Abstract: Traditional machine learning models often fall short in accurately predicting shield tunneling posture because of performance limitations when increasing network depth. To address this issue, a shield posture prediction method that leverages a deep residual long short-term memory(LSTM) model is proposed. This method integrates residual connections into the LSTM framework to address network degradation and enhance the model′s ability to learn long-term dependencies in shield tunneling time-series data. Additionally, a Bayesian optimization algorithm is employed to fine-tune hyperparameters and optimize the shield posture prediction model. Validation conducted in a realworld shield tunneling project in Zhejiang demonstrates that the deep residual LSTM model outperforms conventional LSTM and artificial neural network models. Taking the shield tail horizontal deviation prediction as an example, the deep residual LSTM model has a determination coefficient (R²) of 0.90 and a mean absolute error(MAE) of 0.76 mm. In comparison, the LSTM model yields an R² of 0.64 and MAE of 1.08 mm, whereas the artificial neural network model shows an R² of 0.68 and MAE of 1.93 mm. Furthermore, compared to the LSTM model, the deep residual LSTM model can effectively utilize more network layers (from 5 to 8 layers), demonstrating the significant role of residual connections in preventing network degradation and improving feature learning.

Key words: shield tunnel, long short-term memory, residual connection, machine learning, Bayesian optimization, attitude prediction