Multi-scale spatiotemporal feature lithology identification method based on split-frequency weighted reconstruction

Abstract

The accurate identification of lithology in thin and deep layers is a crucial task in logging. However, traditional logging lithology identification methods are often inefficient for thin and deep layers and sometimes require human intervention, which greatly reduces the efficiency of oil and gas production. Therefore, this study proposes a multi-scale spatiotemporal feature lithology identification method under split-frequency weighted reconstruction. First, split-frequency weighted reconstruction of the logging curves is performed by Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to improve their thin layer resolution. Subsequently, a feature fusion model of multi-scale convolutional neural networks and bidirectional gated recurrent neural networks (MCNN-BIGRU) is constructed to learn the spatiotemporal features of the logging curves. Finally, during feature propagation, the attention mechanism assigns weights to historical features to mitigate the accumulation of error information, thereby improving the lithology recognition effect. To verify the model's performance, we constructed a lithology dataset with five wells and conducted an experiment to show that the reconstructed logging curves have significantly higher vertical resolution than the original curves. Furthermore, the MCNN-BIGRU-AT model developed in this study exhibited a better lithology identification effect than the single-structure model, with a highest accuracy of 96.69%. In summary, the proposed method is a novel and efficient method for lithology identification.