小样本卷积神经网络井震映射反演

Mapping Inversion of Logging and Seismic Data Based on Convolutional Neural Networks with Small Samples

针对常规卷积神经网络(Convolutional Neural Networks,CNN)在井少地区因无法获得大量测井数据而易于出现过拟合现象的问题,提出了一种小样本CNN井震映射反演方法。通过网络结构优化设计,选出了最佳的网络层数、卷积核大小、特征图规模和激活函数,并将优选出的最好网络模型应用于实际资料反演。实际应用表明,小样本CNN井震映射反演方法可以防止过拟合、提高泛化能力和反演精度,为精细刻画薄互层油气藏的空间展布提供了一项智能化的新技术。

Conventional convolutional neural networks(CNN)are prone to over fitting because a large number of logging data cannot be obtained in the area with few wells.Therefore,a small sample CNN mapping inversion method based on well logging-seismic data is proposed.The optimal network layer number,convolution kernel size,feature map scale and activation function are obtained through the optimization of network structure,and the optimal network model is applied to the inversion of actual data.The application results show that the small sample CNN mapping inversion method based on logging-seismic data can prevent over fitting,improve the generalization ability and inversion accuracy,which provides a new intelligent technology for fine describing the spatial distribution of thin interbedded reservoirs.