Abstract
In nuclear spectroscopy, a physical phenomenon known as the pile-up effect distorts direct measurements by causing temporal overlap of detector pulses. Existing deep learning-based pile-up correction methods rely heavily on supervised training with simulated data, which often generalize poorly to real measurements due to simulation-experiment discrepancies. In this work, we propose a contrastive learning framework to learn robust and transferable representations directly from large-scale unlabeled real nuclear pulse signals. The detector output is segmented into physically complete pulse aggregations using a zero-crossing-based strategy, which serve as semantically coherent instances for representation learning. Physics-inspired data augmentations are designed to realistically model detector noise and bandwidth effects while preserving pulse area. A one-dimensional ResNet encoder is employed for efficient representation learning. The learned representations are transferred to pile-up identification and counting-rate estimation tasks. Experimental results on real nuclear radiation detection systems demonstrate that our method achieves strong performance and robustness under high counting-rate conditions, with particularly pronounced advantages in challenging peak pile-up scenarios.