Applications of AI and Machine Learning to Accelerator Operations: Intelligent Knowledge Synthesis and Predictive Maintenance at the Advanced Photon Source
by
Abstract: Modern particle accelerator facilities face operational challenges in managing institutional knowledge across dispersed data sources and maintaining the reliability of critical infrastructure systems. Artificial intelligence and machine learning techniques offer transformative capabilities to address both challenges, enabling a shift from reactive to proactive facility operations. This work presents two complementary AI/ML-driven systems deployed at the Advanced Photon Source (APS) for intelligent knowledge retrieval and predictive equipment maintenance.
First, we introduce APS-RAG, a domain-aware Retrieval-Augmented Generation system that synthesizes operational intelligence from over thousands of documents spanning electronic logbooks, messaging platforms, technical notes, and maintenance logs. The system employs a specialized query preprocessing pipeline coupled with a hybrid retrieval architecture for comprehensive search and grounded response. Second, we present a real-time predictive maintenance framework currently in testing that integrates low-cost vibration sensors with machine learning models to autonomously detect anomalies in cooling system pumps critical to accelerator operations. Deployed on BLS He and water pumps, the system leverages real-time data collection for condition monitoring and unsupervised deep learning to identify anomalies before they result in unplanned downtime.
Together, these systems demonstrate how modern AI/ML methods — spanning natural language processing, information retrieval, and unsupervised anomaly detection — can be integrated into accelerator operations to preserve institutional knowledge and enhance diagnostic capabilities at large-scale scientific user facilities.