DRSN-GAF: Deep Residual Shrinkage Network (DRSN) for Lithology Classification Through Well Logging Data Transformed by Gram Angle Field
Youzhuang Sun, Shanchen Pang, Junhua Zhang, Yong-An Zhang
Abstract
Lithology holds significant importance in reservoir evaluation and geological modeling. However, the complex relationship between logging and lithology leads to strong multisolutions in logging responses, resulting in inaccurate identification of traditional logging lithology methods. Given the impressive performance of deep learning in data classification, we delved further into the technology and presented a deep residual network for lithological classification. The deep residual shrinkage network (DRSN) model incorporates an attention mechanism and a soft thresholding strategy based on the residual network. The residual shrinkage mechanism is the core characteristic of DRSN. It enhances model sparsity by shrinking the weights in the residual block (soft threshold strategy), resulting in a simpler model, and mitigating the overfitting problem. To test the model, we selected data from two wells in the Tarim Oilfield, China. In this article, we innovatively use the Gram angle field (GAF) to convert 1-D logging parameters into 2-D images. These images are then input into the DRSN model, using the idea of image processing to tackle the lithology classification problem. GAF effectively captures time-series information and converts 1-D time-series data into a 2-D matrix representation. Furthermore, GAF enhances the ability to capture nonlinear structures and patterns in time-series data using trigonometric transformations. Experimental results demonstrate that the DRSN-GAF network outperforms both DRSN and convolutional neural network (CNN) networks in terms of accuracy. The lithology prediction tasks for the two wells achieve accuracies of 96.00% and 96.50%, respectively.