Full text: Download
Abstract Astro-rivelatore Gamma a Immagini Leggero (AGILE) is a space mission launched in 2007 to study X-ray and gamma-ray astronomy. The AGILE team developed real-time analysis pipelines to detect transient phenomena such as gamma-ray bursts (GRBs) and react to external science alerts received by other facilities. The AGILE anticoincidence system (ACS) comprises five panels surrounding the AGILE detectors to reject background-charged particles. It can also detect hard X-ray photons in the energy range 50–200 keV. The ACS data acquisition produces a time series for each panel. The time series are merged into a single multivariate time series (MTS). We present a new deep-learning model for the detection of GRBs in the ACS data using an anomaly detection technique. The model is implemented with a convolutional neural network autoencoder architecture trained in an unsupervised manner, using a data set of MTSs randomly extracted from the AGILE ACS data. The reconstruction error of the autoencoder is used as the anomaly score to classify the MTS. We calculated the associated p-value distribution, using more than 107 background-only MTSs, to define the statistical significance of the detections. We evaluate the trained model with a list of GRBs reported by the GRBWeb catalog. The results confirm the model’s capabilities to detect GRBs in the ACS data. We will implement this method in the AGILE real-time analysis pipeline.