Breaking the tunnel vision: Generalizing TBM performance prediction across projects

Accurately predicting the Penetration Rate (PR) in tunneling is crucial for evaluating the performance of Tunnel Boring Machines. However, a significant shortcoming of many existing studies is that the predictive models are developed using data from a single project, which limits their applicability to future tunnelling projects. This research addresses this issue by examining whether the model can generalize to other tunnelling projects and investigates the impact of an incremental learning strategy on its performance. In this study, the Extreme Gradient Boosting (XG Boost) model was initially trained and tested using data from one tunnel project and then generalized to a different tunneling project with similar geological conditions. To improve the model’s generalization performance, an incremental learning strategy was used, which involved iterativelyincor porating new data from the second tunneling project into the existing data from the first tunneling projectto update the model. Additionally, the study analyzed how the generalization ability of XG Boost varies with the incremental size of new data. The findings indicate that there is a promising potential for enhancing the model’sgeneralization ability using incremental learning techniques. However, additional research is needed to address the challenges related with skewed data and mitigating the effects of catastrophic forgetting.