Multi-fidelity knowledge inheritance with active querying for data-driven clogging prediction during mechanized tunneling

Abstract

Muck clogging during shield tunneling often leads to reduced construction efficiency, increased costs and potential safety hazards. Traditional methods for predicting muck clogging primarily rely on the operator’s experience and conventional risk maps, but have limitations in dealing with complex construction conditions. To address these issues, this study presents a Monte-Carlo dropout (MCD)-assisted multi-fidelity neural network (MFNN) framework for effective prediction of muck clogging risk. First, a low-fidelity model is trained based on synthesized data using clogging risk maps. Subsequently, in-situ tunneling data are used as high-fidelity data to train multi-fidelity models. MCD serves to evaluate the uncertainty of the MFNN’s inference, combined with an active learning strategy to refine the low-fidelity model via iterative training of the high-fidelity model. Experimental results show that the MCD-assisted MFNN framework captures clogging features more effectively than traditional machine learning models that use only single-fidelity data, especially in scenarios with imbalanced data. This study provides a viable solution for complex problems in shield tunneling by fully utilizing both experiential knowledge accumulated in engineering practice and field monitoring data, demonstrating the potential of integrating knowledge and data in tackling some challenges that were previously unresolved.