Hybrid Autoencoder-Isolation Forest approach for time series anomaly detection in C70XP cyclotron operation data at ARRONAX

A research team from the ARRONAX facility and Nantes University has published a novel machine learning method designed to prevent costly failures in medical cyclotron operations. Their paper, "Hybrid Autoencoder-Isolation Forest approach for time series anomaly detection in C70XP cyclotron operation data at ARRONAX," tackles a critical industrial challenge: the C70XP cyclotron used for producing medical radioisotopes is a complex, expensive system where unexpected downtime has significant repercussions for healthcare and research. The team's innovation lies in fusing two established techniques—a fully connected Autoencoder (AE) and the Isolation Forest (IF) algorithm—to create a more sensitive detection system.

While Isolation Forest is known for its effectiveness and scalability in spotting outliers, its reliance on axis-parallel splits limits its ability to identify subtle, hard-to-detect anomalies that occur close to the mean of normal operating data. The hybrid approach overcomes this by first using an Autoencoder to reconstruct sensor measurements. The key metric, the Mean Cubic Error (MCE) between the original and reconstructed data, is then fed into the Isolation Forest model. This processed signal amplifies the subtle deviations that a standalone IF might miss.

The method was validated on real-world proton beam intensity time series data from the ARRONAX cyclotron. The experimental results confirmed a clear improvement in detection performance compared to using Isolation Forest alone. This advancement represents a practical step toward predictive maintenance, allowing operators to identify potential system failures earlier and schedule interventions before a catastrophic breakdown occurs, thereby maximizing uptime for vital radioisotope production.