Intelligent early warning system for tunnel boring machine jamming under complex geology

Jamming events during construction using a tunnel boring machine (TBM) are rare but extremely high-risk, posing a significant threat to construction safety, project progress, and cost control. As abnormal samples are extremely scarce, conventional supervised learning methods are insufficient for providing an effective early warning system. This study therefore proposes the TBM jamming early warning theory (TJ-EWT), an unsupervised framework that identifies risk based on reconstruction deviations from normal operational data. Building on the TJ-EWT, we have developed the U-shaped tabular autoencoder (UTabAE). UTabAE employs an end-to-end encoder–decoder architecture that is tailored to multivariate TBM operational data, and is capable of capturing nonlinear coupling relationships and latent feature patterns. Exclusively trained on normal data, UTabAE detects potential anomalies by monitoring real-time reconstruction errors. Moreover, a 60-s sliding time-window strategy dynamically computes reconstruction deviations, thereby reducing false alarms caused by transient fluctuations. Experimental results demonstrate that UTabAE significantly outperforms baseline models such as MLP, LSTM, GRU, and BiLSTM in terms of reconstruction accuracy and generalization performance, and exhibits strong early-warning capability. Finally, this study uses empirical analysis of two typical jamming scenarios, i.e. soft-rock deformation and fault-fracture-zone disturbances, to identify key early-warning indicators and the features most strongly affected during jamming evolution. The aforementioned variables exhibit sustained deviations and amplified reconstruction errors prior to jamming, thus serving as critical indicators of abnormal states. Overall, this study provides a quantitative foundation and practical guidance for intelligent TBM tunneling, early risk identification, and proactive decision-making under complex geological conditions.