Dynamic prediction and multi-objective optimization on driving position of tunnel boring machine (TBM): an automated deep learning approach.

This paper proposes an automated deep learning (AutoDL) framework for dynamic prediction and multi-objective optimization (MOO) on the driving position of the tunnel boring machine (TBM) to enhance tunneling reliability and efficiency. More specifically, AutoDL contains a standard process of data preprocessing, optimal hyperparameter tuning, algorithm selection, and model performance evaluation, which can intelligently learn the time-varying monitoring data collected by the smart sensors installed on TBM. It can return a prediction interval to well consider uncertainty, which is stable enough to raise the reliability of the prediction results. Later, the model from AutoDL serves as a meta-model for MOO-based decision-making to evidently control the four TBM position objectives, including shield head horizontal deviation, shield head vertical deviation, shield tail horizontal deviation, and shield tail vertical deviation. To validate the effectiveness of the proposed approach, it is applied in a TBM project for a natural gas backbone network between Changxing Island and Chongming Island in Shanghai. It is found that the proposed AutoDL can greatly reduce the requirements of professional knowledge of developers on model training. The deviation of the TBM driving position in each segment ring can be dynamically and accurately estimated in the form of both a single point and interval, reaching the average MAE of 1.051. Based on the intelligent model built by AutoDL, Nondominated Sorting Genetic Algorithm II (NSGA-II) is conducted to feasibly adjust TBM operational parameters. The average deviation of the four optimization objectives of the 66-ring data is decreased by 16.05%, contributing to reducing the occurrence of shield tunneling misalignment (STM). [ABSTRACT FROM AUTHOR]