Early warning of tunnel collapse based on Adam-optimised long short-term memory network and TBM operation parameters

Abstract

Collapses are common geohazards during tunnel boring machine (TBM) construction under complex geological conditions. This study proposes a tunnel collapse early warning method based on an adaptive momentum estimation optimised long short-term memory (Adam-LSTM) network and TBM operation parameters. Based on the Songhua River water conveyance project, a sample database containing 7538 TBM excavation cycles, three types of geological information, and 18 tunnel collapse statistics is established. A total of 5440 TBM excavation cycles from stable tunnelling sections are used for model training. The key TBM operation parameters in the first 30 s of the parameter-rising phase are used as inputs to the LSTM cell, and the geology data is considered through fully connected layers outside the LSTM cell. Then, the rock-breaking efficiency index (specific energy, Se) of the stable phase is predicted. Compared with the stable tunnelling section, the prediction accuracy of Se in the collapse section decreases to some degree. In collapse area I (i.e. collapses 15–17), by setting the threshold of the statistical indexes based on 30 consecutive predicted Se, an early warning index system for tunnel collapse is constructed. Collapse area II (i.e. collapses 5–7) and collapse area III (i.e. collapse 18) are used to verify the effectiveness of the proposed method.