Establishment and Empirical Study of Mud Cake Formation Risk Prediction Model for Shield Machines Based on Imbalanced Samples

Objective Predicting the cutterhead mud cake formation based on shield machine data is of significant value for ensuring tunnel construction safety and improving construction efficiency. Traditional machine-learning models struggle to effectively capture minority-class features when handling such small-sample datasets, causing the model to lean towards learning the majority classes while neglecting the minority ones, thereby affecting early-warning performance. Therefore, it is necessary to establish a mud cake formation risk prediction model for shield tunneling based on imbalanced samples and conduct empirical validation.

Method First, the feature engineering is applied to remove shutdown data and identify stable excavation segments. Then, key features for mud cake prediction are selected through feature-importance evaluation and correlation analysis. On this basis, Focal Loss function is embedded into an LSTM (long and short-term memory network) to enhance the model attention to minority-class samples. An actual metro shield tunneling project in Changchun City is used as a case study to empirically verify the model prediction accuracy.

Result & Conclusion  The constructed preprocessing framework for EPB (earth pressure balance) shield raw tunneling data effectively improves the data quality. Through orthogonal tests, the optimal hyperparameter combination for the Focal Loss function is determined: modulation factor γ = 1.000 and class weight corresponding to the true class α_z = 0.750. Under the same dataset and hyperparameter settings, the traditional LSTM model achieves a performance evaluation indicator F₁-score of 0.724, whereas the LSTM model using Focal Loss function achieves an F₁-score of 0.982. The increase in F₁-score indicates that introducing the Focal Loss function effectively enhances the model prediction performance for imbalanced samples.