Conditional WGAN-GP approach for geotechnical dataset augmentation

Description

In this work we propose a conditional generative adversarial network (CGAN)
with gradient penalty (GP) based approach of generating synthetic tabular geotechnical data for dataset augmentation. In geotechnical datasets events of utmost importance, so called rare events (e.g., fault zones during tunnel excavation, sudden increase in displacements shown in geodetic measurement data, etc.) are heavily underrepresented. Conditional GANs present a promising approach to to generate synthetic geotehnical data based on a condition (e.g., underepresented class data) that shows the same characteristics as the original data, but still presents unique samples with no connection to the technical content of the original data. A Wasserstein-GAN algorithm with Gradient Penalty (GP) for training stabilisation is used to generate the synthetic data and augment the existing dataset. The demands on the synthetic data are of a dualistic nature: on the one hand, the data has to be suffciently dissimilar to the original data, so that it does not simply create copies f the existing data (demand for originality). On the other hand, it has to show the same patterns and follow the same rules as the original data, so that it can be used as if it were real data (demand for conformity). The Conditional WGAN-GP model describes how a synthetic dataset is generated, in terms of a probabilistic model based on real data. By sampling from this model, we are able to generate new, unique and conditioned synthetic and realistic data. We show that both imposed demands on the newly generated data are fulfilled and an existing data set’s rare events can be suffciently augmented. Thus, enabling balancing of unevenly distributed classes and enhancing the overall size of a dataset. The authors see big potential for subsequent analysis on enhanced and balanced datasets, which might lead to more accurate predictions of rare events in geotechnical engineering.

Period	8 Aug 2023
Event title	22nd Annual Conferencee of the International Association for Mathematical Geosciences: IAMG 2023
Event type	Conference
Location	Trondheim, NorwayShow on map
Degree of Recognition	International