Lithology identification plays a crucial role in oil and gas exploration.Although drilling data can provides costeffective and realtime lithology characterization,its effectiveness is often constrained by weak lithological signals,strong feature coupling,and complex nonlinear relationships.To overcome these limitations,this study presents a hierarchical stacked autoencoderbased method for real-time lithology identification while drilling.The proposed method constructs a hierarchical feature learning framework composed of raw drilling parameters,shorttime energy intermediate features,and a maximal information coefficientdriven regularization mechanism.This framework enhances the representation of weak lithological information in drilling data and guides the model to preferentially learn features with stronger correlations to lithology.Experimental validation on four drilling datasets demonstrates that the proposed method outperforms seven benchmark methods overall.The proposed model achieved 95.1%accuracy with an F1 score of 93.6%on Well A and 92.1%accuracy with an F1 score of 86.0%on Well B.Moreover,the model exhibited good generalization capability in crosswell prediction,yielding accuracies of 94.2%for Well C and 95.4%for Well D,offering a robust solution for realtime lithology identification.
