Search

Tunnelling in urban areas should be aware of deviations from the design tunnel axis, referred to also as 'misalignment'. Significant misalignment can cause unfavourable soil deformations, threatening adjacent properties, and sustainable development of underground space. In this study, the potential of applying machine learning (ML) and deep learning (DL) algorithms to EPB shield posture prediction and anomaly detection was explored. A framework, including data pre-processing, development of prediction models, performance evaluation, and anomaly detection, was proposed to tackle the said issue. The principal component analysis (PCA) was introduced not only to eliminate outliers but also to isolate salient data features using the Pearson's correlation coefficient (PCC). Also, the partial dependence analysis aimed to marginalize other features and established the relationship between an individual tunnelling parameter and EPB shield position. Then the feature-based subseries were normalized for training and examining the DL-based Long Short-Term Memory (LSTM) and ML-based ensemble EPB shield position predictors. On the other hand, three anomaly detectors founded on three ML algorithms were developed to prevent snakelike advance. Results showed that the DL-based LSTM predictor outperformed the ML-based ensemble predictor owing to its relative insensitivity to lags of unknown duration between important events. The thrust force dominated EPB shield position. The LSTM model prevented the accuracy of EPB shield position prediction from disturbing by gravels surrounding the design tunnel axis or artificial intervention. [ABSTRACT FROM AUTHOR]

Complex geological conditions and/or inappropriate shield tunnel boring machine (TBM) operation can significantly degrade both the excavation and safety of tunnel construction. In recent years, the excavation behaviour of shield TBMs has been a popular topic in the literature given the large volume of data automatically collected by modern shields. These datasets provide an excellent opportunity to apply advanced data analysis techniques to improve predictions of shield tunnelling excavation behaviour. In this study, a framework to develop machine learning (ML)-based regression models for predicting the behaviour of an earth pressure balance (EPB) shield machine using tunnelling parameters is proposed. The feasibility of four ML algorithms, namely Linear Regression (LR), Decision Tree Regression (DTR), Support Vector Regression (SVR), and Gradient Boosting Regression (GBR), to predict EPB shield excavation behaviour is explored through their application to a recent tunnelling case history in sandy soils. The results show that the misestimates were primarily attributed to a reduction of screw conveyor rotational speed (SCRS), induced by a lower injection volume, the artificial manipulation of penetration rate (PR), the local variations of total jacking load, and the use of 'breakout' cutterwheel torque (CT). The GBR model provided the best performance, while LR often performs the worst due to its inability to handle highly nonlinear relationships. DTR prevented the overfitting problem by using a lower max depth parameter towards sacrificing its accuracy. The performance of SVR was seriously affected by loss functions. The proposed optimisation scheme that prevents the over-smoothing problem during the STL decomposition elevates further the prediction accuracy. [ABSTRACT FROM AUTHOR]