Addressing the challenges in carbonate rock lithology identification, such as similar well-logging curve response characteristics, scarcity of transitional lithology samples leading to classification difficulties, and the inability of traditional models to capture dependencies in vertical lithological sequences, we propose a hybrid machine learning framework that integrates feature enhancement, sample balancing, and sequence correction. The method first employs the Dung Beetle Optimizer (DBO) to optimize K-means clustering, thereby enhancing the feature discriminability for similar logging responses. Subsequently, the SMOTE oversampling technique is applied to specifically address the issue of sparse transitional lithology samples while preserving the original data distribution. On this basis, the Hidden Markov Model (HMM), is introduced to vertically correct the preliminary identification results using prior knowledge of stratigraphic sequences, effectively modeling the inter-lithology dependencies. Experimental results show that tree-based ensemble models driven by this framework significantly outperform traditional methods, with all evaluation metrics exceeding 95%. This study demonstrates that by jointly addressing the three major challenges of feature separability, sample balance, and sequence continuity, the proposed framework can significantly enhance the accuracy and geological consistency of carbonate rock lithology identification, providing a reliable solution for intelligent well logging interpretation in complex reservoirs.