Time-domain denoising of CSAMT data base on long short-term memory

Abstract

Controlled-source audio-frequency Magnetotellurics method (CSAMT) partially overcomes the drawbacks of weak natural field signals. However, substantial interference is an inevitable part of field surveys in practice, which negatively impacts signal quality. We require new denoising techniques since traditional techniques, such as Fourier transformation, which compute apparent resistivity directly from frequency-domain data, are insufficient in our situation. CSAMT denoising research is currently lacking, nevertheless. This research proposes the use of long short-term memory (LSTM) neural networks to denoise CSAMT signals in the time domain, given their good performance in processing Magnetotelluric (MT) data as shown by prior studies. We seek to directly extract the desired frequency signal for denoising from the time series data, in contrast to conventional denoising techniques. Since noise and target frequency signals are mixed together in MT data, the only way to suppress noise is to find the characteristics of the noise in the time series. CSAMT, on the other hand, differs from MT in that it uses an artificial transmitting source and fixes the valid signal frequency within a temporal window. This makes it possible to extract target frequency signals directly without taking into account the intricate properties of noise. In order to complete the noise suppression job, we created a neural network in this study that is based on bidirectional LSTM. This method was able to partially handle the difficulty of denoising when the data's SNR falls below 0 dB and, on average, enhance the signal-to-noise ratio (SNR) of CSAMT data by roughly 20 dB after executing both simulated and measured data testing.