<p>Lithological identification has become a core component in mineral resource exploration. In recent years, machine learning-based logging lithology identification has been widely applied. However, traditional machine learning models often exhibit low accuracy when processing logging data due to gradual stratigraphic transitions and sample imbalance issues. To address these issues, we propose a Multi-Scale Attention Sequence Network (MSABNet). This model captures receptive fields at different scales through multi-scale convolutional operations and integrates bidirectional information from upper formation data and lower formation data states using bidirectional gated recurrent units. Simultaneously, the introduction of a scaled dot-product self-attention mechanism and a focal loss function effectively enhances model performance. At the algorithmic level, this study reveals the adaptation mechanism of multi-scale features for stratigraphic gradual and abrupt changes, as well as the pattern of feature enhancement by attention mechanisms for imbalanced lithological samples. To validate the effectiveness of the proposed method, we conducted extensive experiments on the Songliao Basin dataset, achieving an accuracy rate of 88.27%. This study validates the effectiveness of deep sequence models in extracting nonlinear features from logging data, providing a generalizable research approach for lithological identification in complex geological areas. Experimental results demonstrate that this model exhibits outstanding recognition accuracy and stability, making it worthy of promotion and application in the field of lithological identification.</p>