Deep learning models of tunnels for dynamic response prediction under train loads

• Main Author: Zhang, Jingjing
• Other Authors: Fu Yuguang
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Civil engineering
• Online Access: https://hdl.handle.net/10356/172675
• Institution: Nanyang Technological University

Tunnels are critical engineering structures for the transportation system in our built environment. Ensuring tunnels are under desirable working conditions is a major challenge to engineers. Structural Health Monitoring (SHM) systems can provide valuable tunnel responses by real-time sensors. These sensors would capture the tunnel responses such as strain and velocity for the engineer to access the condition of the tunnel to support decision-making processes. To explore the dynamic responses of tunnels under train loads, a high-fidelity finite element (FE) model of physical tunnels was built in ABAQUS (by others) as the baseline, and then deep learning models were developed by feeding the synthetic data extracted from the FE model. Deep learning models are the special types of data-driven models using more advanced deep learning algorithms which are also the subset of AI-based machine learning. The dynamic response types of synthetic data encompassed acceleration, velocity, stress, strain and displacement. It is not an easy task to accurately monitor dynamic displacements in real-time. AI-based machine learning techniques were investigated to support the development of deep learning models in predicting the tunnel responses. Four types of deep learning models using deep learning algorithms including MLP, GRU, LSTM and CNN were discreetly developed as a potential data-driven model of tunnels by training with the multivariate sequence dataset which was prepared and processed from the synthetic data. CNN model was finally selected as the optimal model due to its outstanding performance and reasonable overall training time. Thus, the CNN model was our data-driven model of tunnels for dynamic response prediction under train loads. This data-driven model also achieved satisfactory performance in predicting the dynamic response of tunnel displacement on new datasets. The results of this report have significant implications for building the data-driven model of tunnels. Particularly in predicting dynamic responses of tunnels under train loads with shorter training time and desirable performance.