Towards reinforcement learning - driven TBM cutter changing policies
Tom F. Hansen, Georg H. Erharter, Thomas Marcher
Abstract
Optimizing the cutter changing process for tunnel boring machines (TBMs) is crucial for minimizing maintenance costs and maximizing excavation efficiency. This paper introduces TunnRL-CC, a computational framework that utilizes reinforcement learning to autonomously determine cutter-changing strategies. TunnRL-CC's realistic simulation models cutter wear under varying rock conditions, including hard rock and blockyness. A reinforcement learning agent is trained to learn optimal cutter-changing policies based on a reward function that balances cutter conditions and operational costs. The agent demonstrates innovative decision-making, adapting to changing excavation conditions. TunnRL-CC's proposed methodology significantly differs from traditional cutter changing practices, which rely heavily on operator experience. Although TunnRL-CC has not been applied in practical projects, its theoretical basis and comprehensive computational experiments demonstrate its capability to significantly improve TBM cutter maintenance procedures. • A cutter wear simulation for hard rock tunnel boring machine advance was developed. • A predictive maintenance system for TBM cutters is implemented. • The system is reinforcement learning based and parameter optimization was done. • The learned cutter changing policy considers geotechnical and economic conditions.