A critical evaluation of machine learning and deep learning in shield-ground interaction prediction

Abstract

The interaction between a shield machine and the ground is a complicated problem involving numerous extrinsic and intrinsic factors. Machine learning (ML) algorithms have been recently employed to predict tunnel-soil interactions. This study introduces a more powerful algorithm termed the deep learning (DL) long short-term memory (LSTM) neural network, to identify the interaction between a shield machine and the ground; this network can predict tunnelling-induced maximum settlement, the longitudinal settlement curve, and shield operational parameters. In addition, the generalisation ability of LSTM is comprehensively compared with that of a conventional ML algorithm—random forest (RF)—based on field records collected from two practical tunnel projects. A standard process of developing an ML or DL algorithm-based model, including the pre-processing of raw data, feature selection, determination of optimum hyper-parameters, and evaluation of prediction performance and generalization ability, was introduced. The results indicated that the RF-based model performs capably in terms of predicting tunnelling-induced maximum settlement, and that the LSTM-based model is suitable for predicting longitudinal settlement curves and shield operational parameters. Furthermore, the generalization ability of the LSTM-based model is better than that of the RF-based model. The strong robustness of the LSTM-based model enables its application in different types of tunnel projects.