An adversarial multi-source transfer learning method for the stability analysis of methane hydrate-bearing sediments

Abstract

This study presents an innovative adversarial multi-source transfer learning approach to enhance submarine hydrate slope stability predictions in the face of small and varied datasets. Integrating terrestrial and submarine data, our method significantly improves knowledge transfer and model generalization. Utilizing hydrate triaxial tests and pore pressure models, we construct a comprehensive dataset that bridges the gap between diverse geological environments. Employing the novel Walrus Optimizer and adversarial training techniques, the model substantially outperforms traditional regression methods. It achieves a correlation coefficient of 0.9936 and a mean absolute error of 0.094, indicating high predictive accuracy and robust handling of data anomalies and distribution inconsistencies. These advancements provide crucial insights into slope stability factors and offer potential enhancements for geological hazard monitoring and early warning systems. Our research demonstrates a substantial improvement in slope stability analysis and opens new avenues for intelligent geological hazard assessments in environments characterized by data scarcity.