Transfer learning-based foundation pit deformation prediction

With the rapid urbanization in China, the construction of high-rise residential buildings has become increasingly vital. However, the increasing scale and depth of substructures in tall buildings have posed challenges in the excavation of deep foundation pits. The excavation of such deep foundation pits demands comprehensive monitoring and management to ensure structural stability and safety. This study aims to address the issue of foundation pit deformation prediction using machine learning methods. Traditional mechanics models and numerical analysis methods have demonstrated substantial errors in this context. Therefore, this paper reviews the applications of Gaussian processes, genetic algorithm-optimized BP neural networks, and Long Short-Term Memory (LSTM) networks in foundation pit deformation prediction. However, these methods exhibit limitations, such as the stationary assumption in Gaussian processes, high computational costs associated with genetic algorithms, and LSTM's limited performance with small datasets. To mitigate the challenges posed by limited monitoring data, this study introduces a transfer learning strategy based on the CNN-LSTM-Attention model. Validation results from practical case studies suggest that this strategy effectively enhances the accuracy of foundation pit deformation prediction, especially when dealing with limited data.