Sferics noise minimisation methods for transient electromagnetic systems using neural networks
thesisposted on 2022-03-29, 02:38 authored by Hak Soo Hwang
The time-domain electromagnetic (TEM) method is used generally for conductivity mapping and in particular for mineral exploration. Since the method uses part of the broadband (10⁻² to 10⁵ Hz) spectrum, it is affected by electromagnetic (EM) noise which arises from cultural and natural sources. Thus, sferics, power line and VLF noise provide major limitations to minimum detectable signals in ground and airborne EM surveys. -- New EM noise minimisation techniques described in this thesis use a model training method (MTM), a local noise prediction filter (LNPF), and a remote noise prediction filter (RNPF). All these filters are based on a backpropagation neural network. A major benefit of an artificial neural network for EM noise reduction is its adaptation capability, which arises from its ability to adapt connection weights of the network in order to improve performance based on current results. -- An MTM, which can be viewed as performing a mapping from a noisy transient to a noise-free transient, can filter out EM noise from a very noisy transient with a normalised filter error (NFE) of less than 17% at the late delay times. However, since the MTM is deprendent on the training models for the network, it cannot be generally applied. The neural network should be trained with enough models to cover the geoelectric structure of a given region where EM data are being collected. -- A neural network-based LNPF is used for an in-loop TEM geometry to predict the vertical component of noise from its horizontal components. For very low frequency (VLF) (e.g., 10, ~20, and 44 kHz) background noise, high-frequency (>~5 kHz) sferics, and low-frequency (>~1 Hz) geomagnetic field reductions, the LNPF achieves a noise reduction factor (NRF) of more than 5. In particular, in a frequency range of 5 to 50 kHz, where power is generally contributed by sferics pulses, the LNPF suppresses sferics noise by a factor of 20 dB (i.e., an NRF of 10 in amplitude). -- In the reduction of background EM (VLF) and high-frequency sferics noise, an RNPF attenuates the power of EM noise by more than 20 dB (i.e., a factor of 10 in amplitude). Furthermore, a neural network-based LNPF and RNPF have been assessed with ground-based TEM data obtained with both in-loop and fixed-loop geometries in exploration conditions. The neural network-based noise prediction filters produce smooth TEM profiles compared with the profiles obtained without applying the noise prediction filters. -- A backpropagation neural network has been shown to be effective for noise reduction based on a model training method (MTM) and prediction filters such as an LNPF and RNPF. The neural network can recognise and correct for variable time shifts between local and remote receivers caused by speed of light delays and source directions. The noise reduction achieved with these methods is generally better than those based on alternative methods (e.g., a least-square fitting method, an LNPF based on the tipper method, a simple subtraction method, a transfer function model, and an interpolation method). With currently-available computing power, it is most probably not practical to implement these noise reduction methods as on-line filters. They need to be applied as post-processing procedures.