Multi-stream encoder and multi-layer comparative learning network for fluid classification based on logging data via wavelet threshold denoising.

The article focuses on the development and evaluation of LogMEC-MCL, a novel fluid classification framework for oil and gas exploration that integrates wavelet threshold denoising, a multi-stream encoder, and multi-layer contrastive learning. Designed to address challenges posed by noisy, high-dimensional logging data, LogMEC-MCL preprocesses data to reduce noise, extracts multi-scale spatial and temporal features via attention gated recurrent units and convolutional neural networks, and enhances feature discrimination through instance-level and temporal contrastive learning. Tested on real-world logging datasets from the Tarim Basin in China, the model outperformed existing deep learning approaches—including GRU, BiGRU, Transformer, and hybrid models—achieving classification accuracies above 95% and demonstrating superior precision, recall, and F1 scores in distinguishing rock, water, and oil layers. The study also employs SHAP analysis to interpret feature contributions, highlighting the importance of gamma ray, density, photoelectric effect, acoustic, and neutron logging parameters in fluid identification. The authors suggest future work to improve computational efficiency and real-time applicability, as well as potential extensions to other complex high-dimensional data domains.