Abstract

In this article, we present a dataset of light curves with synthetic signals. BRITE light curves (a constellation of five nanosatellites) are the main source of this dataset. We create the synthetic light curves of exoplanet transit by applying a pre-processing to the BRITE data and an injection of transit according to the Mandel and Agol model with a constraint of stellar radius <3.08[R-sun] and planetary radius between 0.95 and 2.1 [R-jup]. We apply a quality criterion, obtaining 597 Planet Candidate (PC) examples and 3126 Not Planet Candidate examples as a dataset. PCs are injected simulated planets and are not around unique stars. We design a Deep Learning (DL) model to be trained with the created dataset. The DL model is a modified AstroNet Convolutional Neural Network (CNN) from literature to detect possible exoplanets. After evaluation over the testing set we obtain an accuracy of 99.46%, precision of 100% (PCprecision) and a recall of 96.72% for the PC class (PCrecall), and an area under the curve receiver operating characteristics (AUC-ROC) of 100%, overcoming the results of existing networks tested on BRITE data. We ultimately search for potential exoplanets using the pre-processed data from BRITE, finding signals similar to exoplanetary transits in the targets HD 039060, HD 022049, HD 036861 and HD 218396.