ARRONAX (Accelerator for Research in Radiochemistry and Oncology at Nantes Atlantique) is a multi-particle cyclotron capable of delivering high-intensity (2 × 375 μA) and high-energy (70 MeV) proton beams, ensuring precise delivery of ion beams to the target with the required energy and beam characteristics. However, due to the system’s complexity, anomalies may occur, compromising reliability and causing costly interruptions. This study aims to detect such anomalies at an early stage by analyzing beam intensity time series data, using a comparative approach across various anomaly detection techniques. These include traditional machine learning methods like One-Class SVM and Isolation Forest, as well as neural network-based models, with a focus on autoencoders. The performance of each model is assessed using standard metrics, particularly the Area Under the Precision-Recall Curve (AUC-PR). Experimental findings show that neural network-based models consistently outperform classical approaches, likely due to their superior ability to capture complex temporal patterns in the data.
