变窗重建算法和自适应SegNet网络在地层层序划分中的应用  

Variable window reconstruction algorithm and adaptive SegNet network in the application of stratigraphic sequence classification

作　　者：邵广辉 高衍武 张苏利 肖华 杨帆 于龙 姚军朋 陈立东 王绍祥 孟屹明 SHAO Guanghui;GAO Yanwu;ZHANG Suli;XIAO Hua;YANG Fan;YU Long;YAO Junpeng;CHEN Lidong;WANG Shaoxiang;MENG Yiming(Geological Research Institute,China Petroleum Logging Co.Ltd.,Xi’an 710000,Shaanxi;Changqing Branch,China Petroleum Logging Co.Ltd.,Xi’an 710000,Shaanxi;Daqing Branch,China Petroleum Logging Co.Ltd.,Daqing 163000,Heilongjiang;Market production department.,China Petroleum Logging Co.Ltd.,Xi’an 710000,Shaanxi;Shandong Institute of Geophysical&Geochemical Exploration,Shandong Provincial Bureau of Geology&Mineral Resources,Jinan 250013,Shandong)

机构地区：[1]中国石油集团测井有限公司地质研究院,陕西西安710000 [2]中国石油集团测井有限公司长庆分公司,陕西西安710000 [3]中国石油集团测井有限公司大庆分公司,黑龙江大庆163000 [4]中国石油集团测井有限公司市场生产处,陕西西安710000 [5]山东省地质矿产勘查开发局,山东省物化探勘查院,山东济南250013

出　　处：《长江大学学报（自然科学版）》2024年第4期19-31,共13页Journal of Yangtze University(Natural Science Edition)

基　　金：中国石油天然气股份有限公司重大科技专项“陆相中高成熟度页岩油勘探开发关键技术研究与应用”(2019E-2603)。

摘　　要：深度学习技术以批量处理数据、解释时间短、解释精度高等特点,为地层自动划分提供了新的方向。然而由于测井数据维度高、样本数量有限、相邻样本间特征相似等原因,深度学习存在着样本独立性与可靠性等问题。面对复杂的地下结构和不整合面,特别是在样本质量差、样本数量少的情况下,常规深度学习方法很难准确划分地层边界。考虑到测井数据属于小样本数据、数量有限且质量较差,不利于模型的训练和构建,因此拟采用可变窗口波形重建算法增加训练数据量,根据原始波形的特征生成重建波形,模拟不同速度下模型的波形特征,对部分原始测井数据进行人工分层重建,将重建后的数据作为训练样本输入到自适应可变卷积核尺寸的SegNet网络中,使用训练好的SegNet来解决复杂的地下结构问题。实验结果表明,采用自适应可变卷积核尺寸的SegNet网络可以在多个尺度上拟合地震数据中的断层和不整合面,达到更好的分割效果,且具有良好的识别效率和较强的鲁棒性。The emergence of deep learning provides a direction for automatic stratigraphic division,which is characterized by a large amount of batch interpretation data,a short interpretation cycle and high interpretation accuracy.Deep learning requires a large amount of training data,but due to the high dimensionality of logging data,limited sample size,and similar features between adjacent samples,there are issues with sample independence and reliability.In the face of complex underground structures and unconformity surfaces,it is difficult for the general deep learning method to accurately divide the stratum boundary,especially when dealing with poor sample quality and a small number of samples.Considering that the logging data belongs to small sample data with limited quantity and poor quality,which is not conducive to model training and construction,we built a variable window waveform reconstruction algorithm to increase the volume of training data.This algorithm generates reconstructed waveforms based on the characteristics of the original waveforms and simulates the waveform characteristics under different speed models.Manually stratify and reconstruct some of the original logging data,and input the reconstructed data as training samples into a SegNet network with adaptive variable convolution kernel size.Use SegNet with adaptive variable convolution kernel size to solve complex underground structure problems.Through experiments,SegNet with adaptive variable convolution kernel size can fit faults and unconformity surfaces in seismic data on multiple scales to achieve a better segmentation effect.After verification,we found that the model has good recognition efficiency and robustness.

关 键 词：深度学习 自动地层对比 可变窗口波形重建算法 自适应可变卷积核尺寸网络 

分 类 号：TE132.14[石油与天然气工程—油气勘探]