An intelligent prediction method for surface settlement of shield tunnel construction based on CNN-BiLSTM-SA
Quan Yin, Yi Zhou, Junying Rao
Abstract
Aiming to solve the issues of weak feature extraction and inability to fully utilize the bidirectional time-series information of data in the existing methods for predicting surface settlement triggered by shield tunnel construction, a method of maximum ground settlement prediction, named CNN-BiLSTM-SA, is proposed by fusing convolutional neural network (CNN), Bi-directional Long Short-Term Memory (BiLSTM) and Self-attention (SA) with comprehensive consideration of the interactions among engineering geological parameters, spatial parameters and shield parameters. In CNN-BiLSTM-SA, CNN is first introduced to extract the nonlinear feature relationships among the input data, and BiLSTM network is applied to extract the bi-directional timing information, then SA is introduced to assign corresponding weights to the features extracted by CNN to effectively capture the key information in the time series. The results indicate that Mean Square Error (MAE), Root Mean Square Error (RMSE), and R-Square (R2) of CNN-BiLSTM-SA are 0.88, 1.06, and 0.92, respectively, which has better prediction accuracy than Peck formula, analytic method, ANN, RNN, SVM, LSTM, BiLSTM.