Generative adversarial network for geological prediction based on TBM operational data

Abstract

The prediction of tunnel geological conditions plays an important role in underground engineering, such as the tunnel construction and tunnel dynamic design. However, due to the invisibility of underground geological conditions, there remain many challenges in the design of geological prediction models. In this paper, we propose a generative adversarial network for geological prediction (GAN-GP) to accurately estimate the thickness of each rock-soil type in a tunnel boring machine (TBM) construction tunnel based on operational data collected from sensors equipped on the TBM. The generator of the GAN-GP contains feature-extraction (FE) and feature-integration (FI) modules. The former extracts the important features from the TBM operational data, and the latter produces the geological condition prediction, which estimates the thickness of each rock-soil type at a location. The discriminator of the GAN-GP determines whether the FI module’s outputs are true geological data. After adversarial training, if the trained discriminator fails to distinguish them, the outputs of the FI module will accurately approximate the true geological condition. Experimental results support the effectiveness of the proposed GAN-GP model for geological prediction, and show that it outperforms the state-of-the-art models including support vector regression (SVR), feed-forward neural network (FNN) and random forest (RF) models.