High embankment slope stability prediction using data augmentation and explainable ensemble learning

Abstract

The stability of embankment slopes for heavy-haul railway foundations is essential for safe railway operations. Railway embankment slope stability datasets often rely on engineering judgment for analysis. The labor- and resource-intensive processes of data preparation result in small dataset sizes. Machine learning analysis of small-sample potential features is a key low-cost approach for slope prediction. Due to the limited availability of slope failure data, a specialized framework is required for predictive modeling. To address this challenge, the focus is placed on data augmentation and interpretability analysis. A generative adversarial model is constructed using a graph convolutional network-based generator and a discriminator based on Gated Recurrent Unit, accompanied by a quality control method for the generated samples based on maximum mean discrepancy and one-class Support Vector Machine. This approach is designed to more effectively capture the temporal and spatial features of small samples. Three ensemble learning models, namely, XGBoost, random forest, and AdaBoost, are trained with augmented data, and model interpretation is conducted using Shapley Additive exPlanations to identify key factors affecting stability and potential stability improvement strategies. Results indicate that the proposed generative adversarial model surpasses traditional models in generating adequate data; the three enhanced data-trained machine learning models in this study achieved at least a 12% improvement in predictive accuracy, compared to their original small-sample-trained counterparts; The proposed data augmentation method outperformed variational autoencoder and diffusion models in generating high-quality synthetic data. Additionally, the interpretability framework effectively identified primary factors influencing slope stability. These findings provide a robust framework for interpretability-driven assessments of heavy-haul railway slopes with limited sample data.