2022MNRAS.509.1269O

Time series data mining is an important field of research in the era of 'Big Data'. Next generation astronomical surveys will generate data at unprecedented rates, creating the need for automated methods of data analysis. We propose a method of light-curve characterization that employs a pipeline consisting of a neural network with a long-short term memory variational autoencoder architecture and a Gaussian mixture model. The pipeline performs extraction and aggregation of features from light-curve segments into feature vectors of fixed length that we refer to as light-curve 'fingerprints'. This representation can be readily used as input of down-stream machine learning algorithms. We demonstrate the proposed method on a data set of Rossi X-ray Timing Explorer observations of the Galactic black hole X-ray binary GRS 1915+105, which was chosen because of its observed complex X-ray variability. We find that the proposed method can generate a representation that characterizes the observations and reflects the presence of distinct classes of GRS 1915+105 X-ray flux variability. We find that this representation can be used to perform efficient classification of light curves. We also present how the representation can be used to quantify the similarity of different light curves, highlighting the problem of the popular classification system of GRS 1915+105 observations, which does not account for intermediate class behaviour.