Supervised Learning for EM Time Series Modeling and Vehicle-Related Noise Compensation

Le résumé fourni par la source

Over the past few years, machine learning algorithms and methods underpinned by artificial learning have experienced an astounding surge in commercial and academic fields. Machine learning allows patterns, dependencies and regularities to be inferred directly from example data, without relying on preconceived physical concepts, ideas and abstractions shaped by human intuition (LeCun, Bengio, & Hinton, 2015; Andrews, 2025). Even complex, yet intuitively graspable, physical models may no longer offer best predictive performance. The close link between predictive capability and economic efficiency in our societies has fueled the spread of artificial learning and artificially generated content – a development that may erode, or deepen, the understanding of physical phenomena for future generations. In either direction, a rigorous engagement with the potential and operational principles of this technological development is imperative. This contribution is organized into two chapters, each introducing to an application of supervised machine learning in a geophysical context. The examples of application emphasize the capabilities of ordinary neural networks for learning non-linear mappings. Chapter I exemplifies the operation of a simple feedforward neural network (FNN) on the task of continuing a periodic time series, Chapter II investigates the suitability of such a network architecture for predicting sensor readings based on independently measured data. The two issues are intended to provide insight and indicate potential, but also to address two specific practical problems in the field of acquisition and processing of electromagnetic (EM) data: reconstruction of corrupt or incomplete amperage recordings for controlled source EM applications and the prediction of EM noise fields caused by the motion of a carrier vehicle, sensor motion, and fields emitted by vehicle-related operational currents. The data set of this study comprises field data from a drone-based semi-airborne EM survey conducted at a test site in Northern Germany in November 2018 (Stoll et al., 2019). Training was performed on a Lenovo ThinkPad E480 (Intel Core i5-8250U, 8 GB DDR4) using multi-threaded CPU computation.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.