Towards Real Time Ground Forecast for TBM Tunneling: Finding Label Errors in Data Sets

Tunnel construction using tunnel boring machines (TBM) is associated with unforeseen ground conditions. There have been several attempts to use data recorded by the TBM to predict ground conditions ahead of the tunnel face and assist in the automation of the tunneling process. One of the main issues associated with such data, in particular for closed face machines, is that the operator has no view of the ground ahead, compromising ground labelling, which is often estimated from face mappings done at spaced intervals. The use of these “noisy” labels to train ground forecast models affects their performance. In this paper, confident learning (CL) is applied to investigate and detect label errors in a data set from a TBM tunnel at the Porto Metro Project, in Portugal. The detected mislabeled points are further investigated to determine the possible reasons for their mislabeling. Limitations, challenges, and directions for future research are discussed at the end.