Multi-feature SCPTU data driven prediction of undrained shear strength of clay using DBO-optimized CNN-BiLSTM with attention mechanism

Abstract

Accurate estimation of the undrained shear strength (Su) of clay remains challenging due to the complex and nonlinear interactions inherent in multi-source in-situ test data. To address this issue, we propose DBO-CNN-BiLSTM-Attention (DBO-CBA), a novel hybrid deep learning framework integrating (CNN) for local feature extraction, BiLSTM for depth-dependent sequential learning, and a self-attention mechanism to emphasize critical information. A key innovation lies in the integration of the Dung Beetle Optimizer for automatic hyperparameter tuning, which reduces the subjectivity and inefficiency associated with manual configuration. Using 394 normalized SCPTU datasets from Quaternary clays, the DBO-CBA model significantly outperforms standalone deep learning models and traditional empirical approaches, achieving an R 2 of 0.958, RMSE of 0.061, and MAPE of 10.35 %. SHAP and perturbation sensitivity analyses further demonstrate enhanced interpretability and physically consistent feature-response relationships. This study proposes a reliable and interpretable method for non-destructive evaluation of clay strength, delivering a novel intelligent framework for clay strength prediction.