BEACON: Benchmarking Multimodal AI for Explainable Real-Time Anomaly Detection in National Critical Infrastructure

The United States depends on its critical infrastructure systems including power grids and transportation networks and emergency response systems for both national security and daily operations. The systems face rising threats from cyberattacks and natural disasters and extensive failures which spread rapidly between different areas. The current monitoring systems face challenges in detecting early warning signs because they depend on single data types and operate offline and generate complex results that human operators struggle to understand. Researchers and practitioners can develop real-time problem detection tools through BEACON which serves as a benchmark system. The BEACON system integrates three types of event-related data can be integrated for analysis: Sensor data from energy load and traffic flow sources. Geospatial information from maps and hazard layers. Text alerts from trusted U.S. agencies including the National Weather Service. BEACON enables researchers to analyze basic sensor anomalies and complex multi-source patterns through its data stream integration. The dataset includes a reference pipeline which combines machine learning for anomaly detection with large language models (LLMs) to create brief structured explanations. The evaluation assesses system performance through three key metrics which include detection precision and streaming performance speed and explanation clarity for human evaluators. The combination of different data sources decreases false alarm rates while LLM explanations with constraints produce better results than statistical methods for anomaly understanding. The open release of BEACON enables researchers to evaluate and enhance methods which enhance U.S. critical infrastructure resilience.