Speaker
Description
This work investigates the effect of different training strategies on the performance of convolutional neural networks (CNNs) for real-time pile-up identification and correction in single particle counters at the GSI Helmholtz Centre for Heavy Ion Research. Building on a previously developed CNN capable of detecting particle pulses without domain-specific knowledge, we examine the influence of different loss functions and their hyperparameters on the network’s ability to accurately localize overlapping events. We also propose new metrics for particle counting and show how to adapt common metrics for precision and recall to allow a user-defined localization tolerance. Using a dataset of approximately 26,000 manually labeled pulses, we analyze how these training choices impact particle counting and localization performance. The findings provide insights for developing particle counting systems suitable for real-time applications, including potential implementation on FPGA hardware.
| Paper status | Resubmitted proceeding files received and assigned to an editor. Accepted. |
|---|