A deep learning-based parametric inversion for forecasting water-filled bodies position using electromagnetic method

Abstract

The transient electromagnetic method (TEM) is extensively employed for identifying regions with low resistivity ahead of tunnel construction. Nevertheless, performing a 2D or 3D inversion of geo-electric models across the entire subsurface is impractical due to the limited space within the underground tunnel and the non-uniqueness associated with TEM inversion. We design a tunnel electromagnetic joint scan observation system and present a deep learning-based parametric inversion for improved tunnel electromagnetic imaging, designed specifically for tunnel prediction of water-filled structures. It utilizes a configuration wherein transmitters scan along the surface while receivers are positioned within the tunnel, employing time-domain and frequency-domain transmitters and a multi-component receiver. The DL model for the first time provides parametric imaging of two different view, forming a self-checking mechanism, which can help constrain the predictions and reduce the non-uniqueness of the inversion. Trained by synthetic data, our system shows impressive adaptability to predict the 3D spatial position of water-filled anomalies and strong robustness under different tunnel environments including metal inference and undulating terrain conditions.