针对复杂地质环境中全断面岩石隧道掘进机（tunnel boring machine,TBM）施工过程中易发生卡机风险，且传统经验判识方法实时性不足、不同卡机类型难以协同识别等难题，构建一种基于Transformer多任务学习的敞开式TBM卡机风险实时智能预警模型。建立Expert System风险判识规则，采用特定阈值和多指标综合判定方法对掘进数据进行分类，将卡刀盘、卡护盾风险分为0—低风险、1—中风险、2—高风险3个等级。利用Transformer模型的多头注意力机制，分别构建隧道A和隧道B的多任务学习框架，实现卡刀盘与卡护盾2类风险等级的同步预测。其中： 卡刀盘风险预测准确率为97.01%，比次优的门控循环单元网络（gated recurrent unit，GRU）模型高2.73%，精确率为98.21%，召回率为98.44%，受试者工作特征曲线下面积（area under curve,AUC）达到1； 卡护盾风险预测准确率为89.55%，比次优的GRU模型高5.90%，精确率为90.35%，召回率为91.70%，AUC为0.972 6。将模型应用于2座敞开式TBM施工隧道卡机风险进行实时判识与智能预警，结果表明： 风险预测准确率为90%，与现场施工情况吻合较好，能够在卡机发生前实现稳定有效的风险预警。

When full-face hard rock tunnel boring machines (TBMs) bore through complex geological conditions, they are prone to jamming. Traditional experience-based identification methods cannot provide real-time assessment and are unable to collaboratively identify multiple jamming modes. To address these challenges, a real-time intelligent early warning model for open TBM jamming risk was developed based on Transformer multi-task learning. An expert-system-based risk identification rule set is established, in which tunneling data are classified using predefined thresholds and comprehensive multi-indicator evaluation. Cutterhead and shield jamming risks are classified into three levels: 0 (low risk), 1 (medium risk), and 2 (high risk). Based on the Transformer’s multi-head attention mechanism, multi-task learning frameworks are constructed for Tunnels A and B, respectively, enabling simultaneous prediction of cutterhead and shield jamming risk levels. The results are as follows. (1) The prediction accuracy for cutterhead jamming risk reaches 97.01%, which is 2.73% higher than that of the second-best gated recurrent unit (GRU) model. The precision and recall reach 98.21% and 98.44%, respectively, and the area under curve (AUC) reaches 1. (2) For shield jamming risk prediction, the accuracy reaches 89.55%, which is 5.90% higher than that of the second-best GRU model. The precision and recall reach 90.35% and 91.70%, respectively, and the AUC reaches 0.972 6. Application of the proposed model to two open TBM tunnels shows that the jamming risk prediction accuracy reaches 90%, which agrees well with field construction conditions and enables stable and effective early warning before jamming occurs.