Recurrent neural networks for anomaly detection in magnet power supplies of particle accelerators

This research illustrates how time-series forecasting employing recurrent neural networks (RNNs) can be used for anomaly detection in particle accelerators—complex machines that accelerate elementary particles to high speeds for various scientific and industrial applications. Our approach utilizes an RNN to predict temperatures of key components of magnet power supplies (PSs), which can number up to thousands in an accelerator. An anomaly is declared when the predicted temperature deviates significantly from observation. Our method can help identify a PS requiring maintenance before it fails and leads to costly downtime of an entire billion-dollar accelerator facility. We demonstrate that the RNN outperforms a reasonably complex physics-based model at predicting the PS temperatures and at anomaly detection. We conclude that for practical applications it can be beneficial to use RNNs instead of increasing the complexity of the physics-based model. We chose the long short-term memory (LSTM) as opposed to other RNN cell structures due to its widespread use in time-series forecasting and its relative simplicity. However, we demonstrate that the LSTM’s precision of predicting PS temperatures is nearly on par with measurement precision, making more complex or custom architectures unnecessary. Lastly, we dedicate a section of this paper to presenting a proof-of-concept for using infrared cameras for spatially-resolved anomaly detection inside power supplies, which will be a subject of future research. • Particle accelerators rely on magnet power supplies, which can number up to thousands • Pre-failure power supplies can be detected by anomaly detection in their temperatures • RNN-based anomaly detection outperforms a reasonably complex physics-based method • Using RNNs can be more practical than adding complexity to the physics-based model • Infrared cameras can be used to localize the anomalies within each power supply