Robust Errant Beam Prognostics with Conditional Modeling for Particle Accelerators

TL;DR

This work tackles predicting errant beam pulses in the SNS accelerator to reduce downtime. It compares conditional Siamese Neural Networks (CSNN) and Conditional Variational Autoencoders (CVAE), using beam-configuration conditioning to handle data shifts across eight configurations, with CSNN showing superior performance. CSNN outperforms both single-configuration SNNs and CVAE, achieving higher TPR at a fixed FPR and using considerably fewer parameters, demonstrating practical gains for prognostics in complex accelerator systems. The results indicate that conditional, supervised discriminative models can provide robust, configuration-aware anomaly predictions that enhance accelerator availability, with future work aimed at uncertainty quantification and continual learning to address drift and aging components.

Abstract

Particle accelerators are complex and comprise thousands of components, with many pieces of equipment running at their peak power. Consequently, particle accelerators can fault and abort operations for numerous reasons. These faults impact the availability of particle accelerators during scheduled run-time and hamper the efficiency and the overall science output. To avoid these faults, we apply anomaly detection techniques to predict any unusual behavior and perform preemptive actions to improve the total availability of particle accelerators. Semi-supervised Machine Learning (ML) based anomaly detection approaches such as autoencoders and variational autoencoders are often used for such tasks. However, supervised ML techniques such as Siamese Neural Network (SNN) models can outperform unsupervised or semi-supervised approaches for anomaly detection by leveraging the label information. One of the challenges specific to anomaly detection for particle accelerators is the data's variability due to system configuration changes. To address this challenge, we employ Conditional Siamese Neural Network (CSNN) models and Conditional Variational Auto Encoder (CVAE) models to predict errant beam pulses at the Spallation Neutron Source (SNS) under different system configuration conditions and compare their performance. We demonstrate that CSNN outperforms CVAE in our application.

Figures (8)

• Figure 1: Differential Current Monitor (DCM) along with beam demonstration
• Figure 2: Beam macro-pulse pattern and beam current waveform showing the mini-pulses
• Figure 3: The SNN model architecture comprises two input layers for the reference waveforms and the acquired waveform, followed by one common ResNet model and a distance function. Then a MLP block with dropout layers is applied to avoid over-fitting, and finally, an output layer is used to compress the dimensionality and produce the similarity scalar.
• Figure 4: The CSNN model architecture, similar to the SNN model with the addition of a third input for beam configurations and an MLP block for the embedding of the configuration vectors. The configuration embeddings are then concatenated with the latent difference before passing to the common MLP block followed by the output layer.
• Figure 5: TPR at FPR of 0.1% from various architectures of CSNN model during HPO and NAS