Deep Learning–Based Enhancement of Small Sample Liquefaction Data

The liquefaction of sands remains an important topic in geotechnical earthquake engineering. The most widely used evaluation method is based on in situ testing means such as cone penetration test, standard penetration test, and dynamic penetration test. Recently, machine learning has emerged as a promising approach for evaluating liquefaction potential problems. Due to the complexity of the site and the different standards of the available measurement methods, however, the problem of small sample liquefaction data severely restricts the development of machine learning in the prediction and mitigation of soil liquefaction. Here, we propose the Wasserstein Generative Adversarial Networks (WGAN) to expand the sample size of the liquefaction data set. The result shows that the proposed method (WGAN) learns the feature distribution of the original data set effectively and improves the accuracy of the model. By comparing with Synthetic Minority Oversampling Technique, the superiority of Wasserstein Generative Adversarial Networks in data generation is demonstrated, especially for discrete data. The effectiveness of the method (WGAN) on soil liquefaction prediction is further analyzed using the K-means algorithm. The method (WGAN) provides a good solution for earthquake engineering where it is difficult to obtain comprehensive data and improves further the application of deep learning.