Method and system for complementing positioning rule based on large model reasoning capability

Abstract

The invention relates to the technical field of operation and maintenance of communication operators, in particular to a method and system for complementing a positioning rule based on large model reasoning ability, and the method comprises the following steps: fault feature and tool preparation, cue word framework construction, thinking process generation and expert auditing, rule generation and current network verification, and feedback optimization mechanism. The method has the beneficial effects that by virtue of the reasoning capability of the Deepseek large model, the rule data of fault diagnosis is generated by continuously adjusting the input conditions such as the fault feature library and the fault diagnosis tool. Supplementing or perfecting a current network fault positioning rule; the problems of high expert dependency, insufficient rule credibility, difficulty in verification and the like in a traditional rule are solved, the time investment of experts is reduced, and the existing rule is perfected; an efficient, credible and iterable fault positioning mechanism is realized, and the intelligent level of network operation and maintenance is improved.

Description

Technical Field

[0001] This invention relates to the field of telecommunications operator operation and maintenance technology, specifically to a method and system for completing positioning rules based on large model reasoning capabilities. Background Technology

[0002] In the telecommunications industry, the large-scale deployment of NFV 5G networks presents both new challenges and transformation opportunities for network operation and maintenance (O&M). On the one hand, the diversification of business scenarios and the increasing complexity of network architecture place higher demands on end-to-end automated O&M, intelligent decision-making systems, and data-driven operations. On the other hand, issues such as fragmented automation capabilities, insufficient adaptation of intelligent technologies, and the disconnect between O&M capabilities and business needs urgently require systematic solutions. Integrating innovative technologies into the O&M process is of great significance for building an efficient and agile network O&M system.

[0003] Faced with massive amounts of heterogeneous data such as alarms, performance data, and logs in the network domain, traditional rule engines and human experience are insufficient to support real-time decision-making in complex scenarios. It is necessary to integrate expert systems with AI technologies such as large language models and fault propagation graphs to build an intelligent analysis engine and achieve closed-loop optimization of fault prediction, root cause localization, and resource scheduling.

[0004] Current methods for locating the root causes of network failures mainly rely on rule bases developed based on expert experience, which presents the following pain points:

[0005] 1. High reliance on experts: Rule generation requires manual analysis of historical fault data, which is time-consuming and cannot cover complex scenarios;

[0006] 2. Insufficient persuasiveness of rules: Manually generated rules lack the presentation of dynamic reasoning processes, making them difficult to explain in the face of novel faults or cross-system correlation problems;

[0007] 3. Low verification efficiency: Rules need to be verified through live network testing, which takes a long time and may cause secondary failures;

[0008] 4. Lagging updates: Once expert knowledge is solidified, it is difficult to adapt to changes in architecture and new types of faults.

[0009] In existing technologies, large model technology still faces significant application bottlenecks. Its application is mostly limited to RAG or generative AI scenarios. There is a gap between it and the decision-making capabilities required for network operation and maintenance. There is an urgent need for a rule completion method that integrates the reasoning capabilities of large models with expert experience to achieve an efficient, reliable, and iterative fault location mechanism. Summary of the Invention

[0010] The purpose of this invention is to provide a method and system for completing localization rules based on large model reasoning capabilities, so as to solve the problems mentioned in the background art.

[0011] To achieve the above objectives, the present invention provides the following technical solution: a method for completing localization rules based on large model reasoning capabilities, comprising the following steps:

[0012] Fault characteristics and tool preparation: Describe the available data types that are relied upon in the fault scenario, including alarms, logs, performance data, and topology data; encapsulate the available toolset in the fault scenario, the toolset source includes the API capabilities of the business system and specific execution scripts, and represent the encapsulated list of available tools in a specific format;

[0013] The prompt word framework is constructed as follows: The roles of the large model are defined, providing fault scenario descriptions, available data, and toolset information; a multi-level inference structure is defined, which includes "fault phenomenon → possible root cause → verification method → ​​location conclusion"; the inference output format and execution steps are defined; and a constraint mechanism is added to control the inference boundary. This constraint mechanism requires the reference of operation and maintenance domain knowledge, including operation and maintenance manuals, resource topology relationships, and definitions of alarms / performance data.

[0014] Thought process generation and expert review: The large model is called to generate candidate inference chains, which include specific execution steps and the thought process for each step; experts intervene to mark logical loopholes in the candidate inference chains, supplement feature data, and make corrections with the help of the description of the inference chains;

[0015] Rule generation and live network verification: Provide a rule generation tool to convert the approved inference chain into executable rules; build an experimental environment to verify the results of the generated rules;

[0016] Feedback optimization mechanism: Compare the model's inference conclusions with the actual root causes of failures and calculate the confidence deviation; when the deviation exceeds the threshold, optimization is triggered according to the scenario. The optimization includes missing data set optimization to supplement missing features to the training set, candidate inference chain optimization to update thinking steps and tool call conditions.

[0017] Preferably, the fault characteristics and tool preparation are as follows:

[0018] The available data types required for fault scenarios are described in detail, clearly defining that they cover alarms, logs, performance data, and topology information, so that these data can be fully utilized during subsequent large-scale model inference. The available toolsets for fault scenarios are encapsulated, and the toolsets come from a wide range of sources, including API capabilities of business systems and specific execution scripts. The encapsulated list of available tools is represented in the format of "toolName:tool name,toolId:tool number", such as "toolName:circuit-related transmission equipment fault check,toolId:101", providing clear identification for tool calls during subsequent inference.

[0019] Preferably, the prompt word framework is constructed as follows:

[0020] Define a clear role for the large model, making it a professional fault root cause localization assistant with rich communication expertise; provide fault scenario descriptions so the large model understands the background and circumstances of the fault; provide available data to ensure the large model has sufficient information to support its inference process; provide toolset information so the large model is aware of the available tools; define a multi-level inference structure to guide the large model's inference in a logical order of "fault phenomenon → possible root cause → verification method → ​​localization conclusion"; define the inference output format as JSON and clarify the relevant requirements for the execution steps; add a constraint mechanism requiring the large model to reference operation and maintenance domain knowledge during the inference process, including operation and maintenance manuals, resource topology relationships, and alarm / performance data definitions, to ensure the accuracy and reliability of the inference results.

[0021] Preferably, the rule generation and live network verification are as follows:

[0022] A rule generation tool is provided, which transforms the expert-reviewed inference chain into executable rules. These executable rules are represented in a specific format, such as "steps":{{"step": step number, "tool":{"toolId": tool number, "toolName": tool name}, "context": context information},{……}}, which includes step information, tool information, and context information, so that relevant operations can be accurately executed according to the rules in the live network environment. An experimental environment is built to verify the generated rules in the experimental environment. By simulating actual fault scenarios or using historical fault data, the effectiveness and accuracy of the rules are checked to ensure that the rules can run normally in the live network environment and accurately locate the root cause of the fault.

[0023] The preferred feedback optimization mechanism is as follows:

[0024] By comparing the model's inference conclusions with the actual root causes of failures, the accuracy of the model's inference is evaluated by calculating the confidence bias. A bias threshold is set, and when the calculated confidence bias exceeds this threshold, optimization operations are triggered according to the scenario. The optimization operations include dataset missing feature optimization, that is, when it is found that there are missing features in the dataset that cause inference bias, the missing features are added to the training set to improve the model's training effect; and candidate inference chain optimization, that is, updating the thinking steps and tool calling conditions in the candidate inference chain, correcting the unreasonable parts in the inference process, making the inference chain more complete and accurate, thereby continuously improving the quality and reliability of the localization rules based on the inference capabilities of the large model.

[0025] A system for a method to complete localization rules based on large model reasoning capabilities includes:

[0026] Fault Characteristics and Tool Preparation Module: This module describes the available data types required for fault scenarios, including alarms, logs, performance data, and topology data. It also encapsulates the available toolsets for fault scenarios, which are derived from the API capabilities of business systems and specific execution scripts, and presents the encapsulated list of available tools in a specific format.

[0027] The prompt word framework construction module is used to define the roles of the large model, provide fault scenario descriptions, available data, and toolset information; define a multi-level inference structure, which is "fault phenomenon → possible root cause → verification method → ​​location conclusion"; define the inference output format as JSON and clarify the relevant requirements for execution steps; add a constraint mechanism to control the inference boundary, which requires the reference of operation and maintenance domain knowledge, including operation and maintenance manuals, resource topology relationships, and alarm / performance data definitions.

[0028] The thought process generation and expert review module is used to call the Deepseek large model to generate candidate inference chains. The candidate inference chain contains specific execution steps and the thought process for each step. It provides an expert intervention interface so that experts can mark logical loopholes in the candidate inference chain, supplement feature data, and make corrections with the help of the description of the inference chain.

[0029] Rule generation and live network verification module: This module provides a rule generation tool to convert the approved inference chain into executable rules; it also builds an experimental environment to verify the generated rules.

[0030] Feedback optimization mechanism module: used to compare the model's inference conclusions with the actual root causes of failures and calculate the confidence deviation; when the deviation exceeds the threshold, optimization is triggered according to the scenario. The optimization includes missing dataset optimization to supplement missing features to the training set, candidate inference chain optimization to update thinking steps and tool call conditions.

[0031] Preferably, the fault characteristics and tool preparation module specifically includes:

[0032] The data type description unit is used to describe in detail the available data types that are relied upon in the fault scenario, and to clarify that the data types include alarms, logs, performance and topology information, so as to provide a data foundation for subsequent large model inference;

[0033] The toolset encapsulation unit is used to encapsulate the toolsets available in fault scenarios. The toolsets are derived from the API capabilities of the business system and the specific execution scripts. The encapsulated list of available tools is represented in the format of "toolName:tool name,toolId:tool number", such as "toolName:circuit-related transmission equipment fault check,toolId:101", so as to accurately call the tools in the subsequent reasoning process.

[0034] Preferably, the prompt word framework construction module specifically includes:

[0035] The role setting unit is used to assign the role of a professional fault root cause localization assistant to the large model, so that it has rich communication expertise to better perform fault reasoning.

[0036] The information providing unit is used to provide fault scenario descriptions, available data, and toolset information, so that the large model can fully understand the fault background and available resources.

[0037] The reasoning structure definition unit is used to define the multi-level reasoning structure as "fault phenomenon → possible root cause → verification method → ​​location conclusion", guiding the large model to reason in a logical order.

[0038] The output format and step definition unit is used to define the inference output format as JSON and to clarify the relevant requirements for the execution steps, ensuring that the inference results are standardized and executable.

[0039] The constraint mechanism addition unit is used to add constraint mechanisms, requiring the large model to reference operation and maintenance domain knowledge during the inference process, including operation and maintenance manuals, resource topology relationships, and alarm / performance data definitions, in order to improve the accuracy and reliability of inference.

[0040] Preferably, the rule generation and live network verification module specifically includes:

[0041] The rule generation tool unit provides a rule generation tool that can convert expert-reviewed inference chains into executable rules. The executable rules are represented in a specific format, such as "steps":{{"step": step number, "tool":{"toolId": tool number, "toolName": tool name}, "context": context information},{……}}, containing step information, tool information, and context information, so as to be accurately executed in the live network environment.

[0042] The experimental environment construction unit is used to build an experimental environment to verify the generated rules. By simulating actual fault scenarios or using historical fault data, the effectiveness and accuracy of the rules are checked to ensure that the rules can run normally in the live network environment and accurately locate the root cause of the fault.

[0043] Preferably, the feedback optimization mechanism module specifically includes:

[0044] The bias calculation unit is used to compare the model's inference conclusions with the actual root causes of failures, and to evaluate the accuracy of the model's inference by calculating the confidence bias.

[0045] The optimization trigger unit is used to set a deviation threshold. When the calculated confidence deviation exceeds the threshold, the optimization operation is triggered according to the scenario.

[0046] The dataset optimization unit is used to supplement the training set with missing features when it is found that missing features in the dataset cause inference bias, so as to improve the training effect of the model.

[0047] The reasoning chain optimization unit is used to update the thinking steps and tool calling conditions in the candidate reasoning chain, correct unreasonable aspects in the reasoning process, make the reasoning chain more complete and accurate, and thus continuously improve the quality and reliability of the localization rules based on the reasoning ability of the large model.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] This invention proposes a method and system for supplementing fault location rules based on the reasoning capabilities of a large model. By leveraging the reasoning power of the Deepseek large model and continuously adjusting input conditions such as fault feature databases and fault diagnosis tools, it generates rule data for fault diagnosis. This supplements or improves existing network fault location rules; it addresses the problems of high expert dependence, insufficient rule credibility, and difficulty in verification in traditional rules, reducing the time investment of experts and improving existing rules; and it achieves an efficient, reliable, and iterative fault location mechanism, thereby enhancing the intelligence level of network operation and maintenance. Attached Figure Description

[0050] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the present invention clear and complete, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only some, not all, embodiments of the present invention, and are merely illustrative of the embodiments of the present invention. They are not intended to limit 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.

[0052] Example 1, please refer to Figure 1 This invention provides a technical solution: a method for completing localization rules based on large model reasoning capabilities, comprising the following steps:

[0053] A method for completing localization rules based on large model reasoning capabilities includes five stages: fault feature and tool preparation, prompt word framework construction, thought process generation and expert review, rule generation and live network verification, and feedback optimization mechanism.

[0054] a) Fault Characteristics and Tool Preparation: Describe the available data types required for the fault scenario, including alarms, logs, performance data, topology data, etc.; encapsulate the available toolset for the fault scenario, which may originate from API capabilities of business systems, specific execution scripts, etc.; taking base station outage as an example, available tools include query capabilities, analysis capabilities, resource queries, etc.; the encapsulated list of available tools is as follows.

[0055] "toolName: Circuit-related transmission equipment fault check, toolId: 101"

[0056] "toolName: Optical Path Associated Transmission Equipment Fault Check, toolId:102"

[0057] "toolName: Troubleshooting capabilities for device alarms and faults, toolId: 106"

[0058] "toolName: Verify business project status, toolId: 108"

[0059] b) Constructing the prompt word framework: setting the roles of the large model, providing information such as fault scenario descriptions, available data, and toolsets; defining a multi-level inference structure (such as "fault phenomenon → possible root cause → verification method → ​​location conclusion"), defining the inference output format and execution steps, etc.; adding constraint mechanisms to control the inference boundary, such as referencing knowledge from the operation and maintenance domain, including, for example, operation and maintenance manuals, resource topology relationships, and definitions of alarm / significant data.

[0060] Example of prompt word construction:

[0061] You are a professional communication fault root cause locator with extensive communication expertise. You can supplement your knowledge with base station outage fault diagnosis scenarios and manufacturer maintenance manuals. Currently, comprehensive fault information has been collected; you need to use the following tools to check for any issues.

[0062] "toolName: Circuit-related transmission equipment fault check, toolId: 101"

[0063] "toolName: Optical Path Associated Transmission Equipment Fault Check, toolId:102"

[0064] "toolName: Troubleshooting capabilities for device alarms and faults, toolId: 106"

[0065] "toolName: Verify business project status, toolId: 108"

[0066] Please output the complete steps for analyzing the root cause of the failure in JSON format. This JSON will be used to execute the tool; it should include elements such as step information, tool, and context.

[0067] c) Generation and expert review of the thought process: The Deepseek large model is used to generate candidate inference chains. Experts intervene to mark logical flaws and supplement feature data. The candidate inference chains contain specific execution steps and the thought process behind each step. Experts can use the description of the inference chains to make corrections.

[0068]

[0069] Okay, now I need help pinpointing the cause of the fault. First, I should start by checking basic equipment alarms. Use tool 106 to check for any abnormal alarms on my device, such as power issues or hardware malfunctions. Next, check if the transmission link is working properly. Use tools 101 and 102 to check the transmission equipment associated with the circuitry and optical paths, ensuring there are no transmission interruptions. Then…

[0070] By following the steps above to troubleshoot step by step, the root cause of the malfunction can be systematically identified.

[0071]

[0072] d) Rule generation and live network verification: Provide rule generation tools to convert the reviewed inference chain into executable rules; build an experimental environment to verify the generated rules.

[0073] The format of the rule generation based on model inference is as follows:

[0074] "steps":{

[0075] {"step":1,"tool":{"toolId":106."toollame":"Independent device alarm and fault diagnosis capability"},context":"Firstly, check its own alarm and response information to confirm whether there are any device-level anomalies"},

[0076] {"step":2,"tool":{"toolId":103,"toollame":"Uplink BRAS Troubleshooting"},"context":"Check the status of the uplink BRAS to confirm whether there are connection problems or configuration errors to the core network."},

[0077] {……}

[0078] }

[0079] e) Feedback optimization mechanism: compare the model's inference conclusions with the actual root causes of failures and calculate the confidence deviation; if the deviation exceeds the threshold, optimization is triggered according to the scenario, such as supplementing missing features to the training set for missing datasets, updating the thinking steps and calling tool conditions for candidate inference chains, etc.

[0080] Example 2, based on Example 1, proposes a system for completing localization rules based on large model reasoning capabilities, including:

[0081] The Fault Characteristics and Tool Preparation Module describes the available data types required for fault scenarios, including alarms, logs, performance data, and topology data. It encapsulates the available toolsets for fault scenarios, sourced from the API capabilities of business systems and specific execution scripts, and presents the encapsulated list of available tools in a specific format. Specifically, it includes: a data type description unit, which provides a detailed description of the available data types required for fault scenarios, specifying that the data types include alarms, logs, performance data, and topology information, providing a data foundation for subsequent large-scale model inference; and a toolset encapsulation unit, which encapsulates the available toolsets for fault scenarios, sourced from the API capabilities of business systems and specific execution scripts. The encapsulated list of available tools is presented in the format "toolName:tool name,toolId:tool number", for example, "toolName:Circuit-related Transmission Equipment Fault Verification,toolId:101", to ensure accurate tool invocation during subsequent inference.

[0082] The prompt word framework construction module is used to define the roles of the large model, provide fault scenario descriptions, available data, and toolset information; define a multi-level inference structure, which is "fault phenomenon → possible root cause → verification method → ​​location conclusion"; define the inference output format as JSON and clarify the relevant requirements for execution steps; add a constraint mechanism to control the inference boundary, which requires the reference of operation and maintenance domain knowledge, including operation and maintenance manuals, resource topology relationships, and alarm / performance data definitions; specifically, it includes: a role setting unit, used to set the role of a professional fault root cause location assistant for the large model, enabling it to have rich communication expertise to better perform fault inference; and an information providing unit. The system provides fault scenario descriptions, available data, and toolset information, allowing the large model to fully understand the fault background and available resources. The inference structure definition unit defines a multi-level inference structure as "fault phenomenon → possible root cause → verification method → ​​location conclusion," guiding the large model to reason in logical order. The output format and step definition unit defines the inference output format as JSON and clarifies the relevant requirements for execution steps, ensuring the inference results are standardized and executable. The constraint mechanism addition unit adds constraint mechanisms, requiring the large model to reference operational domain knowledge during inference, including operational manuals, resource topology relationships, and alarm / performance data definitions, to improve the accuracy and reliability of inference.

[0083] The thought process generation and expert review module is used to call the Deepseek large model to generate candidate inference chains. The candidate inference chain contains specific execution steps and the thought process for each step. It provides an expert intervention interface so that experts can mark logical loopholes in the candidate inference chain, supplement feature data, and make corrections with the help of the description of the inference chain.

[0084] Rule generation and live network verification module: This module provides rule generation tools to convert approved inference chains into executable rules; it also builds experimental environments to verify the generated rules. Specifically, it includes:

[0085] The rule generation tool unit provides a rule generation tool that can convert expert-reviewed inference chains into executable rules. These executable rules are represented in a specific format, such as "steps":{{"step": step number, "tool":{"toolId": tool number, "toolName": tool name}, "context": context information},{……}}, containing step information, tool information, and context information to ensure accurate execution in the live network environment. The experimental environment construction unit is used to build an experimental environment to verify the generated rules. By simulating actual fault scenarios or using historical fault data, the effectiveness and accuracy of the rules are checked to ensure they can operate normally in the live network environment and accurately locate the root cause of the fault.

[0086] Feedback optimization mechanism module: used to compare the model's inference conclusions with the actual root causes of failures and calculate the confidence deviation; when the deviation exceeds the threshold, optimization is triggered according to the scenario. The optimization includes missing dataset optimization to supplement missing features to the training set, candidate inference chain optimization to update thinking steps and tool call conditions.

[0087] Specifically include:

[0088] The bias calculation unit is used to compare the model's inference conclusions with the actual root causes of failures, and to evaluate the accuracy of the model's inference by calculating the confidence bias.

[0089] The optimization trigger unit is used to set a deviation threshold. When the calculated confidence deviation exceeds the threshold, the optimization operation is triggered according to the scenario.

[0090] The dataset optimization unit is used to supplement the training set with missing features when it is found that missing features in the dataset cause inference bias, so as to improve the training effect of the model.

[0091] The reasoning chain optimization unit is used to update the thinking steps and tool calling conditions in the candidate reasoning chain, correct unreasonable aspects in the reasoning process, make the reasoning chain more complete and accurate, and thus continuously improve the quality and reliability of the localization rules based on the reasoning ability of the large model.

[0092] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for completing localization rules based on large model reasoning capabilities, characterized in that: Includes the following steps: Fault characteristics and tool preparation: Describe the available data types that are relied upon in the fault scenario, including alarms, logs, performance data, and topology data; encapsulate the available toolset in the fault scenario, the toolset source includes the API capabilities of the business system and specific execution scripts, and represent the encapsulated list of available tools in a specific format; The prompt word framework is constructed as follows: The roles of the large model are defined, providing fault scenario descriptions, available data, and toolset information; a multi-level inference structure is defined, which includes "fault phenomenon → possible root cause → verification method → ​​location conclusion"; the inference output format and execution steps are defined; and a constraint mechanism is added to control the inference boundary. This constraint mechanism requires the reference of operation and maintenance domain knowledge, including operation and maintenance manuals, resource topology relationships, and definitions of alarms / performance data. Thought process generation and expert review: The large model is called to generate candidate inference chains, which include specific execution steps and the thought process for each step; experts intervene to mark logical loopholes in the candidate inference chains, supplement feature data, and make corrections with the help of the description of the inference chains; Rule generation and live network verification: Provide a rule generation tool to convert the approved inference chain into executable rules; build an experimental environment to verify the results of the generated rules; Feedback optimization mechanism: Compare the model's inference conclusions with the actual root causes of failures and calculate the confidence deviation; when the deviation exceeds the threshold, optimization is triggered according to the scenario. The optimization includes missing data set optimization to supplement missing features to the training set, candidate inference chain optimization to update thinking steps and tool call conditions.

2. The method for completing localization rules based on large model inference capability according to claim 1, characterized in that: The specific fault characteristics and tool preparation are as follows: The available data types required for fault scenarios are described in detail, clearly defining that they cover alarms, logs, performance data, and topology information, so that these data can be fully utilized during subsequent large-scale model inference. The available toolsets for fault scenarios are encapsulated, and the toolsets come from a wide range of sources, including API capabilities of business systems and specific execution scripts. The encapsulated list of available tools is represented in the format "toolName:tool name,toolId:tool number", such as "toolName:circuit-related transmission equipment fault check,toolId:101", providing clear identification for tool calls during subsequent inference.

3. The method for completing localization rules based on large model inference capability according to claim 2, characterized in that: The specific construction of the prompt word framework is as follows: Define a clear role for the large model, enabling it to act as a professional fault root cause localization assistant with rich communication expertise; provide fault scenario descriptions so that the large model understands the background and circumstances of the fault. Provide available data to ensure the large model has sufficient information to support its inference process; provide toolset information so the large model is aware of the available tools; define a multi-level inference structure to guide the large model's inference in a logical order of "fault phenomenon → possible root cause → verification method → ​​location conclusion"; define the inference output format as JSON and clarify the relevant requirements for the execution steps; add a constraint mechanism that requires the large model to reference knowledge from the operations and maintenance domain during the inference process, including operations and maintenance manuals, resource topology relationships, and definitions of alarms / performance data, to ensure the accuracy and reliability of the inference results.

4. The method for completing localization rules based on large model reasoning ability according to claim 3, characterized in that: Rule generation and live network verification are as follows: A rule generation tool is provided, which transforms the expert-reviewed inference chain into executable rules. The executable rules are represented in a specific format, such as "steps":{{"step": step number,"tool":{"toolId": tool number,"toolName": tool name},"context": context information},{……}}, which contains step information, tool information, and context information, so that relevant operations can be accurately executed according to the rules in the live network environment. Build an experimental environment to verify the generated rules. Test the rules by simulating real-world fault scenarios or using historical fault data to check their effectiveness and accuracy, ensuring that the rules can run normally in the live network environment and accurately locate the root cause of the fault.

5. The method for completing localization rules based on large model reasoning ability according to claim 4, characterized in that: The feedback optimization mechanism is as follows: The accuracy of model inference is evaluated by comparing the model's inference conclusions with the actual root causes of failures and calculating the confidence bias. A bias threshold is set, and when the calculated confidence bias exceeds the threshold, optimization operations are triggered according to the scenario. The optimization operations include dataset missing feature optimization, that is, when it is found that there are missing features in the dataset that cause inference bias, the missing features are added to the training set to improve the training effect of the model. Candidate inference chain optimization involves updating the thinking steps and tool invocation conditions in the candidate inference chain, correcting unreasonable aspects in the inference process, making the inference chain more complete and accurate, and thus continuously improving the quality and reliability of the localization rules based on the large model inference capability.

6. A system for using the method of completing localization rules based on large model reasoning ability according to claim 5, characterized in that: include: Fault Characteristics and Tool Preparation Module: This module describes the available data types that are required in fault scenarios. These data types include alarms, logs, performance data, and topology data. The toolset available in fault scenarios is encapsulated. The toolset comes from the API capabilities of the business system and specific execution scripts, and the encapsulated list of available tools is represented in a specific format. The prompt word framework construction module is used to define the roles of the large model, provide fault scenario descriptions, available data, and toolset information; define a multi-level inference structure, which is "fault phenomenon → possible root cause → verification method → ​​location conclusion"; define the inference output format as JSON and specify the relevant requirements for execution steps; add a constraint mechanism to control the inference boundary, which requires the reference of operation and maintenance domain knowledge, including operation and maintenance manuals, resource topology relationships, and alarm / performance data definitions; The thought process generation and expert review module is used to call the Deepseek large model to generate candidate inference chains. The candidate inference chain contains specific execution steps and the thought process for each step. It provides an expert intervention interface so that experts can mark logical loopholes in the candidate inference chain, supplement feature data, and make corrections with the help of the description of the inference chain. Rule generation and live network verification module: This module provides a rule generation tool to convert the approved inference chain into executable rules; it also builds an experimental environment to verify the generated rules. Feedback optimization mechanism module: used to compare the model's inference conclusions with the actual root causes of failures and calculate the confidence bias; When the deviation exceeds the threshold, optimization is triggered according to the scenario. The optimization includes missing dataset optimization to supplement missing features to the training set, candidate inference chain optimization to update thinking steps and tool call conditions.

7. The system according to claim 6, characterized in that: The fault characteristics and tool preparation module specifically includes: The data type description unit is used to describe in detail the available data types that are relied upon in the fault scenario, and to clarify that the data types include alarms, logs, performance and topology information, so as to provide a data foundation for subsequent large model inference; The toolset encapsulation unit is used to encapsulate the toolsets available in fault scenarios. The toolsets are derived from the API capabilities of the business system and the specific execution scripts. The encapsulated list of available tools is represented in the format "toolName:tool name,toolId:tool number", such as "toolName:circuit-related transmission equipment fault check,toolId:101", so as to accurately call the tools in the subsequent reasoning process.

8. A system according to claim 6, characterized in that: The prompt word framework construction module specifically includes: The role setting unit is used to assign the role of a professional fault root cause localization assistant to the large model, so that it has rich communication expertise to better perform fault reasoning. The information providing unit is used to provide fault scenario descriptions, available data, and toolset information, so that the large model can fully understand the fault background and available resources. The reasoning structure definition unit is used to define a multi-level reasoning structure as "fault phenomenon → possible root cause → verification method → ​​location conclusion", guiding the large model to reason in a logical order. The output format and step definition unit is used to define the inference output format as JSON and to clarify the relevant requirements for the execution steps, ensuring that the inference results are standardized and executable. The constraint mechanism addition unit is used to add constraint mechanisms, requiring the large model to reference operation and maintenance domain knowledge during the inference process, including operation and maintenance manuals, resource topology relationships, and alarm / performance data definitions, in order to improve the accuracy and reliability of inference.

9. A system according to claim 6, characterized in that: The rule generation and live network verification module specifically includes: The rule generation tool unit provides a rule generation tool that can convert expert-reviewed inference chains into executable rules. The executable rules are represented in a specific format, such as "steps":{{"step": step number,"tool":{"toolId": tool number,"toolName": tool name},"context": context information},{……}}, which includes step information, tool information, and context information to ensure accurate execution in the live network environment. The experimental environment construction unit is used to build an experimental environment to verify the generated rules. By simulating actual fault scenarios or using historical fault data, the effectiveness and accuracy of the rules are checked to ensure that the rules can run normally in the live network environment and accurately locate the root cause of the fault.

10. A system according to claim 6, characterized in that: The feedback optimization mechanism module specifically includes: The bias calculation unit is used to compare the model's inference conclusions with the actual root causes of failures, and to evaluate the accuracy of the model's inference by calculating the confidence bias. The optimization trigger unit is used to set a deviation threshold. When the calculated confidence deviation exceeds the threshold, the optimization operation is triggered according to the scenario. The dataset optimization unit is used to supplement the training set with missing features when it is found that missing features in the dataset cause inference bias, so as to improve the training effect of the model. The reasoning chain optimization unit is used to update the thinking steps and tool calling conditions in the candidate reasoning chain, correct unreasonable aspects in the reasoning process, make the reasoning chain more complete and accurate, and thus continuously improve the quality and reliability of the localization rules based on the reasoning ability of the large model.

Images