Abstract:

The recognition and classification of lithological information hold significant importance for categorizing oil and gas reservoirs and evaluating the compressibility of reservoir rocks. This study presents enhancements to the deep learning  network U-Net and conducts a comparative validation using experimental data from  central depression  of  Songliao Basin. We propose a more suitable feature attention fusion Unet (FAF-Unet) designed for well logging data. The selection of logging data primarily involves sensitivity analysis to identify characteristic parameters, including natural potential, acoustic time difference, photoelectric absorption cross-section index, wellbore diameter, density, natural gamma, and deep and shallow lateral resistivity. These parameters are analyzed to understand reservoir rock lithology. FAF-Unet is a network that amalgamates residual blocks and channel attention mechanisms. Residual blocks can better retain the data with lower-level features of the depth direction, and channel attention mechanisms can make up for the problem of ignoring the connection between  horizontal channels during vertical convolution. Comparing the accuracy and recall of six recognition methods, including support vector machine, decision tree, U-Net, U-Net with effective channel attention (ECA) mechanism, U-Net with residual block (Res-Unet), and FAF-Unet with both ECA and residual block, experimental results demonstrate that FAF-Unet achieves an accuracy and recall rate exceeding 89.00%. FAF-Unet outperforms the other five methods in terms of recognition performance and exhibits a narrower fluctuation range between accuracy and recall.

Key words: lithology lecognition')">

lithology lecognition, deep learning, residual blocks, channel attention mechanism, U-Net