Detection of a Quasiperiodic Phenomenon of a Binary Star System Using Convolutional Neural Network

Abstract

Pattern recognition algorithms are commonly utilized to discover certain patterns, particularly in image-based data. Our study focuses on quasiperiodic oscillations (QPO) in celestial objects referred to as cataclysmic variables (CV). We are dealing with interestingly indistinct QPO signals, which we analyze using a power density spectrum (PDS). The confidence in detecting the latter using certain statistical approaches may come out with less significance than the truth. We work with real and simulated QPO data of a CV called MV Lyrae. Our primary statistical tool for determining confidence levels is sigma intervals. The aforementioned CV has scientifically proven QPO existence, but as indicated by our analysis, the QPO ended up falling below 1 − σ , and such QPOs are not noteworthy based on the former approach. We intend to propose and ultimately train a convolutional neural network (CNN) using two types of QPO data with varying amounts of training dataset lengths. We aim to demonstrate the accuracy and viability of the classification using a CNN in comparison to sigma intervals. The resulting detection rate of our algorithm is very plausible, thus proving the effectiveness of CNNs in this scientific area.