Deep Learning-Based Intelligent Prediction Method for Settlement of Buildings on Structurally Unstable Soils
Year: 2026
Type: conference-paper
Abstract: Accurate forecasting of structural settlement on geologically unstable foundations remains a persistent difficulty in geotechnical practice, where conventional regression-based approaches fail to adequately represent the nonlinear coupling between multi-source variables and time-varying deformation responses. This work introduces a hybrid deep neural network that fuses a one-dimensional Convolutional Neural Network (1D-CNN) with a Long Short-Term Memory (LSTM) module augmented by an attention mechanism, enabling joint extraction of cross-parameter spatial correlations and long-range temporal dependencies from heterogeneous monitoring records. An 18-dimensional input space spanning soil mechanical indices, foundation load descriptors, and site environmental variables was systematically constructed and fed into an end-to-end trainable architecture. Evaluation on a 13,104-sample field dataset from a southern China soft-ground project yields a Mean Absolute Error of 2.347 mm, Root Mean Square Error of 3.125 mm, and coefficient of determination R2 = 0.943, surpassing support vector regression and backpropagation networks by margins of 37.6% and 28.3%, respectively. Ablation tests isolate the contribution of each architectural component, and SHAP-guided attention visualization pinpoints compression modulus, preconsolidation pressure, and applied stress as the principal settlement drivers, offering a quantitatively interpretable tool for foundation monitoring and geohazard early warning.
Authors: Xue Li, Yuehan Li, Xinyu Cui, Jie Zhao, U. Dzhusuev
Language: English
Subfield: Civil and Structural Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7224671014