A significant portion of the process of detecting pulsars from radio sky surveys remains a largely manual task. The visual inspection of data in order to detect and validate potential pulsar candidates is by far the most time consuming portion of the overall process. Coupled with the fact that well over a Petabyte of pulsar survey data has been archived, the task of identifying these valuable phenomena is tedious and time consuming.
Using data from a survey performed with the National Radio Astronomy Observatory’s (NRAO’s) Green Bank Telescope (GBT) in 2007, this thesis explores the application of machine learning techniques to mitigate the manual efforts involved in pulsar candidate detection. The performance of three different classifiers is explored – Naive Bayes, C4.5 (J48) Decision Tree, and Support Vector Machine. Preprocessing and feature extraction methods are described and a framework for applying the classifiers to the survey data is presented. Multiple features were extracted from the survey data and used to train the classifiers. Cross-validation results of the various feature sets and classifiers are documented. Experiments suggest the potential of the proposed framework in rapidly detecting pulsars from large amounts of survey data.Read More ›