Deep learning for the classification of transient noises and signals in the LIGO detectors

• Autor(a) principal: Fernandes, Tiago Saraiva
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/36277
• topic: Deep learning; aLIGO; Transient noise; Gravitational waves; Image classification

In this work, data from the aLIGO detectores collected during the first two aLIGO and AdV observing runs (O1 and O2), in the form of spectrograms, were classified using Deep Learning models based on Convolutional Neural Networks. As well as training models from scratch, pre-trained models were also employed, and their performance compared. Initially, a brief theoretical introduction on gravitational wave detection was performed, focusing on the LIGO detectors. In addition, the foundations of Deep Learning and current best practices for the training of image classification models were also presented. The computational experiments showed that encoding information from different time windows in the different colour channels enhanced the performance of the models and that small architectures were capable of separating the 22 classes present in the Gravity Spy dataset. Moreover, transfer learning was able to accelerate the training process and achieve classifiers with competitive performance. The best models obtained a macro-averaged F1 score of 96.84% (fine-tuned model) and 97.18% (baseline trained from scratch), which are in line with the best results in the literature for the same dataset. In addition, these models were evaluated on real gravitational wave signals from Compact Binary Coalescences from the first two aLIGO and AdV observing runs, and they achieved recalls of 75% and 25%, respectively, while only having been trained with a small number of signals from gravitational wave simulations.