Self-supervised machine learning: a new hope for gravitational wave detection?

• Autor(a) principal: Vieira, Samuel Jorge Fernandes
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/39476
• topic: Image classification; Gravitational wave signals; Deep learning; Elf-supervised learning; Supervised learning; Gravitational wave glitches; Resnets

In this thesis, the problem of the appearance of ”glitches” during the detection of gravitational wave signals is addressed. ”Glitches” are classified as noise (excluding background noise), whose origin may be due to the instrumentation used in the detectors or natural causes. Initially, a short introduction is made in this thesis about the operation of LIGO detectors, gravitational waves, and glitches, in the initial chapters. To solve the problem addressed, the use of neural networks is proposed to classify and identify the glitches that appear during data acquisition. Supervised learning methods together with self-supervised learning were used to train neural networks and classify spectrograms that contain frequency versus time information for each corresponding signal. The selected neural network architecture was ”Resnet18”, and the optimizer selected was ”AdamW”. The selected loss function was ”Cross-Entropy”. Three training sessions were conducted for each model: model (I) used supervised learning, and models (II) and (III) used self-supervised learning. The model (III) used self-supervised learning with the knowledge acquired from model (II) through ”transfer learning”. The training of these neural networks, led to high values for the metrics used in this thesis. The highest accuracies obtained were about 96.81%, 96.74% and 96.50% for the corresponding models (I), (II), and (III), while the best macro averaged F1-score was about 94.49%, 96.75% and 94.15%, with hardly any difference in metrics in the method employed.