Domestic intelligent adaptation system and method based on multi-agent interaction

Abstract

The invention relates to a localized intelligent adaptation system and method based on multi-agent interaction, the system is realized based on a large language model and multi-agent cooperation, and the system comprises an offline knowledge extraction module, a diagnosis prompt generation module, a multi-agent cooperation diagnosis module and a man-machine cooperation module. Wherein the offline knowledge extraction module is used for extracting knowledge blocks and dynamically registering an API (Application Program Interface) of an operation and maintenance tool into the system; the diagnosis prompt generation module is used for screening out most relevant knowledge blocks and operation and maintenance tools; the multi-agent cooperative diagnosis module is used for cooperatively generating a diagnosis report among the agents based on asynchronous communication and a cross review mechanism; and the man-machine cooperation module realizes cooperation of manual auditing and automatic repairing through a natural language dialogue interface and a work order system. The invention aims to improve the intelligent level of fault diagnosis and repair through the following innovation breakthrough, and better meet the requirements of localization adaptation.

Description

Technical Field

[0001] This invention relates to an intelligent operation and maintenance tool, specifically to a domestically developed intelligent adaptation system and method based on multi-agent interaction. Background Technology

[0002] Information technology application innovation has become a core path to ensure information security and achieve technological self-reliance and control. The entire chain of replacements, from chips and operating systems to databases and middleware, aims to build a secure and reliable information technology system. However, with the migration and adaptation of numerous applications, multiple challenges arise during the adaptation process:

[0003] Heterogeneous environment adaptation is complex: the compatibility of domestic software and hardware technology stacks with the original system is different, and traditional operation and maintenance tools are difficult to run efficiently across platforms, resulting in high migration and optimization costs.

[0004] Knowledge fragmentation and experience dependence: Enterprise knowledge bases and technical documents are mostly unstructured data, relying on manual experience for retrieval and application, making it difficult to achieve automated decision support.

[0005] Complex faults are difficult to pinpoint: Faults in domestic environments often involve multiple fields (such as network, storage, and application layers), and a single AI model is difficult to cover the entire link analysis and is prone to getting stuck in local optima.

[0006] Against this backdrop, intelligent adaptation to reduce the workload and difficulty of adaptation has become an urgent problem to be solved. Although large language models (LLM) have strong text generation capabilities, they lack the ability to dynamically integrate with tool APIs and cannot handle complex logical reasoning tasks. Existing multi-agent systems mostly adopt fixed collaboration paradigms, which are difficult to adapt to dynamically changing operation and maintenance scenarios.

[0007] With the rapid development of information technology application innovation industries, intelligent operation and maintenance has become key to improving operation and maintenance efficiency and quality. However, current fault diagnosis technologies based on Large Language Models (LLM) still have many shortcomings and are difficult to meet the needs of complex fault diagnosis during the localization and adaptation process.

[0008] (1) Inefficiency of knowledge extraction and application

[0009] Unstructured knowledge is difficult to utilize: Traditional keyword matching (such as TF-IDF) lacks semantic understanding capabilities, resulting in insufficient accuracy in knowledge base retrieval. For example, when a user queries "replace Oracle's NVL functions with DM syntax", it relies on human experience or a pre-set rule base, and cannot dynamically adapt to new scenarios.

[0010] Tools and knowledge are disconnected: API calls to operational tools (such as monitoring and log analysis) are not linked to the knowledge base, resulting in a lack of decision-making closure. Although existing solutions can build a knowledge system through the DeepSeek-R1 model, complex manual intervention and adaptation are still required.

[0011] (2) Decision-making bottleneck of single agent model

[0012] Insufficient domain coverage: When dealing with cross-domain faults, a single LLM model is prone to "illusions" or erroneous inferences due to the large differences between the distribution of pre-training data and operation and maintenance tasks. Although existing solutions can improve accuracy through layered training, they still need to be optimized separately for different scenarios.

[0013] Lack of dynamic reasoning ability: Traditional retrieval-based diagnosis relies on fixed rules and cannot simulate the multi-path exploration and backtracking of human experts. Although existing multi-agent recommendation systems can achieve task decomposition, they lack dynamic path optimization mechanisms such as tree search.

[0014] (3) Insufficient compatibility of domestic software and hardware ecosystem

[0015] Performance and security are difficult to balance: There is a performance gap between domestic hardware and the computing requirements of LLM, and existing optimization methods (such as model pruning) may sacrifice processing accuracy.

[0016] Toolchain integration is challenging: In a domestically adapted environment, plug-and-play frameworks for third-party tool APIs are still immature. While some existing multimodal agents support handles to multiple operating systems, cross-platform adaptation still requires customized development.

[0017] (4) The rigidity of multi-agent cooperation

[0018] Static task allocation mechanism: Although existing agent scheduling methods deconstruct tasks through LLM, they do not design asynchronous communication and cross-review mechanisms, resulting in low collaboration efficiency.

[0019] Lack of self-learning ability: The agent cannot dynamically update the knowledge base after the fault is repaired, resulting in the ineffective reuse of historical experience.

[0020] In summary, with the rapid development of information technology and the increasing demand for domestic adaptation, traditional fault diagnosis and repair methods have gradually revealed many limitations. Existing methods struggle to efficiently utilize unstructured knowledge, lack sufficient tool integration capabilities, have limited diagnostic accuracy for complex faults, and lack flexibility and robustness. Furthermore, in a domestically developed environment, they also face challenges such as insufficient support for domestic toolchains and difficulty in meeting the requirements for secure, controllable, and efficient operation and maintenance. Summary of the Invention

[0021] The main objective of this invention is to provide a domestically developed intelligent adaptation system and method based on multi-agent interaction, aiming to improve the intelligence level of fault diagnosis and repair through the following innovative breakthroughs, and better meet the needs of domestic adaptation.

[0022] To achieve the above objectives, this invention provides a domestically developed intelligent adaptation system based on multi-agent interaction. The system utilizes a large language model and multi-agent collaboration to diagnose and repair abnormal indicators, and includes the following modules:

[0023] The offline knowledge extraction module is used to extract knowledge blocks from unstructured documents and dynamically register the APIs of operation and maintenance tools related to the knowledge blocks into the system.

[0024] The diagnostic prompt generation module is used to select the knowledge blocks and operation and maintenance tools most relevant to the abnormal indicators from the knowledge blocks extracted by the offline knowledge extraction module and the operation and maintenance tools related to the knowledge blocks, and generate prompt templates based on the most relevant knowledge blocks and operation and maintenance tools selected.

[0025] The multi-agent collaborative diagnosis module includes a supervisor agent, a detection agent, and other agents. The supervisor agent assigns the detection agent to analyze the anomaly description in the prompt template generated by the diagnosis prompt generation module. Meanwhile, other agents question the analysis results, thereby triggering multiple rounds of diagnosis. The supervisor agent, detection agent, and other agents collaboratively generate a diagnostic report based on asynchronous communication and cross-review mechanisms.

[0026] The supervisor agent, the detection agent, and other agents are all equipped with a dynamic decision-making module based on tree search. The dynamic decision-making module adopts the Monte Carlo tree search algorithm, combined with the dynamically adjusted UCT algorithm formula to optimize multi-path exploration and backtracking, and improves diagnostic efficiency through a reflection mechanism.

[0027] The human-machine collaboration module achieves collaboration between manual review and automated repair through a natural language dialogue interface and a work order system based on diagnostic reports generated by the multi-agent collaborative diagnostic module.

[0028] Preferably, the offline knowledge extraction module includes a document learning unit and a tool unit;

[0029] The document learning unit is used to split unstructured documents into blocks that do not exceed the input limit of the large language model by chapter, and to build a summary for each chapter, thereby building a summary tree for the document and extracting knowledge blocks from the summary tree.

[0030] The knowledge blocks include:

[0031] Name: The topic of the knowledge block;

[0032] Content: Used to describe the specific content of the knowledge block, using functional description;

[0033] Metrics: A list of system metrics used for knowledge matching;

[0034] Analysis steps: The specific steps for conducting analysis using relevant indicators;

[0035] The tool unit is used to categorize operation and maintenance tools by function, provide usage instructions for each operation and maintenance tool, and dynamically register the APIs of the operation and maintenance tools to the system so that the large language model can call them during diagnosis.

[0036] Preferably, the diagnostic prompt generation module adopts a hybrid retrieval strategy, including knowledge block retrieval and matching retrieval of operation and maintenance tools.

[0037] Further optimized, the knowledge block retrieval uses the BM25 algorithm to retrieve the most relevant knowledge blocks based on anomaly indicators. The formula is as follows:

[0038]

[0039] Where f(q,D) is the word frequency of the anomaly indicator q in knowledge block D;

[0040] |D| represents the length of the knowledge block, calculated in terms of word count;

[0041] avgdl is the average knowledge block length of all knowledge blocks in the offline knowledge extraction module;

[0042] k1 and b are free parameters, and their default values ​​are usually k1 = 1.2 and b = 0.75.

[0043] IDF(q) is the inverse document frequency of the anomaly index q, and its calculation formula is as follows:

[0044]

[0045] Where N is the total number of knowledge blocks extracted; df(q) is the number of documents containing the anomaly indicator q.

[0046] Furthermore, the operation and maintenance tool matching retrieval is based on the fine-tuned Sentence-BERT model. The fine-tuned Sentence-BERT model is used to match the most relevant operation and maintenance tools. The most relevant operation and maintenance tools are selected by calculating the cosine similarity between the embedding vectors of the operation and maintenance tool description and the diagnostic context.

[0047] The Sentence-BERT model uses the cross-entropy loss function during fine-tuning to measure the difference between the predicted and actual values ​​of the classification model. The formula is as follows:

[0048]

[0049] Among them, y ij It is related to the diagnostic context of the operation and maintenance tools, p ij This represents the probability of anomaly prediction related to the diagnostic context of the operation and maintenance tools; n is the number of samples; m is the number of tool types corresponding to each anomaly description.

[0050] The formula for calculating the cosine similarity between the embedding vectors of the computational tool description and the diagnostic context is as follows:

[0051]

[0052] Where emb(.) represents the embedding function of the fine-tuning model. This formula selects the k values ​​with the highest similarity between the diagnostic context s and the diagnostic operation and maintenance tool t, that is, the top k operation and maintenance tools, calculated using the following Top-k formula:

[0053]

[0054] The selected top-k operation and maintenance tools are integrated into the prompt word project. Based on the large language model of these operation and maintenance tools, request APIs are generated, and the tool execution results are obtained to enhance root cause diagnosis.

[0055] Preferably, in the tree-search-based dynamic decision-making module: a dynamically adjusted UCT algorithm is used to search for the node in the tree that best matches the anomaly description. The UCT algorithm formula is as follows:

[0056]

[0057] Among them: w i Node v i Cumulative win rate; n i Node v i N: Total number of visits to the parent node; c: Exploration weight constant.

[0058] Preferably, the multi-agent collaborative diagnostic module includes a decision coordination agent, an anomaly detection agent, a root cause localization agent, a knowledge reasoning agent, a repair execution agent, an effect verification agent, and a human-computer interaction agent;

[0059] Among them, the decision-making and coordination agent is the supervisory intelligent agent, which is used for task scheduling and resource allocation;

[0060] Anomaly detection Agent is a detection intelligent agent used for initial screening of anomalies;

[0061] Root cause localization agent is used for cross-layer fault correlation analysis, knowledge reasoning agent is used for dynamic application of domain knowledge, repair execution agent is used for automated handling implementation, effect verification agent is used for reliable confirmation of repair results, and human-computer interaction agent is used for decision visualization and human intervention.

[0062] In a further optimized manner, the decision coordination agent sends a prompt template to the anomaly detection agent based on the anomaly description. After the anomaly detection agent performs a preliminary screening of the anomaly description, it publishes the anomaly description to the message queue. The root cause localization agent performs root cause analysis on the anomaly description by initiating a tree search diagnosis and generates a corresponding diagnostic report. The knowledge reasoning agent raises questions about the diagnostic report, and the root cause localization agent performs root cause analysis again on the questions, thereby triggering multiple rounds of diagnosis to form the final diagnostic report.

[0063] Furthermore, the human-machine collaboration module supports:

[0064] After manually reviewing the repair suggestions, you can choose to execute them automatically or manually.

[0065] A plug-and-play framework for domestically developed toolchains.

[0066] This invention also provides a domestically developed intelligent adaptation method, based on a large language model and multi-agent collaboration, comprising the following steps:

[0067] S100: The domestically developed intelligent adaptation system based on multi-agent interaction detected a server anomaly and triggered a multi-agent collaborative diagnostic process.

[0068] S200: The offline knowledge extraction module extracts relevant knowledge blocks and related operation and maintenance tools based on abnormal signals, and dynamically registers the API of the operation and maintenance tools to the system;

[0069] S300: The diagnostic prompt generation module selects the knowledge blocks and operation and maintenance tools most relevant to the abnormal signals based on the knowledge blocks and operation and maintenance tools extracted by the offline knowledge extraction module, and generates a prompt template;

[0070] S400: The supervisor agent in the multi-agent collaborative diagnosis module assigns the abnormality description in the prompt template to an agent with the corresponding function. The agent with the corresponding function provides the corresponding diagnosis report based on the dynamic decision-making module of tree search.

[0071] In this step, if the diagnostic path of the tree-search-based dynamic decision-making module fails, a reflection mechanism is triggered to backtrack to the parent node and prune invalid paths.

[0072] S500: An intelligent agent with the ability to raise questions can challenge the diagnostic report, thereby triggering multiple rounds of diagnosis and collaboratively generating the final diagnostic report.

[0073] S600: Based on the final diagnostic report and proposed repair suggestions, the maintenance engineer will reassess the fault situation and determine whether to perform automatic repair or manual repair based on the fault level information.

[0074] The beneficial effects of this invention are as follows:

[0075] (1) Knowledge structuring and dynamic tool integration:

[0076] By leveraging the semantic understanding capabilities of large language models, knowledge blocks with associated metrics are extracted from unstructured documents, and domestically developed operation and maintenance tools are dynamically registered via API to build a "knowledge-tool" linkage system.

[0077] (2) Multi-agent cooperative reasoning:

[0078] Multiple large language model expert agents are generated based on domain knowledge clustering. Through asynchronous communication and cross-review mechanisms, the cross-domain expert consultation process is simulated to improve the accuracy of complex fault diagnosis.

[0079] (3) Tree search-driven dynamic decision-making:

[0080] The Monte Carlo Tree Search (MCTS) algorithm is introduced, and the UCT formula is combined to optimize multi-path exploration and backtracking, ensuring that the diagnostic process is both efficient and robust.

[0081] (4) Human-machine collaboration and localization adaptation:

[0082] The design incorporates a manual review and intervention mechanism, supports plug-and-play functionality for domestically developed toolchains, and meets the dual requirements of domestic adaptation for security, controllability, and efficient operation and maintenance. Attached Figure Description

[0083] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0084] Figure 1 This is a flowchart of the multi-agent collaborative fault diagnosis process of the present invention;

[0085] Figure 2 This invention relates to an asynchronous communication process for multi-Agent collaborative fault diagnosis. Detailed Implementation

[0086] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0087] This embodiment uses a Kylin OS server and a DM database as an example:

[0088] S100: When the system detects anomalies in the Kylin OS server and the DM database node, it triggers a multi-agent collaborative diagnostic process.

[0089] S200: The offline knowledge extraction module extracts relevant knowledge blocks and related operation and maintenance tools based on the abnormal signals, and dynamically registers the API of the operation and maintenance tools to the system.

[0090] The offline knowledge extraction module is initialized by constructing a fault knowledge tree and structuring knowledge blocks.

[0091] Fault Knowledge Tree Construction: The domestically developed intelligent adaptation system based on multi-agent interaction extracts diagnostic documents (such as disk I / O anomaly handling guides) related to the Kylin OS server and DM database nodes from the internal knowledge base. These documents are then divided into knowledge blocks such as "Disk Performance Impact" and "CPU Load Analysis" through chapter segmentation. Specifically, during the segmentation process, the documents are first divided into blocks by chapter, not exceeding the input limit of the large model, and a summary is built for each chapter. Secondly, a summary tree is constructed for the segmented documents, and knowledge blocks are extracted from the summary tree.

[0092] Knowledge blocks are structured: Each knowledge block contains a name (e.g., "Disk I / O anomaly"), content (describes the impact of high disk usage on thread blocking in the DM database), metrics (disk usage, I / O wait time), and analysis steps (calling the iostat tool to check disk status).

[0093] After the offline knowledge extraction module extracts knowledge blocks and operation and maintenance tools, assuming that the extracted operation and maintenance tools are the iostat monitoring tool and the fsck repair tool, the APIs of the iostat monitoring tool and the fsck repair tool are dynamically registered to the domestic intelligent adaptation system based on multi-agent interaction and classified as "monitoring tools" and "repair tools", with accompanying instructions (such as command parameters and output parsing methods).

[0094] S300: The diagnostic prompt generation module selects the knowledge blocks and operation and maintenance tools most relevant to the abnormal signals based on the knowledge blocks and operation and maintenance tools extracted by the offline knowledge extraction module, and generates a prompt template.

[0095] In the above steps, the diagnostic prompt generation module is triggered:

[0096] Knowledge Block Retrieval (BM25 Algorithm): Based on the anomaly indicator "disk usage > 95%", the BM25 algorithm is used to match the knowledge block "disk I / O anomaly" with the highest weight.

[0097] The BM25 algorithm formula is as follows:

[0098] The formula is:

[0099]

[0100] Q is a set of abnormal indicators, which are the relevant indicators of the abnormal situation that needs to be diagnosed.

[0101] Where: f(q,D): the word frequency of the anomaly index q in knowledge block D;

[0102] |D|: The length of the knowledge block, calculated in terms of word count;

[0103] avgdl: The average length of all knowledge blocks in the offline knowledge extraction module;

[0104] The average knowledge block length refers to all knowledge blocks in the knowledge tree before the offline knowledge extraction module performs knowledge block extraction.

[0105] k1 and b are free parameters, and their default values ​​are usually k1 = 1.2 and b = 0.75.

[0106] IDF(q) is the inverse document frequency of the anomaly index q, and its calculation formula is as follows:

[0107]

[0108] Where: N: the total number of knowledge blocks extracted, which is calculated based on all document blocks obtained during the knowledge tree construction phase.

[0109] df(q): The number of documents containing the metric q.

[0110] Operation and maintenance tool matching search:

[0111] Tool matching (Sentence-BERT) model training: The similarity between the anomaly description ("Kylin OS node response latency") and the tool description is calculated using a fine-tuned Sentence-BERT model, prioritizing the matching of the iostat tool's API interface.

[0112] The domain model (i.e., the fine-tuned "Sentence-BERT model") is constructed as follows:

[0113]

[0114] In this embodiment, the Sentence-BERT model is fine-tuned using the cross-entropy loss function to measure the difference between the predicted values ​​and the true values ​​of the classification model (the classification model refers to the Sentence-BERT model used for tool matching, which is fine-tuned based on Sentence-BERT). The formula is as follows:

[0115]

[0116] Among them, y ij This represents the diagnostic context relevance of the operations and maintenance tools, indicating the actual value. A value of 1 indicates relevance to the anomaly description provided by the operations and maintenance tools; otherwise, a value of 0. ij is the probability of anomaly prediction involved in the diagnostic context of the operation and maintenance tool; n is the number of samples, that is, how many anomaly descriptions need to be matched with tools, and these anomaly descriptions constitute the labeled dataset used when fine-tuning the model; m is the number of tool types corresponding to each anomaly description, that is, for each anomaly description, there are m different tools that may need to be matched (including positive sample tools and negative sample tools).

[0117] In this embodiment, the diagnostic context refers to information describing the anomaly or anomaly indicators. It provides specific details of the anomaly in the operation and maintenance scenario, including the anomaly phenomenon and symptoms, to help determine which tools are suitable for the diagnosis and handling of the anomaly.

[0118] The formula for calculating the cosine similarity between the embedding vectors of the computational tool description and the diagnostic context is as follows:

[0119]

[0120] Where emb(.) represents the embedding function of the fine-tuning model. This formula selects the k values ​​with the highest similarity between the diagnostic context s and the diagnostic operation and maintenance tool t, that is, the top k operation and maintenance tools, calculated using the following Top-k formula:

[0121]

[0122] Where m is the total number of tool types corresponding to each diagnostic context.

[0123] The selected top-k operation and maintenance tools are integrated into the prompt word project. Based on the large language model of these operation and maintenance tools, request APIs are generated, and the tool execution results are obtained to enhance root cause diagnosis.

[0124] Based on the above search results, a prompt template is generated, including: an anomaly description (disk alarm trigger time, node ID), a list of available tools (iostat command example), and repair records for similar historical faults.

[0125] S400: The supervisor agent in the multi-agent collaborative diagnosis module assigns the abnormality description in the prompt template to an agent with the corresponding function. The agent with the corresponding function provides the corresponding diagnosis report based on the dynamic decision-making module of tree search.

[0126] In this step, if the diagnostic path of the tree-search-based dynamic decision-making module fails, a reflection mechanism is triggered to backtrack to the parent node and prune invalid paths.

[0127] In this embodiment, the multi-agent collaborative diagnostic module includes a decision coordination agent, an anomaly detection agent, a root cause localization agent, a knowledge reasoning agent, a repair execution agent, an effect verification agent, and a human-computer interaction agent. The decision coordination agent is the supervisor agent, responsible for task scheduling and resource allocation. The anomaly detection agent performs initial anomaly screening. The root cause localization agent performs cross-layer fault correlation analysis and provides repair suggestions. The knowledge reasoning agent dynamically applies domain knowledge. The repair execution agent implements automated procedures and obtains execution results. The effect verification agent confirms the reliability of the repair results and provides a verification report. The human-computer interaction agent visualizes decisions and facilitates human intervention. For example, Figure 1 As shown, the decision coordination agent sends the prompt template to the anomaly detection agent (corresponding to) based on the anomaly description. Figure 1 (Distributing alerts), after the anomaly detection agent performs a preliminary screening of the anomaly description, it publishes the anomaly description to the message queue (corresponding to...). Figure 1 (Regarding the "release event"), the root cause localization agent performs root cause analysis on the anomaly description by initiating a tree search diagnosis and generates a corresponding diagnostic report. The knowledge reasoning agent then challenges the diagnostic report (corresponding to...). Figure 1 ("Knowledge Request"), the root cause localization agent performs another root cause analysis on the question, thereby triggering multiple rounds of diagnosis (corresponding to) Figure 1 ("Return to diagnostic strategy"), thereby generating the final diagnostic report.

[0128] The specific interactions between the various agents are shown in the table below:

[0129]

[0130] Specifically, in this embodiment: the multi-agent collaborative diagnostic mechanism links the supervisor agent to allocate anomaly detection agents, root cause localization agents, and knowledge reasoning agents based on the anomaly type ("disk failure"). The anomaly detection agent publishes the disk detection results to the message queue, and the root cause localization agent subscribes and analyzes the root cause as "disk hardware aging," and verifies it by calling the fsck tool through a knowledge block step. Figure 2 For asynchronous communication processes between multiple agents, such as Figure 2 As shown: After the anomaly detection agent completes the initial screening, it publishes an anomaly description through a message queue; the root cause localization agent subscribes to the message, initiates a tree search, and dynamically selects a path. The knowledge reasoning agent questions the diagnostic report, triggering the root cause localization agent to re-analyze, forming a multi-round loop. After confirming the root cause, the repair execution agent generates operation instructions, and the effect verification agent returns results. The human-computer interaction agent transfers the final report (including repair suggestions) to the operations engineer for review through the work order system, deciding whether to perform automatic repair or manual operation.

[0131] In this step, tree search initialization and path selection include: the anomaly detection agent initializes the diagnostic tree, with the root node being "Execute Disk Detection" and the child nodes being "Detect CPU Load" and "Check Network Latency". The win rate of each node is calculated using the UCT algorithm (70% of historical disk failures), and finally, the root node "Execute Disk Detection" is selected first, thus executing the disk detection path.

[0132] In this step, the dynamically adjusted UCT algorithm formula is calculated as follows:

[0133]

[0134] Among them: w i Node v i Cumulative win rate; n i : Number of visits to node vi; N: Total number of visits to the parent node; c: Exploration weight constant.

[0135] The reflection mechanism is triggered under the following conditions: if the disk detection returns a high degree of abnormality (I / O wait time > 200ms), the CPU detection is skipped, the CPU node is marked as "pruned", and the disk abnormality conclusion is output directly.

[0136] Termination conditions: Under hard termination conditions, if the search depth exceeds 10 layers or the detection agent does not find other abnormal indicators within 30 seconds, or under soft termination conditions, if the confidence increase of 3 consecutive expansions is less than 0.05, the search will be terminated and the diagnostic result (such as "Kylin-OS node disk I / O overload") will be returned.

[0137] S500: An intelligent agent with the ability to raise questions can challenge the diagnostic report, thereby triggering multiple rounds of diagnosis and collaboratively generating the final diagnostic report.

[0138] The knowledge reasoning agent questioned the root cause conclusion ("Could it be a file system misconfiguration?"), triggering a second round of detection, which ultimately confirmed a hardware failure. Finally, the repair execution agent aggregated the outputs of all agents to generate a final report, including:

[0139] Error title: "Kylin OS node disk I / O overload"

[0140] Root cause: Aging disk hardware leads to a surge in I / O wait times.

[0141] Recommended fix: Replace the disk and adjust the log writing strategy (refer to the optimization steps in the knowledge block).

[0142] Explosion radius: Affects the JDBC connection pool of the DM database, increasing the latency of related services X and Y by 20%.

[0143] S600: Based on the final diagnostic report and proposed repair suggestions, the maintenance engineer will reassess the fault situation and determine whether to perform automatic repair or manual repair based on the fault level information.

[0144] In this embodiment, based on the output report and proposed repair suggestions, the operations engineer reassesses the fault situation and, based on information such as the fault level, determines whether to perform automatic repair by the automatic repair agent or manual repair. This multi-agent interaction significantly improves the efficiency of fault diagnosis and reduces labor costs.

[0145] Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

Claims

1. A domestically developed intelligent adaptation system based on multi-agent interaction, characterized in that, The system, based on a large language model and multi-agent collaboration, diagnoses and repairs abnormal indicators, and includes the following modules: The offline knowledge extraction module is used to extract knowledge blocks from unstructured documents and dynamically register the APIs of operation and maintenance tools related to the knowledge blocks into the system. The diagnostic prompt generation module is used to select the knowledge blocks and operation and maintenance tools most relevant to the abnormal indicators from the knowledge blocks extracted by the offline knowledge extraction module and the operation and maintenance tools related to the knowledge blocks, and generate prompt templates based on the most relevant knowledge blocks and operation and maintenance tools selected. The multi-agent collaborative diagnosis module includes a supervisor agent, a detection agent, and other agents. The supervisor agent assigns the detection agent to analyze the anomaly description in the prompt template generated by the diagnosis prompt generation module. Meanwhile, other agents question the analysis results, thereby triggering multiple rounds of diagnosis. The supervisor agent, detection agent, and other agents collaboratively generate a diagnostic report based on asynchronous communication and cross-review mechanisms. The supervisor agent, the detection agent, and other agents are all equipped with a dynamic decision-making module based on tree search. The dynamic decision-making module adopts the Monte Carlo tree search algorithm, combined with the dynamically adjusted UCT algorithm formula to optimize multi-path exploration and backtracking, and improves diagnostic efficiency through a reflection mechanism. The human-machine collaboration module achieves collaboration between manual review and automated repair through a natural language dialogue interface and a work order system based on diagnostic reports generated by the multi-agent collaborative diagnostic module.

2. A domestically developed intelligent adaptation system based on multi-agent interaction according to claim 1, characterized in that, The offline knowledge extraction module includes a document learning unit and a tool unit; The document learning unit is used to split unstructured documents into blocks that do not exceed the input limit of the large language model by chapter, and to build a summary for each chapter, thereby building a summary tree for the document and extracting knowledge blocks from the summary tree. The knowledge blocks include: Name: The topic of the knowledge block; Content: Used to describe the specific content of the knowledge block, using functional description; Metrics: A list of system metrics used for knowledge matching; Analysis steps: The specific steps for conducting analysis using relevant indicators; The tool unit is used to categorize operation and maintenance tools by function, provide usage instructions for each operation and maintenance tool, and dynamically register the APIs of the operation and maintenance tools to the system so that the large language model can call them during diagnosis.

3. A domestically developed intelligent adaptation system based on multi-agent interaction according to claim 1, characterized in that, The diagnostic prompt generation module employs a hybrid retrieval strategy, including knowledge block retrieval and matching retrieval with operation and maintenance tools.

4. A domestically developed intelligent adaptation system based on multi-agent interaction according to claim 3, characterized in that, Knowledge block retrieval uses the BM25 algorithm to retrieve the most relevant knowledge blocks based on anomaly indicators. The formula is as follows: Where f(q,D) is the word frequency of the anomaly indicator q in knowledge block D; |D| represents the length of the knowledge block, calculated in terms of word count; avgdl is the average knowledge block length of all knowledge blocks in the offline knowledge extraction module; k1 and b are free parameters, and their default values ​​are usually k1 = 1.2 and b = 0.

75. IDF(q) is the inverse document frequency of the anomaly index q, and its calculation formula is as follows: Where N is the total number of knowledge blocks extracted; df(q) is the number of documents containing the anomaly indicator q.

5. A domestically developed intelligent adaptation system based on multi-agent interaction according to claim 3, characterized in that, Operation and maintenance tool matching retrieval is based on the fine-tuned Sentence-BERT model. The fine-tuned Sentence-BERT model is used to match the most relevant operation and maintenance tools. The most relevant operation and maintenance tools are selected by calculating the cosine similarity between the embedding vectors of the operation and maintenance tool description and the diagnostic context. The Sentence-BERT model uses the cross-entropy loss function during fine-tuning to measure the difference between the predicted and actual values ​​of the classification model. The formula is as follows: Among them, y ij It is related to the diagnostic context of the operation and maintenance tools, p ij This represents the probability of anomaly prediction related to the diagnostic context of the operation and maintenance tools; n is the number of samples; m is the number of tool types corresponding to each anomaly description. The formula for calculating the cosine similarity between the embedding vectors of the computational tool description and the diagnostic context is as follows: Where emb(.) represents the embedding function of the fine-tuning model. This formula selects the k values ​​with the highest similarity between the diagnostic context s and the diagnostic operation and maintenance tool t, that is, the top k operation and maintenance tools, calculated using the following Top-k formula: The selected top-k operation and maintenance tools are integrated into the prompt word project. Based on the large language model of these operation and maintenance tools, request APIs are generated, and the tool execution results are obtained to enhance root cause diagnosis.

6. A domestically developed intelligent adaptation system based on multi-agent interaction according to claim 1, characterized in that, In the tree-search-based dynamic decision-making module: a dynamically adjusted UCT algorithm is used to search for the node in the tree that best matches the anomaly description. The UCT algorithm formula is as follows: Among them: w i Node v i Cumulative win rate; n i Node v i N: Total number of visits to the parent node; c: Exploration weight constant.

7. A domestically developed intelligent adaptation system based on multi-agent interaction according to claim 1, characterized in that, The multi-agent collaborative diagnostic module includes a decision coordination agent, an anomaly detection agent, a root cause localization agent, a knowledge reasoning agent, a repair execution agent, an effect verification agent, and a human-computer interaction agent; Among them, the decision-making and coordination agent is the supervisory intelligent agent, which is used for task scheduling and resource allocation; Anomaly detection Agent is a detection intelligent agent used for initial screening of anomalies; Root cause localization agent is used for cross-layer fault correlation analysis, knowledge reasoning agent is used for dynamic application of domain knowledge, repair execution agent is used for automated handling implementation, effect verification agent is used for reliable confirmation of repair results, and human-computer interaction agent is used for decision visualization and human intervention.

8. A domestically developed intelligent adaptation system based on multi-agent interaction according to claim 7, characterized in that, The decision coordination agent sends a prompt template to the anomaly detection agent based on the anomaly description. After initial screening of the anomaly description, the anomaly detection agent publishes the anomaly description to the message queue. The root cause localization agent performs root cause analysis on the anomaly description by initiating tree search diagnosis and generates a corresponding diagnostic report. The knowledge reasoning agent raises questions about the diagnostic report, and the root cause localization agent performs root cause analysis again on the questions, thereby triggering multiple rounds of diagnosis to form the final diagnostic report.

9. A domestically developed intelligent adaptation system based on multi-agent interaction according to claim 1, characterized in that, The human-machine collaboration module supports: After manually reviewing the repair suggestions, you can choose to execute them automatically or manually. A plug-and-play framework for domestically developed toolchains.

10. A domestically developed intelligent adaptation method, characterized in that, Based on a large language model and multi-agent collaboration, the implementation includes the following steps: S100: The domestically developed intelligent adaptation system based on multi-agent interaction detected a server anomaly and triggered a multi-agent collaborative diagnostic process. S200: The offline knowledge extraction module extracts relevant knowledge blocks and related operation and maintenance tools based on abnormal signals, and dynamically registers the API of the operation and maintenance tools to the system; S300: The diagnostic prompt generation module selects the knowledge blocks and operation and maintenance tools most relevant to the abnormal signals based on the knowledge blocks and operation and maintenance tools extracted by the offline knowledge extraction module, and generates a prompt template; S400: The supervisor agent in the multi-agent collaborative diagnosis module assigns the abnormality description in the prompt template to an agent with the corresponding function. The agent with the corresponding function provides the corresponding diagnosis report based on the dynamic decision-making module of tree search. In this step, if the diagnostic path of the tree-search-based dynamic decision-making module fails, a reflection mechanism is triggered to backtrack to the parent node and prune invalid paths. S500: An intelligent agent with the ability to raise questions can challenge the diagnostic report, thereby triggering multiple rounds of diagnosis and collaboratively generating the final diagnostic report. S600: Based on the final diagnostic report and proposed repair suggestions, the maintenance engineer will reassess the fault situation and determine whether to perform automatic repair or manual repair based on the fault level information.

Images