A Novel Automated Machine-Learning Model for Lithofacies Rec
Lithofacies represent stratigraphic structure information with high resolution, but the accuracy of lithofacies analysis is often largely affected by the complexity of geological environment. In recent years, machine learning has received increasing attention due to its feasibility in lithofacies recognition. Nevertheless, prevailing machine learning models such as CNN (Convolutional Neural Network) and DNN (Deep Neural Network) still show some disadvantages, such as: significant amounts of human interpreted data are required for data labeling, which means the accuracy level of model interpretation is still largely dominated by human factors and therefore often inefficient or unreliable. In this work, an automated hybrid framework is proposed to overcomes these shortcomings by combining K-means procedure and neural network. The framework consists of two aspects: 1) Images with the same category of lithofacies features are segmented from the resistivity images and they are fed into a network based on the K-means algorithm for clustering, which also uses other available logs to improve the accuracy of the clustering; 2) The clustered images are reconstructed into a matrix that represents extracted image features. This matrix along with the previously labeled dataset or training logs (if available) are fed into a customized CNN and MLP network, and the network assigns a determined lithofacies type to each cluster. Ultimately, automated lithofacies recognition is realized. The proposed model is applied to an oilfield in Jimusar area in Junggar basin, Xinjiang, China. The reservoir sections are heterogeneous and primarily composed of thin interbedded transitional lithologies. Rock types are complex, mainly including dolomitic siltstone, dolarenite and dolomicrite, etc. The lithofacies analysis of some oil wells in this area is carried out. The two-dimensional resistivity images and one-dimensional logs are input into the above model, and the model training results are compared with the core information and artificial analysis results, and good matching results are obtained.
Standard price: 10.00
Discounted price: 1.00
10.00
You could save: 90.0%
Author(s):
Jianhua Gao, Qiong Zhang, Yating Hu, Huilin Wu, Zaoyi Kang
Company(s):
University of Electronic Science and Technology of China
Year:
2022