Zhu, Liping (China University of Petroleum-Beijing) | Li, Hongqi (China University of Petroleum-Beijing) | Yang, Zhongguo (China University of Petroleum-Beijing) | Li, Chengyang (China University of Petroleum-Beijing) | Ao, Yile (China University of Petroleum-Beijing)
Lithology interpretation is a key component of well-log interpretation, which can be viewed as a supervised classification problem from the perspective of machine learning. Recently, various machine-learning algorithms have been applied for borehole lithology interpretation as an alternative way. Convolution neural network (CNN) is a class of deep, feed-forward artificial neural networks, which has been applied to visual imagery analysis successfully. As one of the most popular and effective deep learning structures, CNN has been widely applied in computer vision problems, image target detection, recommendation systems, and natural language processing. However, the machine-application of CNN on logging-curve-based reservoir evaluation bas not received nearly as much attention as expected. The main reason is that CNN is designed to accept input in the form of multilayer images, meanwhile, it's hard to construct the nonsequential curve values into image-style input. In this article, we propose a wavelet-decomposition-based method to construct multilayer image-style input for each logging point, which makes it possible to convert the problem of logging lithological interpretation into a supervised image-recognition task. The proposed method is applied to the wells of the Daqing Oilfield and achieves excellent application effect.