Research on Lithology Identification Method Based on Stacked Capsule Auto-Encoder Network

The capsule network is able to divide the data features into different capsules, which express their invariance and covariation by vectors. Aiming at the insufficiency of traditional machine learning methods for the ability in mapping logging parameter sequence structure and lithological feature diversity, a stacked capsule auto-encoder network (SCAE-Net) is proposed to improve the lithology identification in complex carbonate rocks. Firstly, multiple capsule auto-encoders are stacked and the capsules in the encoder can express the internal relationship of the object from part to whole. Secondly, SCAE-Net mines sample subspace by capsules, where sample representations are similar. Attention mechanisms are introduced into the model and renew the loss function to improve the convergence speed of the model. Finally, taking the carbonate reservoir in the Sulige gas field as an example, the accuracy of the SCAE-Net lithology identification model is up to 95.92%, which provides a new idea for complex carbonate lithology identification.