基于机-岩状态识别和融合注意力的盾构姿态多步预测  

Multi step prediction of shield tunneling posture based on machine-stratum state recognition and fusion attention

作　　者：熊栋栋 刘哲 许超 XIONG Dongdong;LIU Zhe;XU Chao(CCCC Second Harbour Engineering Company Ltd.,Wuhan 430040,China;Research and Development Center of Transport Industry of Intelligent Manufacturing Technologies of Transport Infrastructure,Wuhan 430040,China;Key Laboratory of Large-Span Bridge Construction Technology,Wuhan 430040,China)

机构地区：[1]中交第二航务工程局有限公司,湖北武汉430040 [2]交通运输行业交通基础设施智能制造技术研发中心,湖北武汉430040 [3]长大桥梁建设施工技术交通行业重点实验室,湖北武汉430040

出　　处：《铁道科学与工程学报》2024年第9期3809-3821,共13页Journal of Railway Science and Engineering

基　　金：国家自然科学基金资助项目(52379114);湖北省重点研发计划项目(2023BAB094)。

摘　　要：为准确预测未来多个时刻的盾构姿态,帮助盾构操作人员提前识别盾构姿态变化趋势,以进行早期决策,提出基于机-岩状态识别和融合多尺度的特征-时域注意力机制的知识-数据双驱动盾构掘进姿态多步预测方法。引入实时反映机-岩工作状态关系的场切入指标FPI、单位贯入度的刀盘扭矩指标TPI、分区可掘性指标FPIR、螺旋机性能指标STP等复合参数指标,作为模型的输入特征参数,并构建以GRU为基础单元的Encoder-Decoder网络结构。在特征维度上,采用融合多尺度一维卷积的特征注意力机制,自适应捕捉不同层次、不同尺度的盾构掘进参数特征;时间维度上,在解码器中引入时域注意力机制,充分挖掘盾构掘进数据中的长期历史信息和短期输入输出序列的隐藏依赖关系。广州地铁12号线盾构掘进历史数据的模型测试结果分析表明,该预测方法在不增加网络复杂性的前提下,既缓解了算法自身在寻优、迭代和匹配时产生的可解释性差、效率低等问题,又大幅提升模型的特征提取、时间序列的相关性捕获及长期趋势挖掘的能力,实现盾构掘进姿态的精准多步预测,其性能明显优于门控循环神经网络LSTM、GRU及其经典组合模型GRU-SelfAttention、GRU-MultiheadAttention等。研究结果为进一步完善盾构掘进姿态预测方法、提升盾构掘进姿态优化控制水平提供参考。To predict the shield tunneling posture at multiple future moments and help shield operators identify the trend of shield tunneling posture changes in advance for early decision-making.A knowledge data dual drive shield tunneling posture multi-step prediction method based on machine-formation state recognition and fusion of multi-scale feature attention mechanism time-domain attention mechanism was proposed.The method introduced composite parameter indicators that reflects the real-time relationship between machine-stratum working state such as field penetration index FPI,cutterhead torque per unit penetration index TPI,partition excavatability index FPIR,and screw machine performance index STP as input feature parameters for the model.An Encoder Decoder network structure based on GRU was constructed,with the features as the basis unit.In terms of dimensions,a feature attention mechanism that integrates multi-scale CNN was adopted to adaptively capture different levels Characteristics of shield tunneling parameters at different scales.In terms of time dimension,a time attention mechanism was introduced into the decoder to fully explore the hidden dependencies between long-term historical information and short-term input-output sequences in shield tunneling data.Through the analysis of model testing results on historical data of shield tunneling on Guangzhou Metro Line 12,it is shown that this prediction method not only alleviates the problems of poor interpretability and low efficiency generated by the algorithm itself during optimization,iteration,and matching without increasing network complexity,but also significantly improves the model’s ability to extract features,capture correlation in time series,and mine long-term trends,achieving accurate multi-step prediction of shield tunneling posture.Its performance is significantly better than that of gated recurrent neural networks LSTM,GRU,and their classic combination models GRU SelfAttention,GRU MultiheadAttention,etc.The research results can provide reference

关 键 词：盾构 盾构姿态 多步预测 复合参数指标 一维卷积 特征注意力 时域注意力 

分 类 号：U455.43[建筑科学—桥梁与隧道工程] TP18[交通运输工程—道路与铁道工程]