Combining Deep Learning and Transformations for Seismic Data Processing

During seismic acquisition, reflected waves from the subsurface are recorded by sensors in seismic cables in the water column. The measured data in these sensors are used to create images and find the properties of the subsurface rocks. To achieve a clean image of the subsurface, advanced signal process removes unwanted energy, such as wave noise from the sea surface, background noise, noise from other seismic surveys, water surface echoes (ghosts) and waves bouncing up and down (so-called multiples). In addition, we need to correct for the effect of the source wave and the distortions of the signal in the sensors. The Ph.D. candidate, Jing Sun, has used Machine Learning and Deep Neural Network (DNN) and a series of supervised training sequences to solve some of the fundamental problems in signal processing. The focus of Jing?s research has been on two topics. The first is the removal of noise from other seismic surveys, so-call Seismic Interference (SI). The second topic is so-called deblending and has become a serious problem in modern seismic acquisition and processing as the shooting rate has increased to improve efficiency. Previously the shooting rate was typically 5-8 seconds, while now the shooting rate may be down to 2.5 to 3 seconds which means that the reflected signal from the previous shot is still recorded when the new shot is fired. In her Ph.D. work, Jing has trained the networks in several ways. The key problem is to establish clean, uncontaminated shots as a target for the training. For this she has used both clean shots at the end of the sail lines, clean shots from conventional deblending and the clean part of the shot gathers before the next shot. She has also done some ground-breaking experiments on how to use noise in the training process. The work Jing has done has led to an improved workflow for SI removal in two commercial projects (see Paper 3), plus a patent application where Jing is the second author. Jing was supposed to deliver her thesis in December 2022 but has worked very hard and submitted her thesis to the University of Oslo (UiO) in October 2021. UiO approved it and delivered it to the adjudication committee consisting of external opponent Dr. Aina Juell Bugge from Lundin energy Norway and Professor Johan Robertson from ETH Zurich and internal UiO opponent Professor Jan Inge Faleide. In December 2022 the adjudication committee approved the thesis. The defense was scheduled for 28th January 2022. The core of Jing?s Ph.D. thesis consists of four papers: Paper 1: Attenuation of marine seismic interference noise employing a customized U-Net, by Jing Sun, Sigmund Slang, Thomas Elboth, Thomas Larsen Greiner, Steven McDonald and Leiv Jacob Gelius, published in Geophysical Prospecting, 2020, 68, no. 3, 845-871. Paper 2: An exploratory study toward demystifying deep learning in seismic signal separation, by Jing Sun and Song Hou, (submitted to Geophysics in October 2022) Paper 3: DNN-based workflow for attenuating seismic interference noise and its application to marine towed streamer data from the North Sea, by Jing Sun, Song Hou, and Alaa Triki. (submitted to Geophysics in November 2022) Paper 4: Deep learning-based shot-domain seismic deblending, by Jing Sun, Song Hou, Vetle Vinje, Gordon Poole and Leiv Jacob Gelius, submitted to Geophysics 30th November 2020 (accepted for publication in Geophysics)

Arbeidet Jing har gjort har ført til en forbedret resultat for fjerning av esktern støy (SI) i to kommersielle prosjekter (se artikkel 3), pluss en patentsøknad der Jing er andreforfatter.

Easier access to affordable and powerful hardware (CPU and GPU) solutions together with user-friendly open-source software such as TensorFlow, has been the key to the accelerated use of machine learning within various areas of technology. In seismic data processing, artificial neural networks (ANNs) have the potential to be applied to many of the key processing steps (deblending, seismic interference attenuation, deghosting, etc.) which today involve significant testing time and computational power. Once trained, ANNs are computationally very light and potentially adaptable to different datasets. Sorting the data from common source domain to another by using mathematical transformation such as Wavelet transform or Shearlet transform may make the target signals and noise have more differences in characteristics. Their use could, therefore, save processing time and, in the long term, impact the whole business sector. The proposed doctoral work is about the usage of ANNs and transformations for processing of marine seismic data. The goal is to achieve similar or better quality results compared to conventional processing methods. If this is achieved, deep learning-based methods can save significant computer resources and time.