to all specific fault information, the specific fault information corresponding to the first m target root cause localization results is selected as the specified fault information corresponding to the current time point; the semantic similarity between the target root cause localization result and the second initial root cause localization result corresponding to the specified fault information at the current time point is calculated; if the semantic similarity is not less than the preset semantic similarity, then the specified fault information is marked as the current time point's pair. The system identifies key fault information. If the ratio of the number of key fault information items corresponding to the current time point to the number of specified fault information items corresponding to the current time point is not less than a preset percentage threshold, it indicates that the root cause localization results obtained in the most recent m times are mostly similar to the root cause localization results generated by the data-driven large model. The labels of the data-driven large model are relatively stable and accurate recently, and the content generated by the data-driven large model is more reliable. Therefore, the model label of the current data-driven large model is determined to be a trustworthy large model label, and the model label of the current knowledge-driven large model is determined to be an untrustworthy large model label. Conversely, if the ratio is not greater than the specified percentage threshold, the content generated by the knowledge-driven large model is considered more reliable, and the model label of the current knowledge-driven large model is determined to be a trustworthy large model label, and the model label of the current data-driven large model is determined to be an untrustworthy large model label. This achieves a dynamic, quantitative, and semantically aware evaluation of the relative trustworthiness of knowledge-driven and data-driven large models.
[0101] Furthermore, if the cumulative number of iterations corresponding to the target root cause localization result is greater than the preset maximum number of iterations, then when the model label of the current knowledge-driven large model is a credible large model label, the updated first root cause localization result will be used as the target root cause localization result corresponding to the fault information; when the model label of the current data-driven large model is a credible large model label, the updated second root cause localization result will be used as the target root cause localization result corresponding to the fault information; this is beneficial to improving the accuracy and credibility of the target root cause localization result.
[0102] Embodiments of the present invention also provide a non-transitory computer-readable storage medium that can be disposed in an electronic device to store a computer program related to implementing a method in the method embodiments, the computer program being loaded and executed by the processor to implement the method provided in the above embodiments.
[0103] Embodiments of the present invention also provide an electronic device, including: a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method provided in the above embodiments.
[0104] Embodiments of the present invention also provide a computer program product including program code, which, when the program product is run on an electronic device, causes the electronic device to perform the steps of the methods described above in various exemplary embodiments of the present invention.
[0105] This invention provides a fault root cause localization method, device, and medium based on a large model. In the method, fault information is input into a knowledge-driven large model and a data-driven large model respectively to obtain a first root cause localization result and a second root cause localization result corresponding to the fault information. A semantic consistency analysis is performed on the first and second root cause localization results by an expert agent to generate a semantic consistency judgment result. If the semantic consistency judgment result indicates semantic consistency, then the first or second root cause localization result is used as the target root cause localization result corresponding to the fault information; otherwise, the fault information and the second root cause localization result are input into the knowledge-driven large model to obtain... Extract the text of the reasonableness challenge corresponding to the second root cause location result; input the fault information and the first root cause location result into the data-driven large model to obtain the text of the reasonableness challenge corresponding to the first root cause location result; input the fault information and the text of the reasonableness challenge corresponding to the first root cause location result into the knowledge-driven large model to obtain the updated first root cause location result; input the text of the reasonableness challenge corresponding to the fault information and the second root cause location result into the data-driven large model to obtain the updated second root cause location result; perform semantic consistency analysis again on the updated first root cause location result and the updated second root cause location result to obtain the target root cause location result corresponding to the fault information. As can be seen, this invention achieves multi-perspective and differentiated root cause reasoning through collaborative reasoning using a knowledge-driven large model and a data-driven large model. Furthermore, by combining semantic consistency analysis with an expert intelligent agent, a cross-questioning mechanism is triggered when semantic inconsistencies occur. One large model generates a reasonableness questioning text for the root cause localization result generated by another large model, allowing all models to revise the generated root cause localization result based on the reasonableness questioning text, generating an updated root cause localization result. This enables the large model to dynamically optimize the reasoning process. Compared to using a single large model to obtain the target root cause localization result, this invention effectively improves the accuracy and credibility of the target root cause localization result.
[0106] While specific embodiments of the invention have been described in detail by way of examples, those skilled in the art should understand that the examples are for illustrative purposes only and are not intended to limit the scope of the invention. Those skilled in the art should also understand that various modifications can be made to the embodiments without departing from the scope and spirit of the invention.
Claims
1. A fault root cause localization method based on a large model, characterized in that, The method includes the following steps: S1. Input the fault information into the knowledge-driven large model and the data-driven large model respectively to obtain the first root cause localization result A and the second root cause localization result B corresponding to the fault information; wherein, A is the root cause localization result output by the knowledge-driven large model; B is the root cause localization result output by the data-driven large model; the knowledge-driven large model is connected to the knowledge database. S2. Based on the semantic consistency analysis of A and B by the expert intelligent agent, a semantic consistency judgment result is generated. S3. If the semantic consistency determination result indicates semantic consistency, then A or B is taken as the target root cause localization result corresponding to the fault information, and the process ends; otherwise, the fault information and B are input into the knowledge-driven large model to obtain the reasonableness questioning text corresponding to B; the fault information and A are input into the data-driven large model to obtain the reasonableness questioning text corresponding to A. S4. Input the fault information and the reasonableness questioning text corresponding to A into the knowledge-driven large model to obtain the updated first root cause localization result A. 0 The fault information and the text questioning the reasonableness of B are input into the data-driven large model to obtain the updated second root cause localization result B. 0 ; S5, A 0 As A, B 0 As B, proceed to step S2.
2. The fault root cause localization method based on a large model according to claim 1, characterized in that, The knowledge database includes a keyword index database and a vector database. The keyword index database contains several keywords and at least one text corresponding to each keyword; the vector database contains several semantic vectors and the text corresponding to each semantic vector.
3. The fault root cause localization method based on a large model according to claim 2, characterized in that, The method also includes the following steps to construct a knowledge database: S01. For each document to be added to the database, obtain the document type corresponding to the document to be added to the database; S02. If the document to be added to the database is a semi-structured document, then the document to be added to the database is segmented into text according to the preset keyword dictionary to obtain several texts corresponding to the document to be added to the database and proceed to step S03; if the document to be added to the database is an unstructured document, then the document to be added to the database is segmented into text according to the preset text segmentation method to obtain several texts corresponding to the document to be added to the database and proceed to step S04. S03. Obtain the keywords of each text corresponding to the document to be added to the database and store each keyword and its corresponding text in the keyword index database; S04. Obtain the semantic vector corresponding to each text of the document to be added to the database and store each semantic vector and its corresponding text in the vector database.
4. The fault root cause localization method based on a large model according to claim 3, characterized in that, In step S1, while inputting the fault information into the knowledge-driven large model, several target recall texts corresponding to the fault information are also input into the knowledge-driven large model, and the target recall texts are obtained from the knowledge database.
5. The fault root cause localization method based on a large model according to claim 4, characterized in that, The method further includes the following steps to obtain the target recall text corresponding to the fault information: S10. Based on the fault keywords, perform keyword recall in the keyword index database to obtain several first candidate texts; the fault keywords are extracted from the fault information. S20. Based on the semantic vector corresponding to the fault information, perform vector recall in the vector database to obtain several second candidate texts; S30. The first and second candidate texts are used as the third candidate texts, and all third candidate texts are sorted by the re-ranking model. The re-ranking model scores each third candidate text and sorts all third candidate texts in descending order of the scores. S40. Select the top n third candidate texts from all sorted third candidate texts as n target recall texts, where n is the preset number of target recall texts.
6. The fault root cause localization method based on a large model according to claim 4, characterized in that, In step S4, while inputting the fault information and the reasonableness questioning text corresponding to A into the knowledge-driven large model, several target recall texts corresponding to the fault information are also input into the knowledge-driven large model.
7. The fault root cause localization method based on a large model according to claim 3, characterized in that, The documents to be added to the database must include at least historical fault reports.
8. The fault root cause localization method based on a large model according to claim 1, characterized in that, The process after step S3 also includes: S31. Based on the target root cause localization result corresponding to the fault information, execute the fault handling method corresponding to the target root cause localization result to obtain the fault handling result. S32. Based on the fault information, the target root cause localization result corresponding to the fault information, and the user's feedback score on the target root cause localization result, a training sample is constructed and added to the training sample set; wherein, the training sample includes the fault information and the target root cause localization result corresponding to the fault information, and the user's feedback score on the target root cause localization result corresponding to the fault information is greater than the preset score; the training sample set is used to fine-tune the data-driven large model.
9. A non-transitory computer-readable storage medium, characterized in that, The storage medium stores a computer program, which is loaded and executed by a processor to implement the large model-based fault root cause localization method as described in any one of claims 1-8.
10. An electronic device, comprising: A processor, a memory, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the computer program, it implements the fault root cause localization method based on a large model as described in any one of claims 1-8.