Ada-attention mechanism for intelligent parameter optimization in TBM rock fragmentation: A deep learning approach

Wencan Guana,b,f, Suran Wanga,c,*, Youliang Chena,e, Siyu Chend

a Department of Civil Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China

b Department of Engineering Geology and Hydrogeology, RWTH Aachen University, Aachen D-52064, Germany

c Key Laboratory of Geotechnical and Underground Engineering of the Ministry of Education, Tongji University, Shanghai 200092, China

d College of Civil and Transportation Engineering, Shenzhen University, Shenzhen 518060, China

e Department of Engineering Geology, School of Civil and Environmental Engineering, University of New South Wales, Kensington, NSW 2052, Australia

f Institute of Structural Mechanics (ISM), Bauhaus University Weimar, Weimar 99423, Germany

Abstract

Tunnel boring machine (TBM) rock breaking parameter optimization is a technical challenge in underground engineering. Traditional numerical simulation methods have limitations in computational efficiency and accuracy when dealing with multi-parameter coupling optimization under complex geological conditions. This study proposes a deep learning method based on the Ada-Attention mechanism for predicting and optimizing parameters such as confining pressure, penetration depth, and cutting tool spacing in TBM rock breaking processes. The method employs a hybrid attention architecture that combines global window mechanisms with local sliding windows, reducing the computational complexity of traditional self-attention mechanisms from O(n²) to O(n(w+α)). Additionally, the Newton-Gauss optimization algorithm is introduced to improve the softmax normalization process, enhancing numerical stability and convergence performance. The research constructs a prediction framework covering a temperature range from 25°C to 500°C, using 800 experimental samples for model training and validation. Experimental results show that the Ada-Attention model achieves R² values of 0.92, 0.93, and 0.94 for torque, rolling force, and specific energy predictions respectively, obtaining 2–10 times computational speedup compared to traditional Transformer architectures. Generalization capability validation demonstrates that the model exhibits high prediction accuracy in soft sedimentary rock and medium sandstone (R²>0.95), maintains moderate performance levels in hard limestone and crystalline rock (R²=0.80–0.90), while prediction accuracy decreases in complex geological environments such as composite formations and fractured rock masses. This method provides a feasible technical solution for TBM parameter optimization under complex geological conditions.

主要内容

提出了一种基于Ada-Attention机制的深度学习方法，用于隧道掘进机（TBM）破岩过程中的参数智能优化。研究以围压、贯入度和刀具间距为输入变量，以扭矩、滚动力和比能为输出指标，结合PFC3D模拟与实验数据进行建模。通过引入全局-局部混合注意力机制和Newton–Gauss优化的Softmax归一化算法，模型在保持高精度（R²=0.92–0.94）的同时，将计算复杂度从O(n²)降低到O(n(w+α))，计算速度提升2–10倍。研究还将**温度梯度（25–500°C）**及其对岩石和刀具性能的影响纳入模型，构建了多场耦合预测体系。结果显示，该模型在软岩和中硬岩条件下具有较强的泛化能力，可为复杂地质环境下的TBM参数优化与智能掘进提供高效可靠的技术方案

创新点

  • 提出基于Ada-Attention机制的TBM参数智能优化框架；

  • 实现了自适应的局部-全局注意力融合与Newton-Gauss优化Softmax；

  • 将热-力-化多场耦合效应纳入TBM破岩预测模型；

  • 显著提高了预测精度与计算效率，为复杂地质条件下的智能掘进提供新方法

引用格式

Guan W., Wang S., Chen Y., et al. (2025). Ada-attention mechanism for intelligent parameter optimization in TBM rock fragmentation: A deep learning approach. Intelligent Geoengineering, 2, 192–215. https://doi.org/10.1016/j.ige.2025.10.001