A Case Study: Monitoring and Prediction for Convergence of Shield Tunnel with Wireless Sensor Network and Long Short-Term Network.

Convergence is key indicator of tunnel security and it is of great significance to predict the trend of convergence and give early warning in time based on the previous monitoring data. Wireless tilt sensors have the advantages of real-time monitoring, flexibility and ease of installation, no interruption to tunnel operation and no demand of inter-visibility. A wireless sensor network system is placed in Shanghai Metro Line 1 and Line 2 and the over-10-months returned monitoring data is collected. This paper builds a prediction model for convergence variation of tunnel in soft soil based on wireless tilt monitoring returned data with Long Short-term Memory (LSTM) network and proposes appropriate prediction time step and sampling interval. LSTM method is introduced to analyze the returned monitoring data time series. Prediction results shows that there is a spearman's ρ greater than 0.7 and a Pearson correlation coefficient greater than 0.8 between monitored data and the prediction values 7 steps ago, which demonstrates the prediction model's feasibility. By utilizing the data by certain interval to simulate different sampling interval, the network shows better performance with lager interval and make more detailed prediction with small interval. This paper demonstrates that LSTM network can meet the requirement of accuracy and cost with time steps 7 and sampling interval 12 hours.

Keywords: convergence monitoring, shield tunnel, wireless sensors network, Long Short-term Memory(LSTM) network