基于一维卷积神经网络的岩石物理相识别

Identification of Petrophysical Facies Based on One-Dimensional Convolutional Neural Networks

为解决岩石物理相识别问题,提出了一种基于可解释一维卷积神经网络的识别方法。该方法通过引入全局平均池化层,突出了测井曲线波形的动态变化部分;并且通过分类激活映射增强了方法的可解释性;通过引入扩张卷积和批量归一化,弥补了由全局平均池化层引起的性能下降。实验结果表明,测试集中4种岩石物理相的平均F1分数为0.97,相比其他同类方法提升了0.15左右。研究表明,该方法可用于识别岩石物理相,并可提高分类过程中的可解释性,从而为预测优质致密砂岩储层提供了一种新的深度学习方法。

Aiming at the problem of rock physical facies identification, a identification method based on the interpretable one-dimensional convolutional neural network is proposed. By introducing the global average pooling layer, the dynamic variation part of the logging waveform is highlighted. The interpretability of the method is enhanced by the classification activation mapping. By introducing dilated convolution and batch normalization, the performance degradation caused by the global average pooling layer is compensated. According to the experimental results, the average F1 score of the four petrophysical facies in the test set is 0.97, which is about 0.15 higher than that of other similar methods. The results show that the proposed method can be used to identify petrophysical facies and improve the interpretability in the classification process, providing a new deep learning method for the prediction of high quality tight sandstone reservoirs.