Integrating numerical modeling and deep learning with electrical resistance tomography for rock mechanics

Abstract

The integration of physical testing and numerical modeling is becoming increasingly important in rock mechanics. This study leverages deep learning techniques to combine numerical modeling with an electrical resistance tomography (ERT) device. A dataset of complex conductivity distributions is first generated using numerical modeling with multi-point spline curves. A normalized data preprocessing method is then employed to transform measured physical signals into simulated signals while preserving their intrinsic characteristics. This approach enables transfer learning, allowing the trained network derived from numerical modeling to be effectively applied to the physical device. Building on this foundation, a one-dimensional convolutional neural network (1D-CNN) model is developed, demonstrating significant advantages in terms of image reconstruction accuracy, computational efficiency, and robustness. The effectiveness of the 1D-CNN model is validated through its application in monitoring changes in electrical conductivity distributions during rock seepage, crack propagation, and failure processes in red sandstone specimens. This methodology offers a robust framework for integrating numerical modeling with physical experiments, providing a promising solution to address complex challenges in rock mechanics.