Application of Deep Learning Methods for Distinguishing Gamma-Ray Bursts from Fermi/GBM TTE Data 后印本

摘要: To research the burst phenomenon of gamma-ray bursts (GRBs) in depth, it is necessary to explore an effective and accurate identification of GRBs. Onboard blind search, ground blind search, and target search method are popular methods in identifying GRBs. However, they undeniably miss GRBs due to the influence of threshold, especially for sub-threshold triggers. We present a new approach to distinguish GRB by using convolutional neural networks (CNNs) to classify count maps that contain bursting information in more dimensions. For comparison, we design three supervised CNN models with different structures. Thirteen years Time-Tagged Event (TTE) format data from Fermi/GBM is employed to construct useful data sets and to train, validate and test these models. We find an optimal model, i.e. the ResNet-CBAM model trained on the 64 ms data set, which contains residual and attention mechanism modules. We track this deep learning model through two visualization analysis methods separately, Gradient-weighted Class Activation Mapping (Grad-CAM) and T-distributed Stochastic Neighbor Embedding (t-SNE) method, and find it focused on the main features of GRBs. By applying it on one-year data, about 96% of GRBs in the Fermi burst catalog were distinguished accurately, six out of ten GRBs of sub-threshold triggers were identified correctly, and meaningfully thousands of new candidates were obtained and listed according to their SNR information. Our study implies that the deep learning method could distinguish GRBs from background-like maps effectively and reliably. In the future, it can be implemented into real-time analysis pipelines to reduce manual inspection and improve accuracy, enabling follow-up observations with multi-band telescopes.