Single Pulse Candidate Identification Based on Machine Learning
Large scale radio astronomical surveys often generate massive amounts of data,and conducting single pulse searches on them will yield billions of single pulse candidate signals,including many non periodic pulse signals that cannot be found through periodic searches.The 500 m Aperture Spherical Radio Telescope(FAST)for simultaneous scanning of multiple scientific targets(CRAFTS)generates a large amount of data.CRAFTS data has less signal volume,interferes with other irrelevant signals,and takes a huge amount of time to screen one by one.In addition,existing AI models have low accuracy in identifying CRAFTS pulse signals.On the basis of FETCH,DenseNet121 and Xception networks are used to train and optimize existing models,improve the recognition rate of AI models for weak signals,reduce the impact of interference on classifiers when interference and pulse signals coexist,and apply them to the single pulse screening process.The results showed that using DenseNet121 and Xception networks to train optimization models improved the accuracy of weak signal recognition and to some extent eliminated the influence of interference.In datasets with weak signals,interference,and signals coexisting,it showed a recall rate of 77.78%,a false positive rate of 1.50%,and an accuracy rate of 88.49%.
single pulse searchpulsarsignal-to-noise ratiomachine learning
NETL
NSTL
万方数据
