针对全断面隧道掘进机（TBM）穿越软岩变形段落时支护效率低、围岩变形风险高的问题，构建一种基于注意力卷积神经网络-长短期记忆网络（ACNN-LSTM）融合模型的风险识别方法，旨在实现软岩变形风险的精准分级预警。以某高原铁路隧道为背景，构建“感知-预测-决策”闭环控制体系。通过灰色关联分析从TBM掘进参数和地质参数中筛选核心学习数据；联合围岩变形收敛监测值与支护结构，将软岩变形风险划分为5个等级；通过CNN神经网络提取数据特征，借助注意力机制增强LSTM对时序动态的捕捉能力，实现TBM穿越软岩变形段落的风险识别。研究结果表明： ACNN-LSTM模型的整体准确率达到90.91%，宏观精确率达到90.77%，召回率达到88.66%，综合预测性能显著优于CNN-LSTM模型和LSTM模型，注意力机制能够有效提升关键特征捕获能力，减少长序列信息衰减；同一模型在不同风险等级划分方式后整体准确率不同，说明标签分类方式会影响模型性能与泛化能力; ACNN-LSTM模型可为深埋TBM隧道软岩变形段施工过程中的支护决策供数据驱动依据。

To address the challenges of low support efficiency and high surrounding rock deformation risk during full-face tunnel boring machine (TBM) tunneling in soft rock deformation zones, a risk identification method based on an integrated attention convolutional neural network-long short-term memory (ACNN-LSTM) model is developed. The proposed approach aims to achieve accurate hierarchical early warning of soft rock deformation risks. A case study is conducted on a plateau railway tunnel, and a closed-loop control system of “perception-prediction-decision” is constructed. Core learning data are selected from TBM tunneling parameters and geological parameters through Grey correlation analysis; the risk of soft rock deformation is classified into five levels based on the convergence monitoring values of surrounding rock deformation and support structures; data features are extracted via CNN neural networks, and the attention mechanism is employed to enhance LSTM’s ability to capture temporal dynamics, achieving risk identification for TBMs traversing soft rock deformation sections. The research results show that the overall accuracy of the ACNN-LSTM model reaches 90.91%, with a macro accuracy of 90.77% and a recall rate of 88.66%. Its comprehensive prediction performance is significantly better than that of the CNN-LSTM and LSTM models. The attention mechanism effectively enhances the capture of key features and reduces information decay in long sequences; the model exhibits different overall accuracy under different risk classification methods, indicating that label classification approaches affect model performance and generalization ability. The research results indicate that the ACNN-LSTM model can provide a data-driven basis for support decision-making during the construction of soft rock deformation sections in deep-buried TBM tunnels.