Prediction of penetration rate by GRU neural network

• Main Author: Leo, Kew Xun Wei
• Other Authors: Zhao Zhiye
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Civil engineering::Geotechnical; Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/158680
• Institution: Nanyang Technological University

Mechanised tunnelling projects using boring machines (TBM) are becoming widespread. Here in Singapore multiple projects are underway that expands our subterranean space. The productivity and financial viability of these projects have since become an important topic. Much research, past and ongoing, are focused on accurately predicting the penetration rate of a TBM because of its importance. Of the multitude of prediction models, neural networks showed great potential for practical use. This study proposes an approach to develop and deploy a neural network based on the Gated Recurrent Unit (GRU), a type of neural network that is suited to prediction of time-based datasets. The network model is trained and tuned for accuracy, and the mean absolute error (MAE) and root-mean squared error (RMSE) analysed to set the most optimal combinations of hyperparameters. The analysis from the results based on a case study of DTSS2 Project T10 indicates that the network model is sufficiently accurate for prediction – with MAE and RMSE at 4.63mm/min and 5.76mm/min respectively. Sensitivity analysis suggest that uniaxial compressive strength, duration of excavation, and grout volumes, are key input parameters that contribute most to the accuracy of the model. As such, there can be further studies to examine the corresponding relationships with a view on extending theoretical understanding of the kinematic behaviours and responses of a TBM during excavation. Introducing larger datasets to the network model can enhance the accuracy of the network model. Additionally, the design of the network model can be improved by incorporating theoretical frameworks to augment the hyperparameters and refine the overall performance of the network.