为解决高黏性地层条件下盾构结泥饼程度难以精细化分级及预警能力不足的问题，以长春地铁5号线硅谷大街—繁荣路西区间为研究对象，提出一种融合K-means聚类与随机过采样（random oversampling, ROS）、沙普利可加解释（SHapley Additive exPlanations,SHAP）值法与长短期记忆网络（long short-term memory, LSTM）的结泥饼分级与智能预警框架。首先，筛选与结泥饼演化高度相关的掘进参数和渣土改良参数，采用K-means聚类方法自动划分结泥饼等级，并对分级结果的合理性与工程意义进行验证；其次，基于聚类结果构建结泥饼等级LSTM时序预测模型，同时引入ROS策略缓解样本类别不平衡问题；最后，使用SHAP值法对模型的预测行为进行解释。研究结果表明： 1）K-means聚类方法划分了4类结泥饼等级，分级结果具有较强的工程可解释性； 2）相较于LSTM和人工神经网络（artificial neural network, ANN）模型，ROS-LSTM模型在各预测时步下均取得了最高的加权F1分数，并且在各类别间实现了更均衡的性能，其中，将结泥饼类别误判为正常掘进的样本比例始终低于6%，可为实际施工提供多时步结泥饼预警参考； 3）SHAP分析结果表明，各特征均有效参与结泥饼分级决策，模型的决策行为与结泥饼演化过程中的工程规律一致，进一步验证了特征选取与聚类分级结果的合理性。
Accurate quantification and classification of muck caking remain challenging, limiting the effectiveness of early warning during earth pressure balance shield tunneling in a highly cohesive stratum. To address these challenges, a case study was conducted on a section between Guigu Street Station and West Fanrong Road Station of the Changchun Metro Line 5, and an intelligent muck caking grading and early warning framework integrating K-means clustering, random oversampling (ROS), Shapley additive explanations (SHAP), and long short-term memory (LSTM) networks is proposed. First, tunneling parameters and muck conditioning parameters highly correlated with the evolution of muck caking were screened and selected. K-means clustering was then used to automatically classify the muck caking grade, and the rationality and engineering significance of the grading results were verified. Based on the clustering results, an LSTM time-series model for predicting muck caking grades was constructed, and the ROS strategy was introduced to mitigate class imbalance problems for the samples. Finally, the effectiveness of the model prediction was evaluated using SHAP analysis. The results show the following: (1) The K-means clustering method divides muck caking into four grades, and the grading results present strong engineering interpretability. (2) Compared with the LSTM and artificial neural network models, the ROS-LSTM model achieves the highest weighted F1 scores at all prediction time-steps and delivers more balanced performance across all categories. Specifically, the false-negative rate of misclassifying muck caking samples as normal tunneling samples is consistently below 6%, which can provide a reference for multistep early warning of muck caking in onsite construction. (3) SHAP analysis indicates that all factors contribute to the muck caking grading classification, and the decision-making behavior of the model is consistent with the engineering laws in the evolution of muck caking, further verifying the rationality of feature selection and clustering-based grading results.
