Abstract
Abstract
Nowadays, astronomy has entered the era of Time-Domain Astronomy, and the study of the time-varying light curves of various types of objects is of great significance in revealing the physical properties and evolutionary history of celestial bodies. The Ground-based Wide Angle Cameras telescope, on which this paper is based, has observed more than 10 million light curves, and the detection of anomalies in the light curves can be used to rapidly detect transient rare phenomena such as microgravity lensing events from the massive data. However, the traditional statistically based anomaly detection methods cannot realize the fast processing of massive data. In this paper, we propose a Discrete Wavelet (DW)-Gate Recurrent Unit-Attention (GRU-Attention) light curve warning model. Wavelet transform has good effect on data noise reduction processing and feature extraction, which can provide richer and more stable input features for a neural network, and the neural network can provide more flexible and powerful output model for wavelet transform. Comparison experiments show an average improvement of 61% compared to the previous pure long-short-term memory unit (LSTM) model, and an average improvement of 53.5% compared to the previous GRU model. The efficiency and accuracy of anomaly detection in previous paper work are not good enough, the method proposed in this paper possesses higher efficiency and accuracy, which incorporates the Attention mechanism to find out the key parts of the light curve that determine the anomalies. These parts are assigned higher weights, and in the actual anomaly detection, the star is detected with 83.35% anomalies on average, and the DW-GRU-Attention model is compared with the DW-LSTM model, and the detection result f1 is improved by 5.75% on average, while having less training time, thus providing valuable information and guidance for astronomical observation and research.