摘要
针对石窟寺岩体裂隙发育缓慢、影响因素多样化,难以对裂隙发育情况进行预测的问题,提出了一种新的基于深度学习的岩体裂隙发育预测网络(FDPNet:Fracture Development Prediction Network),即并行自注意力机制混合网络。FDPNet通过局部卷积模块和全局循环模块对序列数据的时间相关性进行建模,使模型能准确捕捉到不同时间尺度的时间模式。同时通过引入自注意力机制对多元时序数据中不同序列之间复杂的依赖关系进行建模。在此基础上,利用传统的自回归模型进一步提高模型的鲁棒性。此外,以Q市北石窟寺32号窟6个月的裂隙发育相关因素监测数据构建了国内外首个该领域的数据集。在该数据集上进行的对比实验结果表明,该模型在石窟寺岩体裂隙发育预测场景下具有更好的性能表现。
Aiming at the problem that the development of fissures in the rock mass of grotto temple is slow and the influencing factors are diverse,it is difficult to predict the development of fissures.A new prediction network for the development of fissures in rock mass based on deep learning is proposed.It is a hybrid network with parallel self-attention mechanism.It models temporal correlations through local convolution modules and global recurrent modules to capture temporal patterns at different time scales accurately.Self-attention mechanism is introduced to model the complex dependencies between different sequences in multivariate time series data.To further improve the robustness of the model,traditional autoregressive processing is followed.We constructed the first dataset in this field based on the monitoring data of fissure development-related factors in Cave No.32 of North Grotto Temple in Q City.Comparative experiments on this dataset show that the proposed model has a better performance in fracture development prediction of grotto rock mass.
作者
孙美君
郭红桐
王征
刘洋
张及鹏
张景科
李黎
SUN Meijun;GUO Hongtong;WANG Zheng;LIU Yang;ZHANG Jipeng;ZHANG Jingke;LI Li(College of Intelligence and Computing,Tianjin University,Tianjin 300350,China;College of Civil Engineering and Mechanics,Lanzhou University,Lanzhou 730000,China;Chinese Academy of Cultural Heritage,Beijing 100029,China)
出处
《吉林大学学报（信息科学版）》
CAS
2022年第6期994-1002,共9页
Journal of Jilin University（Information Science Edition）
基金
国家重点研发计划基金资助项目(2019YFC1520602)
国家自然科学基金资助项目(61876125,62076180)。
关键词
岩体裂隙发育
多元时序数据预测
自注意力机制
岩体裂隙监测数据集
rock fissure development
multivariate time series data forecast
self-attention mechanism
rock mass fissure dataset
