基于深度学习优化稀疏射频干扰预测
天体物理仪器与方法
2019-07-17 v1 宇宙学与河外天体物理
摘要
射频干扰（RFI）即使在最偏远的观测地点，也是射电望远镜中始终存在的限制因素。在再电离纪元研究中，为了保留最大灵敏度并减少污染，RFI的识别与去除尤为重要。除了改进RFI识别外，我们还必须考虑RFI识别算法的计算效率，因为诸如氢再电离纪元阵列（HERA）等射电干涉仪阵列的接收机数量不断增长。为此，我们提出了一种深度全卷积神经网络（DFCN），该网络全面利用干涉测量数据，联合使用振幅与相位信息来识别RFI。我们使用包含模拟RFI的HERA可见度数据训练该网络，从而得到已知的“地面真值”数据集，用于评估各种RFI算法的准确性。DFCN模型在HERA-67（67面天线的构建规模）的观测数据上进行了评估，实现了每GPU每小时1.6×10^5个HERA时间序1024通道可见度数据的数据吞吐量。我们确定，相对于仅使用振幅的网络，包含可见度相位增添了重要的相邻时频上下文，从而提高了RFI与非RFI之间的区分能力。在HERA-67观测数据上应用预测时，包含相位的模型达到了0.81的召回率（Recall）、0.58的精确率（Precision）以及0.75的F_2分数。
引用
@article{arxiv.1902.08244,
title = {Optimizing Sparse RFI Prediction using Deep Learning},
author = {Joshua Kerrigan and Paul La Plante and Saul Kohn and Jonathan C. Pober and James Aguirre and Zara Abdurashidova and Paul Alexander and Zaki S. Ali and Yanga Balfour and Adam P. Beardsley and Gianni Bernardi and Judd D. Bowman and Richard F. Bradley and Jacob Burba and Chris L. Carilli and Carina Cheng and David R. DeBoer and Matt Dexter and Eloy de Lera Acedo and Joshua S. Dillon and Julia Estrada and Aaron Ewall-Wice and Nicolas Fagnoni and Randall Fritz and Steve R. Furlanetto and Brian Glendenning and Bradley Greig and Jasper Grobbelaar and Deepthi Gorthi and Ziyaad Halday and Bryna J. Hazelton and Jack Hickish and Daniel C. Jacobs and Austin Julius and Nicholas Kern and Piyanat Kittiwisit and Matthew Kolopanis and Adam Lanman and Telalo Lekalake and Adrian Liu and David MacMahon and Lourence Malan and Cresshim Malgas and Matthys Maree and Zachary E. Martinot and Eunice Matsetela and Andrei Mesinger and Mathakane Molewa and Miguel F. Morales and Tshegofalang Mosiane and Abraham R. Neben and Aaron R. Parsons and Nipanjana Patra and Samantha Pieterse and Nima Razavi-Ghods and Jon Ringuette and James Robnett and Kathryn Rosie and Peter Sims and Craig Smith and Angelo Syce and Nithyanandan Thyagarajan and Peter K. G. Williams and Haoxuan Zheng},
journal= {arXiv preprint arXiv:1902.08244},
year = {2019}
}
备注
11 pages, 7 figures