<p>Engineering geological conditions vary significantly across regions. Accurate and real-time stratigraphic identification is essential for efficient and safe construction of pile foundations, especially for intelligent control of the construction process. Rotary drilling has become a mainstream method for pile foundation projects. However, common rotary drilling rigs collect construction parameters at a relatively low frequency, and this brings challenges for stratigraphic identification based on these data. This paper proposes a DDE-LSTM-SWAM model to tackle this issue. The data-dynamic enhancement (DDE) method is used to establish correlations between low-frequency sampling data and stratigraphic sequence, thereby improving data utilization. A long short-term memory network (LSTM) is employed to capture the global variation of construction parameters, and it is further equipped with a sliding-window attention mechanism (SWAM) to assign adaptive local feature weights based on stratal continuity. A dataset comprising over 7000 data samples is established from the construction of 85 piles, and it is used to train and evaluate the proposed DDE-LSTM-SWAM model. The results indicate that the DDE method can enhance the prediction accuracy of machine learning models by approximately 20%, and an overall identification accuracy of 85.63% can be achieved by the proposed model. The proposed model offers a reliable approach for real-time stratigraphic identification during the rotary drilling process, enabling the stratum-adaptive optimization of construction parameters for rotary drilling operations.</p>