A deep multimodal fusion model for real-time identification of mixed strata in super-large-diameter shield tunnelling

Geological uncertainty is a major risk in shield tunnel construction, especially in upper-soft lower-hard mixed strata, where lithological differences can cause machine faults and construction hazards. Traditional drilling lacks spatial continuity, and single-modality methods based only on tunnelling parameters often show limited accuracy in mixed strata. To address this issue, this study proposes a multimodal fusion deep learning model that integrates numerical modality and image modality for mixed strata identification in shield tunnelling. In the proposed framework, a multilayer perceptron (MLP) is used to extract nonlinear features from tunnelling parameters and vibration root mean square (RMS) values in three orthogonal directions, while a convolutional neural network (CNN) is used to extract spatial features from vibration spectrograms. Middle-level feature concatenation is adopted to achieve deep fusion of heterogeneous data. To obtain realistic vibration responses, triaxial accelerometers were installed in the central chamber of a super-large-diameter shield machine. Under a chronologically constrained evaluation protocol, the proposed model achieved an average test accuracy of 88.71% ± 7.50% across the blocked temporal forward-chaining folds. On the independent later-stage test set, it achieved an overall accuracy of 97.36%. Ablation results show that multimodal fusion outperforms single-modality settings. Horizontal and axial vibration spectrograms contribute most to strata discrimination, and cutterhead thrust is the most sensitive numerical feature. These findings demonstrate the feasibility and effectiveness of multimodal fusion and vibration data for mixed strata identification.