Artificial intelligence for tunnel boring machine penetration rate prediction

Abstract

Penetration rate prediction of Tunnel Boring Machines (TBM) is critical for understanding excavation performances. In this paper, we investigate the possibility of developing machine learning models that accurately predict the Penetration Rate of a TBM using only machine parameters. We leveraged two datasets collected from the Exploratory and Main excavation of the Lot Mules 2–3 of the Brenner Base Tunnel Project. We compared the performance of two different Artificial Neural Network architectures, one based on feedforward architecture and the other on long short-term memory (LSTM). We also studied which features lead to a good estimation of the penetration rate using SHAP, an explainable AI tool, discovering that the Specific Energy (SE) and the Cutterhead Power (CP) are the most impactful features. We also explore the possibility of performing cross-tunnel prediction by training the model on the Exploratory tunnel and testing it on the Main Tunnel, obtaining promising results.