Conceptual

Deep-Learning Multivariate Time-Series Prediction for Accelerator Beam-Outage Forecasting

An AI framework that replaces a reactive threshold-based alarm system at the Fermilab accelerator complex with predictive analytics: it benchmarks recurrent, attention-based, and linear deep-learning architectures for forecasting beam outages from multivariate sensor time series (2,703 Linac devices, 80 labeled outages) and adds a Random Forest classifier that emits consistent, confidence-scored outage-cause labels, characterizing each architecture's strengths and the data/modeling gaps blocking reliable predictive downtime reduction.