Evaluating The Classification Of Fermi Source Using Machine Learning

The theoretical and statistical analysis research on high-energy radiation mechanism and statistical analysis of objects require a large sample of ?-ray sources.So far,the Fermi ?-ray Telescope has discovered 5787 gamma-ray sources,of which mainly belong to active galaxy nuclei(AGN)and pulsars.The classification diversity of ?-ray sources and the massive observations of the source increase the difficulty of data processing.A large number of dark ?-ray sources cannot be identified accurately,thus,the completeness of ?-ray source samples is destroyed.As a consequence,there is an urgent need simple and efficient data mining and data analysis methods.The paper briefly introduced the Ge V-TeV ?-ray observation results,machine learning and Fermi data processing,and used machine learning to evaluate the classification of the ?-ray sources in the fourth Fermi-LAT ?-ray source catalog(4FGL).The specific research work is as follows:(1)The feature parameters of 3207 active galactic nuclei,239 pulsars and 190 indentified ?-ray source of other types were selected from 4FGL using K-S test and random forest(RF)feature importance test methods,And these feature parameters are used for the training,optimization and testing of two unsupervised machine learning(SML)classification models of artificial neural network(ANN)and RF,respectively.Then the models are used to evaluate the classification of 1336 unassociated sources to find possible AGN and pulsar candidates.In order to reduce the classification error caused by the large difference in the number of samples,we first select AGN candidates from 1336 unassociated source,and the model accuracy is about 95%;and then select pulsar candidates from the remaining samples,the model accuracy is about 80%.Considering the classification results of the two models,583 AGN candidates,115 pulsar candidates and 154 other types of ?-ray source candidates are obtained.(2)Based on two SML algorithms of support vector machine(SVM)and logistic regression(LR),SVM and LR classification models are established respectively.Collected 180 high-synchrotron-peaked frequency BL Lacertae objects(HBL)with redshift parameters in the fourth AGN catalog(4LAC)of Fermi-LAT,together with the Radio foundational Source Catalog(RFC)and Gaia Optical source catalog data release two(Gaia-DR2)is cross-matched to obtain radio,optical and Ge V ?-ray band observations.Using 180 HBL with multi-wavelength observations we train and optimize the classification model.According to the evaluation results of the two classification models on the samples,the likelihood probability space is constructed,and 24 possible TeV candidates(PTC)are obtained.Using Fermitools data processing software environment,we analyze the 12-year Fermi average energy spectrum of 24 PTCs,and perform EBL correction on the energy spectrum.The results show that the predicted TeV fluxes of 20 PTCs are above the sensitivity curve of CTA 50 hrs/1yr,while there are only 9 PTCs' predicted fluxes are above the sensitivity curve of LHAASO one year of operation among 14 PTCs whose coordinates are within the field of view of LHAASO.