盾构岩机作用载荷（盾构推力、刀盘转矩）对于装备设计、现场施工等尤为重要，但现有理论或统计模型参数多、计算繁杂，为提升盾构数字化设计、数字孪生水平，高效精准地进行载荷赋值仍需探索。针对已有机器学习算法重视各地层比例与力学参数等，却忽视地层空间位置关系的现状，融合图像、数据序列进行多特征信息表征，提出一种基于卷积神经网络(CNN)、双向长短期记忆(BiLSTM)与注意力机制(Attention)的数据驱动型超大直径泥水盾构岩机作用载荷预测方法。首先，利用K-means算法对地质纵剖面图进行地层信息提取并切片，对切片图像进行CNN特征提取，结合装备运行数据形成多特征融合数据集；其次，采用BiLSTM网络提取输入特征的双向时序信息，以注意力机制为后续特征进行权重再分配；最后，利用灰狼优化算法（gray wolf optimizer,GWO）对深度学习网络超参数寻优。研究结果表明： GWO-CNN-BiLSTM-Attention模型相比BP（BP 神经网络）、PSO-BP（粒子群优化的BP神经网络）、LSTM（长短期记忆网络）对盾构推力、刀盘转矩的预测效果更佳，R2均超过0.9，平均相对误差＜1%；由于融入了地层位置信息，模型对岩机作用载荷问题具有更强的适用性。
Interaction loads (shield thrust and cutterhead
torque) between rock and shield are essential for equipment design and
tunneling efficiency. Existing theoretical and statistical models involve
numerous parameters and are computationally cumbersome. Further research is
required to improve the efficiency and accuracy of load assignment in the
digital design and digital twin of the shield. Although machine learning
algorithms emphasize the proportion and mechanical parameters of geological
formation, they often overlook the spatial relationships of strata. To address
this, a data-driven load prediction method is proposed between rock and machine
interaction based on a convolutional neural network (CNN), a bidirectional long
short-term memory (BiLSTM), and an attention mechanism
to characterize multifeature information from images and data sequences. First,
the K-means algorithm is employed to extract and slice
the geological longitudinal profile. Features are extracted from the sliced
images using CNN, and a multifeature fusion dataset is formed by combining
equipment operation data. Second, the BiLSTM network is employed to extract
bidirectional time series information of the input features, and the attention
mechanism is used to redistribute the weights of subsequent features. Finally,
the grey wolf optimizer is used to optimize the hyper parameters of the deep
learning network. The results indicate that the GWO-CNN-BiLSTM-Attention model exhibits better
predictive performance for shield thrust and cutterhead torque—with R2 values exceeding 0.9 and an average relative error below 1%—than the BP, PSO-BP, and LSTM models. By
incorporating geological spatial information, the proposed model offers
improved applicability for load prediction in rock-machine
interactions.
