Identification of Fast Radio Bursts using Transfer Learning Approach with Data Augmentation

Abstract

The universe has many mysteries, such as pulsars, dying stars, supernovae, and fast radio bursts (FRBs), FRBs are millisecond long radio signals, detected as a spike in radio-telescope data. Identification of Fast Radio Bursts from available data involves manual inspection of exhaustive data/plots. Radio Frequency Interference in pose a major challenge in identification of Fast Radio Bursts due to their abundance in the observatory data. We present a machine-learning-aided system, which screens telescope-generated data and identifies potential Fast Radio Burst candidates in it. Proposed system employs Convolutional Neural Networks and Transfer Learning to classify potential Fast Radio Bursts from Radio Frequency Interference from data recorded by the uGMRT. We have used data simulation tools to synthesize additional samples in order to make up for the paucity of data. The VGG16-based model displayed the best receiver operating characteristics curve with the area under curve being 0.90 along with an accuracy of 90.67%.