VGGNet网络结构下的地震数据增强方法研究  

Research on Seismic Data Augmentation Based on VGGNet architecture

作　　者：冯庆华[1] FENG Qinghua(Jiangsu Vocational Institute of Architectural Technology,Xuzhou 221000,Jiangsu,China)

机构地区：[1]江苏建筑职业技术学院,江苏徐州221000

出　　处：《矿产与地质》2022年第5期1089-1096,共8页Mineral Resources and Geology

基　　金：江苏省建设系统科技基金项目(2021ZD15)资助。

摘　　要：断层识别是油气勘探中至关重要的工作内容,针对目前断层识别效果不好的问题,基于VGGNet神经网络算法实现了断层的自动识别。对原始的VGGNet网络进行了改进,使其适用于本次的断层识别研究。通过添加缩放系数对学习率进行了优化,使得网络识别准确率有了提高。在此基础上为了解决断层自动识别神经网络算法中训练数据过少、而网络识别效果对标签数据要求较高的问题,文章尝试使用了Color Jitter和增加高斯噪声两种新的数据增强方法。使用不同的训练数据在改进的VGGNet网络结构上进行训练,通过对比其准确率和最终的断层识别效果,可以发现使用的两种新的数据增强方法训练的网络模型准确率更高、断层识别效果更好。从而证明了本次所引入的数据增强方法的有效性,为神经网络算法应用到断层识别研究中提供了一种有效的数据增强方法。Fault identification is an important work for the oil and gas exploration. In view of the poor effect of fault identification, this paper realizes the automatic fault identification based on Vggnet neural network algorithm. The original Vggnet network is improved to make it suitable for the fault recognition research. Then the learning rate is optimized by adding scaling coefficient, which improves the accuracy ratio of network recognition. On this basis, in order to solve the problem that there are too few training data in the neural network algorithm for automatic fault recognition and the network recognition effect has high requirements for label data, two new data enhancement methods, color jitter and adding Gaussian noise, are proposed in this paper. Then different training data are used to train on the improved Vggnet network structure. By comparing its accuracy ratio and the final fault recognition effect, it can be found that the network model trained by the two new data enhancement methods proposed in this paper has higher accuracy ratio and better fault recognition effect. The effectiveness of the proposed data enhancement method is proved. It provides an effective data enhancement method for the application of neural network algorithm to fault recognition.

关 键 词：VGGNet Color Jitter 地震数据 高斯噪声 缩放系数 断层识别 

分 类 号：P631.4[天文地球—地质矿产勘探]