Research On Lithology Identification Using Well Logs Based On The Semi-supervised Ladder Network

Lithology identification is an important basis for reservoir characterization and modeling,and the core description is the most accurate method to identify lithology.Due to the high cost of obtaining cores,the number of well log samples for core calibration labels is small,and there are a large number of unlabeled well log samples.Commonly,the supervised learning methods use a small number of labeled samples when establishing a lithology prediction model.They ignore a large number of unlabeled samples,so that the model has limited accuracy.To handle this problem,in this thesis we introduce the semi-supervised ladder network algorithm for lithology identification using well logs.This algorithm integrates supervised learning and unsupervised learning by adding lateral connections.Also,it uses both the labeled samples and unlabeled samples to train the model.In order to further strengthen the feature propagation ability of the model and improve the lithology identification accuracy,we add oblique lateral connections to improve the network structure on the basis of the ladder network.We establish the improved ladder network model for lithology identification using well logs.Through the experimental analysis of the improved ladder network optimizer and key parameters,such as the number of network layers,initial learning rate and noise variance,the optimal parameter range of the model is obtained.For the problem of lithology identification using well logs,we compare the model recognition effect of the improved ladder network,the original ladder network and four supervised learning.It is concluded that the ladder network has an average accuracy improvement of about 7.1%over the supervised model.In the improved ladder network,the model of the oblique lateral connection averaging method has the best effect,which is about 3.2%higher than the original ladder network.The F₁-scores on each lithology are also improved compared to the other models.Experiments show that the improved semi-supervised ladder network method can effectively identify lithology using well logs.