A Multivariate Deep Learning-Based Risk Prediction Model for Shield Machine Mud Cake Formation

The study aims to propose a risk prediction model for shield machine mud cake formation based on multivariate deep learning and online learning, addressing the issues of reduced construction efficiency and increased safety risks caused by mud cake formation during shield operation. The paper first preprocesses the data and removes outliers by effectively cleaning the multi-dimensional sensor data collected by the shield machine using an isolation forest and the 3σ criterion, while also eliminating data corresponding to idle states. For multi-modal data—including numerical, textual, and temporal types—feature fusion is achieved through dynamic normalization, Word2Vec word vector extraction, as well as both learnable and sinusoidal-cosine encoding methods. Additionally, a multi-head attention mechanism focusing on the slurry inlet-outlet flow difference is introduced to further capture the instantaneous pressure fluctuations resulting from slurry outlet blockages. Experimental results indicate that when a continuous weighting function is applied to process the slurry inlet-outlet thresholds, the model achieves an accuracy of up to 0.8935 on the test set, demonstrating significant advantages in capturing risk features of the shield machine. This provides an efficient and robust solution for risk early-warning in shield construction.