Prediction method of longitudinal surface settlement caused by double shield tunnelling based on deep learning

Area of Science:

• Geotechnical Engineering
• Artificial Intelligence in Civil Engineering
• Tunneling and Underground Construction

Background:

• Shield excavation poses challenges for predicting surface settlement due to limited data and high-dimensional operational parameters.
• Accurate prediction of surface settlement is crucial for mitigating risks associated with underground construction projects.

Purpose of the Study:

• To develop an optimized deep learning framework for predicting longitudinal surface settlement caused by shield excavation.
• To address the challenges of small sample data and high-dimensional parameters in settlement prediction.
• To incorporate the influence of twin tunnels into the settlement prediction model.

Main Methods:

• Compared various optimization algorithms, selecting the slime mould algorithm (SMA) for hyperparameter optimization of random forest (RF).
• Utilized SMA-optimized RF (SMA-RF) for dimensionality reduction and feature contribution analysis.
• Proposed a double-input deep neural network (D-DNN) framework to account for twin tunnel influences and enhance data fidelity.

Main Results:

• SMA demonstrated superior performance over other optimization algorithms.
• Input features with a cumulative contribution exceeding 90% yielded high prediction accuracy.
• Reduced shield operational parameters exhibited a strong nonlinear relationship with surface settlement.
• The D-DNN model, considering twin tunnels, expanded the database by over 1.5 times, improving R² by 27.85% and reducing MAE by 53.2% compared to S-DNN.

Conclusions:

• The proposed D-DNN framework effectively predicts surface settlement, outperforming traditional methods by incorporating twin tunnel effects.
• Feature selection using SMA-RF is vital for accurate settlement prediction, though feature contribution analysis shows uncertainty with small datasets.
• The study highlights the potential of optimized deep learning for complex geotechnical engineering problems.