Reducing the Effect of Incorrect Lithology Labels on the Training of Deep Neural Networks for Lithology Identification

The identification of lithology is a crucial step in determining the characteristics of petroleum reservoirs, and many studies have investigated the application of deep neural networks in lithology identification. However, incorrect lithology labels as a result of manual interpretation can seriously affect network training when deep learning is used to identify lithology from well logging data. To address this problem, a method of learning with noisy labels (probabilistic end-to-end noise correction in labels, PENCIL) is applied to the network training process. Experiments are conducted on two real well logging datasets, and two types of label noise, random and pattern, are added to the lithology labels of the training data to simulate the lithology label noise that may exist in the actual data. To demonstrate the effect of this method, this study trains four network models, namely residual network (ResNet), bidirectional gated recurrent unit (Bi-GRU), ResNet-PENCIL, and Bi-GRU-PENCIL. The results of the experiments show that pattern label noise has a more serious effect on network training than random label noise, and network models that use the PENCIL framework effectively mitigate the effect of incorrect lithology labels on lithology identification results.