DEEP CONTINUAL LEARNING FOR ANOMALY DETECTION IN PARTICLE PHYSICS EXPERIMENTS

Data quality monitoring (DQM) is the process of identifying usable data in par- ticle physics experiments. Automation of the DQM process can lead to substan- tially improved data quality with considerably reduced human effort. This work presents a deep continual learning approach for automatic DQM. We introduce the Continual Shifting Transformer (CST), a Transformer-based model designed for anomaly detection in shifting data distributions. The continual learning method is interpretable, adaptable, and significantly surpasses competing algorithms on key metrics. The code is available at: https://github.com/eiriksteen/deep-DQM.