关键词: 盾构隧道, 时空特性, 沉降预测, 深度学习, 自注意力机制

Abstract: To improve the prediction accuracy of soil settlement induced by shield tunneling excavation, a deep bidirectional recurrent neural network model based on self-attention mechanism (SAM-Bi-RNN) is proposed, which can capture spatiotemporal characteristics and vital information from settlement data. The SAM-Bi-RNN model utilizes time series data from multiple sensors as input and adopts the multi-layer bidirectional recurrent neural network architecture to capture the spatiotemporal correlations and long-distance dependencies in the settlement data. The self-attention layers are embedded between the adjacent recurrent layers to strengthen the model for the extraction of crucial data features and the capture of internal autocorrelations. The prediction effects of two variant models including bidirectional long short-term memory (Bi-LSTM) and bidirectional gated recurrent unit (Bi-GRU) on soil settlement before and after the introduction of self-attention mechanism are compared. The dataset collected from a tunnel project in Hangzhou is selected to verify the SAM-Bi-RNN model. The results exhibit that: compared with other models, the SAM- Bi-LSTM model has the best predictive performance at various monitoring points, with a total average absolute error of 0.036 6 mm and total average root mean square error of 0.034 8 mm. Furthermore, the mean absolute percentage error of SAM-Bi-LSTM model for each monitoring point in the dataset of a tunnel project in Suzhou is below 7.0%, indicating good generalization. Statistical analysis of the confidence interval shows that the results have 95% reliability. However, it is advisable to fine-tune hyperparameters according to the engineering needs to meet the prediction accuracy requirements in practical applications.

Key words: shield tunneling, spatiotemporal characteristics, settlement prediction, deep learning, self-attention mechanism