AgenticTwin: An Agentic LLM Framework Integrated with Digital Twin for Anomaly Detection

Digital twins are widely deployed to monitor and simulate cyber-physical systems, but interpreting detected anomalies from raw sensor streams remains exceptionally difficult for human operators. Traditional anomaly detection models flag abnormal states accurately yet provide little insight into root causes, historical context, or actionable mitigation strategies. To bridge this gap, the authors propose AgenticTwin, an agentic framework that replaces monolithic language model prompting with a coordinated team of specialized LLM agents. Operating on structured evidence generated by a hybrid physics-guided and data-driven digital twin, the framework translates abstract anomaly flags into grounded, human-interpretable diagnoses and mitigation plans.

The AgenticTwin pipeline begins with real-time sensor data fed into both a digital twin regression model and an anomaly classifier. The digital twin computes an expected state, and the resulting residual vector enables the classifier to detect and categorize anomalies. Once an anomaly event is structured, it is handed off to an LLM-powered agentic interface coordinated by a Supervisor Agent. The Diagnosis Agent investigates underlying root causes using residual patterns and domain knowledge, the Retrieval Agent surfaces similar historical anomalies from a repository, and the Mitigation Agent recommends corrective actions. This modular task decomposition ensures that reasoning remains tightly grounded in system evidence.

To rigorously assess the framework, the authors introduce a benchmark-oriented evaluation pipeline built by injecting synthetic anomalies—such as spikes and drifts—into a real-world weather sensor dataset. This setup enables controlled, reproducible generation of operator queries covering diverse operational anomaly scenarios. Furthermore, the authors evaluate the feasibility of deploying lightweight, open-source large language models within the framework. Experimental findings indicate that structured agent collaboration significantly enhances diagnosis quality, contextual retrieval precision, and mitigation relevance compared to monolithic LLM baselines, while remaining practical for resource-constrained environments.