基于像素偏移值信号增强算法的单脉冲候选体分类方法研究

Single Pulsar Classification Research Based on OffsetHighlights Signal Augmentation Algorithm

时域射电领域观测活动经常产生海量的数据,而随着深度学习技术近几年来的不断发展,以在海量的数据流中识别可能的单脉冲侯选体成为主流方法。相对于传统的脉冲星搜寻方法,使用深度学习领域的技术可以缩短搜寻时间,提高搜寻效率,同时能达到极高的预测精度。本文提出了一种像素偏移值信号增强算法,采用经ImageNet预训练、以EfficientNet-B0为骨干网络的迁移学习模型,可以直接处理来自射电望远镜的原始、高时间分辨率的数据集,不需要消色散,此方法亦可以应用于FRB侯选体的发现。实验证明,在来自FAST官方网站的数据集测试中,该模型可达到92.0%的召回率以及90.1%的精确率。经对比,本文提出的像素偏移值信号增强算法使得模型的Recall提高了45.8%,Precision提高了17.5%,Accuracy提高了26.3%,以及F1-score提高了31.1%。

Observational activities in the field of time-domain radio often generate huge amounts of data.With the continuous development of deep learning techniques in recent years,it has become a mainstream method to develop automated methods to identify possible single-pulse candidates in the massive data stream.Compared with traditional pulsar search methods,using techniques in the field of deep learning can shorten the search time and improve the search efficiency,while achieving extremely high prediction accura-cy.In this research,we present a pixel offset highlight approach which utilize EfficientNet-B0 as the backbone network with a pre-trained model on ImageNet.It can directly handle the raw,high-resolution temporal information from the FAST radio telescope without dispersion,and this method can also be applied to the discovery of FRB candidates.Experimentally,our method is proven to attain a recall rate of 92%and an accuracy rate of 90.1%on FAST datasets,and the OffsetHighlights signal augmentation algorithm proposed in this paper improves the Recall of EfficientNet by 45.8%,Precision by 17.5%,Accuracy by 26.3%,and F1-score by 31.1%.