Deep learning-based approach for reservoir fluid identification in low-porosity, low-permeability reservoirs.

This article focuses on a novel reservoir fluid identification method designed for low-porosity, low-permeability reservoirs, combining a symmetric synthetic minority over-sampling technique (Sym-SMOTE) with a hybrid model integrating revisiting mobile convolutional neural network from vision transformer perspective (RepViT) and transformer architectures. The approach processes conventional well logging data using a sliding window to preserve stratigraphic continuity and transforms sequence samples into multi-channel feature maps for improved local and global feature extraction. Tested on logging data from the Dagang Oilfield, the Sym-SMOTE+RepViT-transformer method achieved superior classification accuracy (94.41%) compared to traditional models, particularly enhancing minority class (oil-poor layer) recognition. The study highlights the method's potential applicability to other well-logging tasks while acknowledging challenges related to data quality and parameter tuning in real-world scenarios.