摘要
针对隧道地质超前预报过程中,探地雷达(ground penetrating radar,GPR)线测图解释仅依靠专家经验,且存在准确率不高的问题,通过对GPR成像原理和隧道地质特性的研究,以及对深度置信网络(deep belief networks,DBN)计算复杂度的分析,提出一种改进的基于压缩感知和DBN的GPR线测图分类解释模型。该模型首先利用压缩感知技术对原始GPR线测图进行压缩处理,通过选择图像压缩比得到合理的压缩图像;然后将压缩后的图像送入DBN模型进行分类,根据分类结果对原始GPR线测图进行解释;最后利用广西六宜(六寨—宜州)高速公路隧道实测数据对模型的有效性进行验证,试验数据共20 000幅GPR图像,包括6种隧道地质类型,其中15 000幅图像作为训练样本集,5 000幅图像作为测试样本集。研究结果表明:当GPR线测图压缩比为256,反向微调数据为1 000幅图像,DBN模型迭代次数为30时,模型对测试数据中6类探地雷达线测图的分类准确率达100%,单次训练时间降低为原DBN模型的8%左右;大量仿真试验发现GPR线测图的合理图像压缩比区间为64~1 024,在此区间压缩的图像能最大限度地降低图像维度并且保留原始图像信息。该模型具有解释准确率高、训练速度快等优点,可为制定隧道施工和开挖计划提供合理依据。
In the process of tunnel geological prediction, Interpretations of GPR （ground penetrating radar） line images are mainly determined by experience of experts, so that the accuracy is low. Through in-depth study on the principle of GPR, the geological characteristics of tunnel as well as the analysis on computational complexity of DBN （deep belief networks）, an interpretation model of GPR line images was proposed based on compressed sensing and DBN. Firstly, the raw GPR line images were compressed by compressed sensing to obtain reasonableimage with compression ratio. Secondly, the compressed images were sent to DBN model for classification. Then, the raw GPR line images were interpreted according to the classification results. Finally, the measured data of Liuyi （Liuzhai to Yizhou） Highway tunnels in Guangxi were used to verify the effectiveness of the proposed model. The experimental data contained 20 000 images which could be classified into 6 types of tunnel geology. 15 000 images of the experimental data were used as training dataset, and the other 5 000 images were used as testing dataset. The results show that when the compression ratio of GPR line images is 256, the reserve fine turning data is 1 000 images and the iteration number of DBN model is 30, the classification accuracy of the proposed model on the 6 types in testing dataset is 100%, and the single training time reduces to about 8% of that of the original DBN model. According to a large number of simulation experiments, the reasonable range of image compression ratio is 64 to 1 024, among which, the size of image can be greatly reduced and the information of raw image can be effectively preserved. Thus, it can be seen that the proposed model has the advantages of high accuracy of interpretation and fast training speed, and can provide reasonable basis for construction and excavation plan of tunnel. 3 tabs, 5 figs, 25 refs.
出处
《长安大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2017年第3期90-96,共7页
Journal of Chang’an University(Natural Science Edition)
基金
国家自然科学基金项目(61271143)
作者简介
李宝奇（1985-），男，天津宝坻人，工学博士研究生，E-mail：bqli@mail．nwpu．edu．cn。通讯作者：贺昱曜（1956-），男，陕西富平人，教授，博士研究生导师，E-mail：heyyao@nwpu．edu．cn。
