Abstract
In the last couple of decades, tremendous progress has been achieved indeveloping robotic telescopes and, as a result, sky surveys (both terrestrialand space) have become the source of a substantial amount of new observationaldata. These data contain a lot of information about binary stars, hidden intheir light curves. With the huge amount of astronomical data gathered, it isnot reasonable to expect all the data to be manually processed and analyzed.Therefore, in this paper, we focus on the automatic classification of eclipsingbinary stars using deep learning methods. Our classifier provides a tool forthe categorization of light curves of binary stars into two classes: detachedand over-contact. We used the ELISa software to obtain synthetic data, which wethen used for the training of the classifier. For evaluation purposes, wecollected 100 light curves of observed binary stars, in order to evaluate anumber of classifiers. We evaluated semi-detached eclipsing binary stars asdetached. The best-performing classifier combines bidirectional Long Short-TermMemory (LSTM) and a one-dimensional convolutional neural network, whichachieved 98% accuracy on the evaluation set. Omitting semi-detached eclipsingbinary stars, we could obtain 100% accuracy in classification.