思维链微调大语言模型赋能网络故障自动诊断OA

In order to solve the problems of difficult configuration and unexecutable output of the universal large language model in network fault diagnosis,this paper proposes an automated network fault diagnosis framework based on Chain-of-Thought(CoT).The framework first integrates network topology,device configuration,logs and fault phenomena into LLM-friendly inputs through a unified YAML serialization method;secondly,it builds a thought chain fine-tuning dataset covering 20 typical faults,using QLoRA(Quantized Low-Rank Adaptation)technology fine-tuned the Qwen3-8B model to enable the model to master hierarchical troubleshooting logic and generate executable solutions;finally,a"diagnosis-execution-verification"closed-loop evaluation system linked with the GNS3 network simulation tool was designed and implemented,with functional repair success rate as the core indicator.Experiments show that the model repair success rate after fine-tuning reaches 82.6%,an increase of more than 54 percentage points from the baseline,proving that this framework can effectively improve the ability of large models to solve actual network failures.