CNN算法的损失函数优化及在低信噪比资料中的应用  

Optimization of loss function in CNN algorithm and its application in low signal-to-noise ratio data

作　　者：李辉 阎建国[1] 陈榆桂 孟辉 LI Hui;YAN Jianguo;CHEN Yugui;MENG Hui(College of Geophysics,Chengdu University of Technology,Chengdu 610059,China;Bureau of Geophysical Prospecting Inc.,China National Petroleum Corporation,Zhuozhou 072750,China)

机构地区：[1]成都理工大学地球物理学院,四川成都610059 [2]中国石油集团东方地球物理勘探有限责任公司,河北涿州072750

出　　处：《断块油气田》2023年第1期107-113,共7页Fault-Block Oil & Gas Field

摘　　要：深度学习CNN算法的核心之一是其利用损失函数完成反传机制达到各层网络之间的优化,因此,不同的损失函数及反传机制带来训练阶段人工神经网络模型不同的网络优化效果,其影响了机器学习算法的泛化能力及预测效果。基于此,文中提出了一种改进的带惩罚系数的损失函数,解决了在断层识别问题中因正负样本的数量高度不均衡导致的网络朝着错误方向收敛的问题。将其用于网络中指导训练,通过不同损失函数下的网络模型对理论数据和实际数据的识别结果,证明了这种方法的有效性和适用性。在低信噪比资料中的断层识别中,这种改进的优化网络,能够得到更稳定和更可靠的断层识别结果,为研究区潜山内幕小断层及断缝系统识别提供一种高效可靠的方法技术。One of the core of CNN algorithm based on deep learning is to use the loss function to complete the back propagation to achieve the optimization of each layer of network,so different loss function and back propagation mechanism bring different network optimization effect of artificial neural network model in the training stage,which affects the generalization ability and prediction effect of machine learning algorithm.Based on this,an improved loss function with penalty coefficient is proposed,which solves the problem that the network converges in the wrong direction due to the highly unbalanced number of positive and negative samples in fault identification.It is applied in network to guide the training,and its effectiveness and applicability are proved by theoretical data and practical data by network models under different loss functions.In the fault recognition of low signal-to-noise ratio data,this improved optimization network can obtain more stable and reliable fault recognition results,which provides an efficient and reliable method and technology for the recognition of small faults and fault fracture systems inside the buried hill in the study area.

关 键 词：Unet 反传机制 惩罚系数 断层识别 损失函数 低信噪比资料 

分 类 号：TE132.14[石油与天然气工程—油气勘探] P618.13[天文地球—矿床学]