A Deep Learning Approach to Large-Scale Light Curve Prediction and Real-Time Anomaly Detection with Grubbs Criterion

Abstract

In light curves (LCs), the brightness of stars is associated with time, and so is its image. The traditional data processing methods cannot effectively handle real-time and large-volume data of various LCs. To address this issue, this work develops a deep neural network named Dropout-based Recurrent Neural Networks (DRNN). It extracts complicated features of all images captured by Mini Ground-based Wide-Angle Camera array (Mini-GWAC) for point source extraction and cross-certification through Long Short-Term Memory units. DRNN can also produce warnings for abnormal values of light change curves. Furthermore, this work optimizes the training model by combining a dropout method with an adaptive moment estimation algorithm to iteratively update the network weight of the RNN based on the LCs data. Extensive experiments with a Mini-GWAC dataset demonstrate that DRNN outperforms several typical methods in terms of prediction performance of star brightness in large-scale astronomical LCs.