Knowledge-assisted deep transfer learning for muck clogging identification during mechanized shield tunneling

Predicting the risk of muck clogging in shield tunneling is challenging due to data scarcity and class imbalance. This study develops a knowledge-assisted transfer learning strategy to tackle this issue. The muck-clogging evaluation experience stored in a well-established risk diagram is extracted by constructing extensive low-fidelity data. Beyond the feature types considered in the diagram, additional tunnelling-related features are incorporated by using kernel density estimation, following their actual distribution space unveiled by high-dimensional data visualization. These synthesized data serve to pre-train machine learning models, which are subsequently fine-tuned with high-fidelity on-site data. Results show that transfer learning significantly improves prediction accuracy and efficiency, with knowledge consistency being crucial for success. Furthermore, the study demonstrates robust generalization when the test set surpasses the training set but stays within the source data range. The proposed method offers an innovative solution for broadening practical applications of data-driven models in complex engineering problems.