A Deep Learning Approach for Transient Electromagnetic Data Denoising, Inversion and Uncertainty Analysis With Monte Carlo Dropout Technique

Abstract

A comprehensive deep learning approach was introduced, encompassing data denoising, inversion imaging and uncertainty analysis. For denoising transient electromagnetic (TEM) data, we utilized a Bidirectional Long Short-Term Memory (BiLSTM) network. In the data inversion process, a combination of convolutional neural network (CNN) and BiLSTM structures was employed, and their outputs were consolidated using a multi-head attention mechanism. To ensure robust performance under challenging noise conditions, we implemented a specialized multi-channel noise training protocol during model optimization. The framework incorporates Monte Carlo (MC) dropout techniques to systematically evaluate prediction reliability throughout the inversion pipeline. This approach has not only been validated on test datasets but has also been successfully applied to the field dataset collected at the Narenbaolige Coalfield in Inner Mongolia, China. The deep learning inversion results obtained from both raw and denoised data exhibit reduced vertical continuity and increased roughness characteristics. In contrast, the Occam's inversion method with smoothness constraints yields results demonstrating superior lateral continuity and vertical smoothness. It is noteworthy that both inversion approaches show consistent interpretations regarding the scale of basalt formations and their contact interfaces with underlying sedimentary layers. Further uncertainty analysis reveals relatively higher uncertainty characteristics in the transition zones between basalt and sedimentary layers, as well as in deeper formations. The elevated uncertainty at interface regions may be attributed to model resolution limitations and inversion ill-posedness issues, whereas the higher uncertainty in deeper formations is more likely caused by the volumetric effects of electromagnetic field detection and the influence of observational data noise.