Neural network interpretation of electromagnetic ellipticity data in a frequency range from 1 kHz to 32 MHz

Abstract

A new real-time in-field interpretation and visualization scheme and software, using neural networks for the detection and localization of buried waste, and the boundaries of waste sites, has been developed. The capabilities and limitations of the high-frequency (1 kHz to 1 MHz and 31 kHz to 32 MHz) electromagnetic ellipticity systems are analyzed by numerically studying the sensitivity of the acquired 3D-ellipticity to model parameters describing the geometry of the systems and the electrical parameters of layered-earth models. Changes in ellipticity due to coil misalignment in standard operating mode are typically just 1% to 2%. Changes due to variations in layered-earth model parameters (resistivity, relative dielectric constant, and thickness) are typically at least one order of magnitude higher. Hence, it will be possible to resolve these parameters. For conductive models (resistivity < 50 Ωm) it will be hard to determine the relative dielectric constant and for models with high relative dielectric constants it will be hard to determine the resistivity, especially if it is greater than 1000 Ωm. The results of the sensitivity analysis contribute considerably to the training of several neural networks to determine the electrical properties of the subsurface. The two classes of artificial neural network paradigms utilized in this study are the radial basis function and the modular neural network algorithms. One-dimensional layered-earth inversions are performed by neural networks using ellipticity data. The three-dimensional localization of metallic objects (e.g. drums) is done by visualizing the results of one particular halfspace neural network technique. Several small conductive objects have been detected by applying this technique to data collected in controlled physical modeling field experiments. Classification neural networks are trained on field data to categorize ellipticity soundings into either a target or a background class. Two environmental geophysics field case studies were analyzed using the developed interpretation system and the visualization software. The first case study involves mapping subsidence areas caused by an underground coal mine fire in Wyoming. The neural network interpretations from the mine survey match comparable inversion results. The second study documents the successful characterization of a simulated hazardous waste pit at the Idaho National Engineering Laboratory.