Prediction of disc cutter wear of shield machines based on transfer learning

Disc cutters serve as the primary device in shield machines for rock breaking during tunnel construction. Assessing the wear state of disc cutters is crucial for making timely replacement decisions. Several researchers have successfully predicted disc cutter wear with acceptable accuracy using Machine Learning (ML). However, ML models are often project-specific. The model trained on one project cannot be applied to the other project if the two projects have significant deviations, resulting in a waste of resources and effort. To address this issue, we propose a domain-adversarial-based transfer learning method to improve the generalization performance of ML models. In particular, we integrate the Domain-Adversarial Neural Network (DANN) with the Transformer. The proposed model makes domain discrimination and regression prediction for input parameters. The domain-adversarial mechanism makes the extracted features from input parameters share many commonalities and confuses data from different domains, which can improve the generalization performance of the model. The hyperparameter λ of the proposed model is used to balance the importance of domain discrimination and regression prediction. We validate the effectiveness of the proposed method in the second subsea tunnel in Qingdao, China. The south and service tunnels of the project are in similar strata conditions but have a significant difference in tunnel diameter. They share many commonalities in the wear characteristics of disc cutters. We set the service tunnel as the source domain and the south tunnel as the target domain. The model is trained and tested in the service tunnel, learning wear characteristics under different strata. Then, the pre-trained model is transferred to the south tunnel and fine-tuned with limited data to adapt to wear characteristics under different shields. Finally, the fine-tuned model is used to predict the wear values of the target data in the south tunnel. The domain-adversarial-based Transformer model outperforms FCN, LSTM and Transformer without the domain-adversarial mechanism and requires even limited target project data, unlike traditional models, which require extra training data. The proposed method can be applied in the early stage of related projects when data are scarce.