Scholarlink

<p>Liquefaction prediction using conventional approaches often relies on empirical correlations and involves costly, time-consuming field studies. Using large databases of post-liquefaction observations, machine learning methods have recently been developed to evaluate liquefaction potential. However, the availability of adequate real-world data limits the efficacy of these approaches. This study investigates the efficacy of the deep learning technique, i.e., Conditional Tabular Generative Adversarial Networks (CTGAN), for generating synthetic data. A comparison between original and synthetic data is made based on the absolute log mean, numeric data standard deviation, cumulative sums per feature, a correlation matrix, principal component analysis (PCA), distributional features and <i>p</i>-values from the Kolmogorov–Smirnov (KS) test. It is found that the synthetic data statistically resembles the original, making it viable for developing predictive models. There is a notable increase in the accuracy of liquefaction predictions when using 10,000 synthetic datasets generated from 288 original datasets. The synthetic data outperforms the original datasets across various machine learning methods, including Logistic Regression, Random Forest, SVM, KNN, and Decision Tree, with improvements in liquefaction classification accuracy of 89%, 98%, 92%, 98%, and 98%, respectively.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Assessment of Soil Liquefaction Potential Prediction Using Synthetic Data and Soft Computing Techniques

  • 期刊论文

  • 发表日期：

    2026年2月13日

查看全文
  • Jajati Keshari Naik,
  • Pradyut Kumar Muduli,
  • Gopal Charan Behera

摘要

Liquefaction prediction using conventional approaches often relies on empirical correlations and involves costly, time-consuming field studies. Using large databases of post-liquefaction observations, machine learning methods have recently been developed to evaluate liquefaction potential. However, the availability of adequate real-world data limits the efficacy of these approaches. This study investigates the efficacy of the deep learning technique, i.e., Conditional Tabular Generative Adversarial Networks (CTGAN), for generating synthetic data. A comparison between original and synthetic data is made based on the absolute log mean, numeric data standard deviation, cumulative sums per feature, a correlation matrix, principal component analysis (PCA), distributional features and p-values from the Kolmogorov–Smirnov (KS) test. It is found that the synthetic data statistically resembles the original, making it viable for developing predictive models. There is a notable increase in the accuracy of liquefaction predictions when using 10,000 synthetic datasets generated from 288 original datasets. The synthetic data outperforms the original datasets across various machine learning methods, including Logistic Regression, Random Forest, SVM, KNN, and Decision Tree, with improvements in liquefaction classification accuracy of 89%, 98%, 92%, 98%, and 98%, respectively.