探地雷达钢筋检测数据深度学习反演

Deep learning of ground penetrating radar data inversion in reinforcement detection domain

实际工程中探地雷达采集的数据量大,目前对其数据解译仍然以人工为主,成本高、效率低,且解译结果受解译员的专业知识及经验等主观因素影响.因此,为快速、准确且智能化对探地雷达实测数据进行处理,文中提出一种探地雷达钢筋检测数据深度学习反演方法.文中首先运用时域有限差分方法对不同尺寸、不同埋深的钢筋模型进行正演计算,构建样本数据集;其次搭建一种端到端的深度学习网络架构,其输入端为归一化后的探地雷达数据,输出端为介电常数模型;然后针对网络进行监督学习;最后,利用已训练好的网络对未知探地雷达数据进行反演.模型试验表明,反演结果与模型中钢筋位置及尺寸均高度吻合,模型一、模型二和模型三的介电常数反演值精度分别为99.28%、99.23%和98.84%.对于含噪声数据的反演,仍然具有较好的效果.实测探地雷达数据的反演也验证了文中方法的有效性.

In actual engineering,the amount of data collected by ground penetrating radar is large.At present,the data interpretation is still mainly manual,with high cost and low efficiency.The interpretation results are affected by subjective factors such as the professional knowledge and experience of the interpreter.Therefore,in order to process the measured data of ground penetrating radar quickly,accurately and intelligently,this paper proposed a deep learning inversion method of ground penetrating radar steel bar detection data.In this paper,the time-domain finite difference method is used to carry out forward calculation on the steel bar models of different sizes and different burial depths,and a sample data set is constructed;Secondly,an end-to-end deep learning network architecture is built,the input end is the normalized ground penetrating radar data,and the output end is the dielectric constant model;Then perform supervised learning for the network;Finally,use the trained network to invert the unknown GPR data.The model test shows that the inversion results are highly consistent with the position and size of the steel bars in the model.The inversion accuracy of the dielectric constant values of model 1,model 2 and model 3 are 99.28%,99.23%and 98.84%,respectively.For the inversion of noisy data,it still has a good effect.The inversion of the measured ground penetrating radar data also verifies the effectiveness of the method in this paper.