Tunnel-face rock mass classification using transfer learning based on multi-task adversarial network

Tunnel-face rock mass (TFRM) is the object in direct contact with the shield machine during shield tunnelling. Reliable TFRM classification enables shield drivers to adjust tunnelling parameters appropriately. Machine learning (ML) models have achieved acceptable classification performance. Still, their parameters are often highly correlated with the data distribution of the training sets, limiting their generalization to other projects with different data characteristics. To address this issue, this paper proposes a transfer model integrating multi-task learning and adversarial training to learn commonalities across projects. The commonalities possess continuous and discrete characteristics that support regression and classification and are weakly dependent on training set distribution, thereby improving model generalization. A key hyperparameter λ balances the importance of different tasks. The effectiveness of the proposed model is validated on the second subsea tunnel project in Qingdao, China, where the south and service tunnels share similar strata conditions but differ substantially in tunnel diameter and shield configuration. The model is trained on the service tunnel and transferred to the south tunnel, with DNN, CNN, and RF serving as baselines. Results show that the F1 score of MTAN is 50 % higher than that of the baselines without fine-tuning, and it saved >10 % of data with the same performance. Ablation experiments are conducted to quantify the contributions of key components, thereby supporting the reliability of the proposed model. These results indicate that the proposed model is suitable for early-stage deployment in related projects when data are scarce.