【摘要】 隧道掘进机(TBM)开挖的深埋隧道岩爆灾害频发，对施工安全造成极大威胁．针对深埋TBM隧道岩爆案例分散、预警模型固化和预警时效性不足等问题，本研究集成多个TBM隧道工程岩爆案例、机器学习和智能算法，自主研发了深埋TBM隧道岩爆智能动态预警系统．系统构建了包含TBM隧道岩爆基本信息、地质特征及微震监测指标的专用岩爆数据库，设计了项目及工区管理、岩爆案例和模型管理、预警结果管理等多个功能，实现了不同TBM工程岩爆案例分区块存储、统一管理、多源数据动态整合和更新．基于极致梯度提升(XGBoost)、贝叶斯优化(BO)和超参数带宽(Hb)优化构建了TBM隧道岩爆智能预警模型，将准确率从基础XGBoost模型的78.2%提升至83.4%(BO-XGBoost优化模型)和84.0%(Hb-XGBoost优化模型)，实现预警准确率的有效提升．此外，通过数据库岩爆案例调用和模型增量更新实现了预警模型动态更新机制，系统响应时间控制在30～60 min，且更新过程不影响系统的使用，这使得模型具备了持续学习的能力，其预警效果将随案例的积累不断提升．在中国西部某深埋TBM隧道的工程应用表明，该系统相比传统预警方式，有效提升了预警准确率和时效性，具有显著的实用价值与预警效能．研究成果可为深埋TBM隧道岩爆防控提供依据，以期推动岩爆预警向自动化、智能化发展．更多还原

【Abstract】 Rockbursts occur frequently in deep-buried tunnels excavated by tunnel boring machines(TBMs),posing a significant threat to construction safety. To address the issues such as scattered rockburst cases in deep-buried TBM tunnels,fixed early-warning models and insufficient early-warning timeliness,multiple rockburst cases from TBM tunnel projects,machine learning and intelligent algorithms are integrated to develop a deep-buried TBM tunnel rockburst intelligent dynamic early-warning system. This system establishes a specialized rockburst database,which includes the basic information,geological characteristics and microseismic monitoring indexes for TBM tunnel rockbursts. It features multiple functions such as project and work area management,rockburst cases and model management,and early-warning result management,enabling block storage,unified management,and dynamic integration and updates of multi-source data for different TBM engineering rockburst cases. An intelligent early-warning model for TBM tunnel rockbursts is constructed using extreme gradient boosting(XGBoost),Bayesian optimization(BO),and Hyperband(Hb) optimization. The accuracy rate is improved from 78.2% for the basic XGBoost model to 83.4% for the BO-XGBoost optimized model and 84.0% for the Hb-XGBoost optimized model,achieving a significant improvement in early-warning accuracy. Additionally,the system implements a dynamic update mechanism for the early-warning model through the invocation of rockburst cases from the database and incremental model updates. The system’s response time is controlled within 30??60 min,and the update process does not affect the system usage. This endows the model with the capability of continual learning,enabling its early-warning performance to improve progressively with the accumulation of cases. The application of this system in a deep-buried TBM tunnel in western China demonstrates that,compared with the traditional early-warning methods,the system significantly improves both the accuracy rate and timeliness,showing remarkable practical value and early-warning effectiveness. The research findings in this paper can provide a basis for rockburst prevention and control in deep-buried TBM tunnels,aiming to promote the automation and intelligence of rockburst early-warning systems.更多还原