聚类、迁移学习与集成学习耦合的输水隧洞TBM刀盘推力预测及掘进安全预警

尹鹏博; 何芳婵; 江生金; 成荃; 寇磊

doi:10.13928/j.cnki.wrahe.2026.08.019

水利水电技术（中英文） ›› 2026, Vol. 57 ›› Issue (8) : 260 -273. DOI: 10.13928/j.cnki.wrahe.2026.08.019

工程施工

聚类、迁移学习与集成学习耦合的输水隧洞TBM刀盘推力预测及掘进安全预警

尹鹏博 ¹ ,
何芳婵 ²^,³ ,
江生金 ⁴ ,
成荃 ⁵ ,
寇磊 ⁵

作者信息 +

A Coupled clustering, transfer learning, and ensemble learning approach for cutterhead thrust prediction and excavation safety early warning in water conveyance tunnel TBM operations

Pengbo YIN ¹ ,
Fangchan HE ²^,³ ,
Shengjin JIANG ⁴ ,
Quan CHENG ⁵ ,
Lei KOU ⁵

Author information +

文章历史 +

PDF (20199K)

摘要

【目的】在长距离输水隧洞施工中,精准预测隧道掘进机(TBM)刀盘推力对保障施工安全与提升掘进效率具有重要意义。传统深度学习模型如LSTM、GRU与Transformer虽能提供较高精度,但普遍存在参数量大、训练时间长、现场数据适应性差等问题,难以满足工程现场快速响应与实时更新的需求。【方法】提出了一种聚类、迁移学习和集成学习耦合的长距离输水隧洞TBM刀盘推力快速预测和掘进安全预警框架:首先,通过聚类算法对刀盘推力上升段数据进行变化形态分类,构建子数据库;随后,基于迁移学习训练变化形态分类模型,以提升新数据的适配效率;最后,结合集成学习方法,在各变化形态子数据库中实现高精度刀盘推力回归预测,并引入推力贯入滑动平均前导比(LAFPI)作为敏感特征,以提供有效预警。【结果】结果显示:刀盘推力预测框架能够在保持极低参数量(0.015B)和2分钟训练时间的情况下,实现对超过1000个时间节点的刀盘推力数据的高精度预测(R²>0.85)。【结论】与表现最优的传统大型神经网络模型相比,训练时间缩短了86.9%,静态参数量降低了97.4%。通过融合聚类分析、迁移学习和集成学习技术,显著提升了实时训练和应用效率,具备强大的现场实时预测能力,能够有效应对长距离隧道施工中的复杂地质条件,在提高施工安全性和掘进效率方面具有重要的工程应用前景。

Abstract

[Objective] Accurate prediction of Tunnel Boring Machine(TBM) cutterhead thrust is crucial for ensuring construction safety and improving excavation efficiency in long-distance water tunnel projects. However, traditional deep learning models such as LSTM, GRU, and Transformer, while capable of achieving high prediction accuracy, generally suffer from large parameter sizes, long training times, and poor adaptability to on-site data, making them unsuitable for real-time updates in practical engineering scenarios. [Methods] A rapid prediction and excavation safety early warning framework for TBM cutterhead thrust in long-distance water conveyance tunnels is proposed, based on the coupling of clustering, transfer learning, and ensemble learning. First, clustering algorithms are applied to classify the ascending phase data of cutterhead thrust into different variation forms, constructing sub-databases. Then, transfer learning is used to train variation form classification models, enhancing the adaptability to new data. Finally, an ensemble learning method is employed to achieve high-precision regression prediction of cutterhead thrust within each variation form sub-database, incorporating the Leading Averaged Force-Penetration Index(LAFPI) as a sensitive feature for identifying surrounding rock transitions, thus providing effective early warning. [Results] The result show that the cutterhead thrust prediction framework, with only 0. 015 billion parameters and just two minutes of training time, can achieve high-precision prediction(R² > 0. 85) for cutterhead thrust data containing over 1, 000 time points. [Conclusion] Compared with traditional large neural network models, the training time is reduced by 86. 9%, and the model size is reduced by 97. 4%. The proposed framework, by integrating clustering analysis, transfer learning, and ensemble learning techniques, significantly improves real-time training and application efficiency. It possesses strong real-time prediction capabilities, effectively addressing the complex geological conditions in long-distance tunnel construction, and holds significant engineering application prospects, particularly in enhancing construction safety and excavation efficiency.