해석 데이터 기반 흙막이 벽체 거동 예측에서의 현장 변동성을 고려한 가우시안 노이즈 주입 데이터 증강 기법

초록

This study presents a Gaussian noise–based data augmentation strategy to improve the predictive performance of a multi-resolution ConvLSTM framework for predicting retaining wall deformation during excavation. Although an initial training database comprising 4,200 samples was generated using finite element analysis (PLAXIS 2D), the dataset was inherently limited by the idealized, smooth deformation patterns typically produced by numerical simulations. Consequently, it could not adequately represent the localized and irregular deformation characteristics observed under field conditions. To address this limitation, Gaussian noise with varying spatial resolutions and amplitudes was injected into the simulation-based displacement data, expanding the database to 16,800 samples. The trained model was subsequently validated using 34 inclinometer datasets collected from 11 excavation sites in South Korea. The results demonstrated that the proposed data augmentation strategy effectively mitigated overfitting to idealized simulation data, significantly improving the prediction accuracy and reproducibility of diverse, irregular deformation patterns observed in the field while reducing prediction uncertainty.

키워드