Data-driven multi-step robust prediction of TBM attitude using a hybrid deep learning approach

Abstract

A robust multi-step TBM attitude prediction approach named convolutional gated-recurrent-unit neural network (C-GRU) is proposed in this research and the random balance design Fourier amplitude sensitivity test method is used for sensitivity analysis to reveal the interaction between input and output of the C-GRU model. A tunnel construction project in Singapore is taken as an example to prove the robustness and effectiveness of the proposed approach. Results indicate that the length of the output sequence of the model can maintain high robustness and accuracy within 21 steps. In the 21-step prediction, the highest R2 can reach 0.9652 while the mean R2 is 0.9004 even though some attitude parameter is with large fluctuations. Each step in the 21-step prediction can maintain a stable accuracy. The data of the past 11 time steps of the TBM attitude parameters are the most sensitive. The proposed method has higher accuracy and robustness than state-of-art time-series based methods.