LangGraph and LLM-based AI Agent for Real-time Anomaly Detection

Recent advancements in Large Language Models (LLM) have drawn significant attention to agent technologies capable of autonomously controlling AI models. However, practical methods for applying such technologies to real-time anomaly detection in manufacturing environments remain limited. This paper presents a LangGraph-based real-time anomaly detection agent that controls two anomaly detection models through natural language. To achieve this, we constructed a custom dataset reflecting real-world environmental conditions and applied training optimization techniques for fine-tuning. The two models were then integrated as nodes within LangGraph and connected to an LLM-based framework, enabling model selection or multi-step reasoning based on text commands. Experimental results demonstrate that both models maintained stable detection performance even in noisy environments, while the LLM provided detailed explanations of the detection results. This AI Agent framework is expected to accelerate AI adoption in manufacturing and contribute to enhanced product quality and productivity.