AgenticTwin Enhances Anomaly Detection with LLM-Digital Twin Integration

Digital twins are increasingly vital for monitoring and simulating complex cyber-physical systems, but interpreting anomalies detected within these systems remains a significant challenge, even for skilled operators, due to the sheer volume and complexity of sensor data. While large language models (LLMs) offer strong reasoning and explanation capabilities, their integration into digital twin-driven anomaly analysis has been largely unexplored. Researchers propose AgenticTwin, an agentic framework that bridges this gap by combining LLM-driven reasoning with digital twin technology for anomaly detection. This framework grounds the LLM's explanations in the outputs from a digital twin's anomaly classifier, allowing human operators to interact with the system using natural language to query and understand anomalies. To evaluate AgenticTwin, a benchmark-oriented pipeline was created using synthetic anomalies injected into a real-world weather sensor dataset, facilitating controlled testing of operator queries. The results indicate that structured agent collaboration and knowledge-grounded reasoning significantly improve the quality of anomaly diagnosis, contextual retrieval, and mitigation strategies across various scenarios, even when using lightweight, open-source LLMs.