Autocodificadores convolucionais para atenuação de ruídos em dados sísmicos

• Ano de defesa: 2020
• Autor(a) principal: Salvador Neto, Carlos
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Rio de Janeiro; Brasil; Instituto Alberto Luiz Coimbra de Pós-Graduação e Pesquisa de Engenharia; Programa de Pós-Graduação em Engenharia Civil; UFRJ
• Palavras-chave em Português: Redes neurais profundas; Autocodificadores convolucionais; Engenharia Civil
• Link de acesso: http://hdl.handle.net/11422/23170

The seismic method stands out for being essential for the oil industries in the identification and characterization of hydro-carbide reservoirs, being able to provide important information regarding the geological structure, thus leading to a detailed picture of the subsurface geology. Seismic data are physical observations, measurements or estimates of seismic sources, seismic waves and their means of propagation. The objective of acquiring and processing seismic data is to learn something about the Earth’s Interior and to understand certain aspects of the Earth, it is necessary initially to establish some specific relationships between the intended objectives and the measurable parameters. The first step is to perform the data acquisition projected for the problem, then the data is processed to identify and improve the desired signal, and finally, interpretations of the data are performed based on the processed data. This work has as main focus one of the first stages of seismic processing, and aims to improve the signal-to-noise ratio of seismic data. For this purpose, the work proposes an approach based on deep learning models to directly attenuate noise such as Swell, linear and random noises in pre-stack data. The network learns how to directly detect noise and then obtains attenuated data, removing noise from the corrupted raw data set. The evaluation of the results of this method will be based on the use of evaluation functions such as the signal-to-noise ratio or mean quadratic error and through a visual analysis of the seismic image quality in order to verify the recovery of seismic signals damaged by noise. The tests were performed using fully connected networks and the results of the signal-to-noise ratio and the mean square error indicate that the deep learning approach has the ability to attenuate noise without damaging the primary signals.