Sequential excavation method (SEM) is commonly used for the underground rock cavern construction. One of the major focuses in the SEM process is the selection of the excavation sequence parameters including the subdivision of cavern cross-section and the round length. In this paper, the parameters of excavation design were going to be optimized by adopting the approximate excavation performance using the response surface generated by artificial neural network (ANN) model. Firstly, the training data was generated using numerical studies. Multi-staged 2D plain strain models were adopted to conduct the numerical simulations, and further associated with tunnel advance processes using the convergence and confinement method (CCM). The parameter studies were involving the studies of rock types, cavern sizes, excavation methods and cavern performance. Then, a 3-layer ANN model was used to mapping the relationship between the excavation design parameters and the tunnel performance. At last, by adopting the proposed ANN model with the optimizing function in EXCEL, a revised excavation chart was proposed to help the engineers to quickly find the optimized sequential excavation parameters.
The sequential excavation method (SEM) is widely used for the construction of rock caverns, shafts and other underground structures. It takes advantage of the capacity of the rock mass to support itself by deliberately controlling and adjusting the stress and deformation field which takes place in the surrounding rock mass during the excavation. Federal highway administration (2009) has proposed four essential processes in the SEM design, include: the classification of ground condition and excavation, the definition of excavation method and support classes, the instrumentation and monitoring, and the ground improvement prior to rock cavern excavation. One of the major focuses in the SEM process is the selection of the excavation sequence parameters including the subdivision of cavern cross-section, the round length (maximum unsupported excavation length) and the supports installation time. Subdividing the cavern cross-sections could heavily reduce the risk of the cavern instability during excavation (Graziani et al., 2005; Lunardi and Barla, 2014; Zhang and Goh, 2012). However, too many subdivisions will increase the required equipment and manpower and thus increase total construction costs.
To optimize the excavation designs, it is important to approximate the performance of tunneling under specified SEM parameters. Response surface method (RSM) has been proposed as a useful method to predict the tunneling performance. It has been studied by researchers to present the performance in explicit form (Lü et al. 2017; Hamrouni et al, 2018). Artificial neural network (ANN) is one of the effective ways to approximate the response surface. It has been widely used in data analysis in civil engineering (Zhao and Ren, 2002; Zhao et al, 2008). Essentially, the network is trained by adapting the weights and biases using optimization methods to minimize the mean square error between the predicted and the target values. Some commercial codes such as MATLAB have provided for convenient use of ANN.
