摘要
岩爆是大型地下岩土和深部资源开采工程中面临的关键问题之一。为准确可靠地预测岩爆灾害,本研究提出一种基于Dropout与改进的Adam的深度神经网络(DNN)岩爆预测模型(DA-DNN)。根据岩爆的影响因素、特点及成因,选取硐壁围岩最大切向应力、岩石单轴抗压强度、岩石单轴抗拉强度和岩石弹性能量指数构成岩爆预测指标体系。在国内外岩爆研究成果的基础上,搜集289组岩爆工程实例数据,并以此作为岩爆预测的样本数据,然后采用深度学习技术建立基于DA-DNN岩爆预测模型。DA-DNN模型避开了指标权重确定问题,完全由数据驱动,减少了人为因素影响,可实现不完全、不精确并带有噪声的有限数据集中复杂且微妙的深层关系的学习。考虑到岩爆样本数据量有限,根据深度学习领域常用的数据集划分方法,即训练集、验证集、测试集按照6∶2∶2划分。先从样本数据中随机抽取58组作为测试集(预测样本),在模型最终训练完成后,评估其泛化能力,测试其真正的预测准确率。剩余的231组样本数据作为DA-DNN模型的学习样本,在训练过程中随机采样,抽取学习样本的80%作为训练集,20%作为验证集。当训练次数(epochs)取60时,学习样本为231组时,预测样本的预测准确率达到了98.3%。锦屏二级水电站、秦岭隧道和冬瓜山铜矿岩爆预测等3个工程应用实例来自于预测样本中,预测结果验证了DA-DNN岩爆预测模型的有效性与正确性。
Rockburst is one of the most critical problems in large-scale underground geotechnical engineering and deep mineral resource mining.In order to accurately and reliably predict rockburst disasters,a Dropout and improved Adam-based deep neural network(DA-DNN)rockburst prediction model is proposed.According to the affecting factors,characteristics,and genesis of the rockburst,this paper establishes a rockburst prediction index system composed of four evaluation indices,i.e.,tunnel-wall surrounding rock’s maximum tangential stress,rock uniaxial compressive strength,rock uniaxial tensile strength,and rock elastic energy index.Based on the literature on rockburst,this paper has collected 289 groups of rockburst engineering case data and adopted them as the sample data for rockburst prediction.Then,the rockburst prediction model based on DA-DNN is established by deep learning technology.The DA-DNN model avoids the problem of the determination of index weights,takes a completely data-driven approach,weakens the influence of human factors,and is capable of learning complex and subtle deeper relationships in incomplete,imprecise,and noisy finite data sets.Considering the finiteness of the rockburst sample data,this paper,based on the common data set segmentation method used in the deep learning field,segments these data into training set,validation set,and test set by the ratio of 6∶2∶2.It randomly draws 58 groups from the sample data of Formation 289 as the test set(prediction samples),which is used to evaluate the generalization ability of the model after the completion of the final training and test its real prediction accuracy.The remaining 231 groups of learning sample data serve as the learning samples of the DA-DNN model.During training,80%of the learning samples are randomly drawn as the training set,while the remaining 20%are used as the validation set.When the number of training epochs is set as 60 and there are 231 groups of learning samples,the prediction accuracy of prediction samples reaches 98.3%.The following three engineering application cases,i.e.,Jinping II Hydropower Station,Qinling Tunnel,and Dongguashan Copper Mine,come from the prediction samples.The prediction results validate the effectiveness and correctness of the DA-DNN rockburst prediction model.
作者
田睿
孟海东
陈世江
王创业
张飞
TIAN Rui;MENG Haidong;CHEN Shijiang;WANG Chuangye;ZHANG Fei(Institute of Mining Engineering,Inner Mongolia University of Science and Technology,Baotou 014010,China)
出处
《煤炭学报》
EI
CAS
CSCD
北大核心
2020年第S01期191-201,共11页
Journal of China Coal Society
基金
国家自然科学基金资助项目(51564038,51464036)
内蒙古自治区自然科学基金资助项目(2018MS05037)
作者简介
田睿(1988-),男,内蒙古四子王旗人,工程师,博士研究生。E-mail:tianrui6251@126.com;通讯作者:孟海东(1958-),男,内蒙古托克托县人,教授,博士生导师,博士。E-mail:haidongm@imust.cn
