Attention-based LSTM predictive model for the attitude and position of shield machine in tunneling

Shield machine may deviate from its design axis during excavation due to the uncertainty of geological environment and the complexity of operation. This study therefore introduced a framework to predict the attitude and position of shield machine by combining long short-term memory (LSTM) model with attention mechanism. The data obtained from the Wuhan Rail Transit Line 6 project were utilized to verify the feasibility of the proposed method. By adding the attention mechanism into the LSTM model, the proposed model can focus more on parameters with higher weights. Sensitivity analysis based on Pearson correlation coefficient was conducted to improve the prediction efficiency and reduce the irrelevant input parameters. Compared with LSTM model, LSTM-attention model has higher accuracy. The mean value of coefficient of determination (R2) increases from 0.625 to 0.736, and the mean value of root mean square error (RMSE) decreases from 3.31 to 2.24. The proposed LSTM-attention model can provide an effective prediction for attitude and position of shield machine in practical tunneling engineering.