针对盾构姿态预测模型存在易过拟合、预测精度低的问题，提出一种基于融合注意力机制的盾构姿态组合预测模型。为强化有效特征的提取，抑制冗余特征信息的表达，引入基于选择性卷积核网络（selective kernel networks，SKNet）的特征注意力机制提取网络，消除固定尺寸卷积核带来的限制，并自适应形成带有注意力的特征映射。为更好地捕捉长期信息和特征模式，通过双向长短期记忆网络（bidirectional long short-term memory，BiLSTM）、门控循环单元（gated recurrent unit, GRU）得到2组隐含输出结果，再利用多头注意力机制，捕获组合模型输出的隐含特征与模型输出的盾构姿态之间的依赖关系，进一步提高预测模型对重要隐含特征的信息抓捕能力； 同时,为解决地质勘察钻孔数据连续性差、精确性不足，难以应用于机器学习模型训练的问题，将基于人工先验知识的二级特征引入模型特征输入，提升模型对地层信息的感知能力。最后，基于广州地铁12号线官洲站—大学城北站盾构实例，对模型不同参数结构下的性能进行研究，并进行对比试验验证模型性能，采用可解释性试验评估特征对预测结果的影响。试验结果表明，相比其他预测模型，所提出的预测模型优越性更好，预测精度更高，解决了长时间序列高特征维度数据在传统模型下易过拟合且预测精度较低的问题。

Existing shield attitude prediction models suffer from overfitting and low prediction accuracy. Thus, a combined prediction model for shield attitude based on fusion attention mechanism is proposed. A feature extraction network using a selective kernel network attention mechanism is introduced to enhance the extraction of relevant time-domain features and suppress redundant information. This approach eliminates the limitations of fixed-size convolution kernels and adaptively forms a feature map with time-domain attention. To capture long-term information and feature patterns, two sets of implicit output results are obtained through a bidirectional long- and short-term memory network and a gated recurrent unit. A multihead attention mechanism is then used to capture the dependence between the implicit features output by the combined model and the shield attitude output, further improving the model′s ability to capture critical implicit features. Furthermore, to address the issue of insufficient continuity and accuracy in geological survey drilling data, which complicates its application in machine learning model training, secondary features based on artificial prior knowledge are introduced to improve the model′s perception of stratum information. Finally, the model′s performance is evaluated using a shield example from the Guangzhou metro line 12, specifically between the Guangzhou station and the Higher Education Mega Center North station. The model′s performance is tested under different parameter structures and verified through comparative experiments. Furthermore, the influence of features on the prediction results is assessed through interpretable experiments. The experimental results demonstrate that the proposed model outperforms other prediction models, showing improved accuracy and addressing the issues of overfitting and low prediction accuracy with long time-series, high-dimension feature data in traditional models.