Application of attention mechanism-based LSTM neural network in stratigraphy identification

• This study establishes a machine-rock coupling model grounded in contact mechanics and rock mechanics, formulating a mathematical framework that correlates drilling dynamics with rock properties. By adopting uniaxial saturated compressive strength as the classification criterion and extracting multi-scale features through wavelet analysis, the model significantly enhances lithology prediction accuracy. The coupling equations integrate Hertzian contact theory and Hoek-Brown failure criteria, enabling real-time mapping between drilling parameters (torque, RPM) and rock strength, achieving a 25% improvement over conventional methods. • To ensure high-fidelity data, an optimized acquisition scheme is designed for real-time sampling at 50 Hz, capturing 50+ critical parameters (e.g., thrust, vibration) per borehole. Each drilling operation generates approximately 10 million data rows, with rigorous field validation to maintain authenticity. Advanced data cleansing techniques, including outlier removal and non-drilling phase filtering, are applied to eliminate noise and retain valid segments, reducing the dataset to 2 million rows while preserving essential lithological signatures. This refined dataset provides a robust foundation for downstream analysis. • The study further addresses signal processing challenges by proposing a snake optimizer-enhanced TVF-EMD model, where permutation entropy guides the adaptive optimization of threshold and B-spline order parameters. This approach effectively decouples overlapping frequency bands in drilling signals, outperforming PSO and GA in convergence speed and decomposition precision. The optimized TVF-EMD isolates intrinsic mode functions (IMFs) with high spectral resolution, enabling accurate lithology-related feature extraction from complex vibrational and acoustic emissions. • For temporal feature learning, an improved LSTM-transformer hybrid model is introduced, leveraging multi-head masked attention to capture long-range dependencies in drilling time series. The attention mechanism prioritizes critical lithology-indicative segments (e.g., sudden torque fluctuations), while the transformer’s parallel computation enables millisecond-level inference. Combined with the cleansed dataset and decoupled IMFs, the model achieves over 90% prediction accuracy, surpassing standalone LSTM or CNN architectures in field trials. • The integrated system—spanning physics-driven modeling, adaptive signal processing, and deep learning—delivers a closed-loop solution for lithology recognition. Validated against field data from shale and limestone formations, it demonstrates superior generalizability and computational efficiency. Future work will extend the framework to heterogeneous rock layers and integrate reinforcement learning for autonomous drilling optimization, further bridging the gap between theoretical models and industrial deployment. This paper proposes a lithology recognition system based on a Long Short-Term Memory (LSTM) neural network, which incorporates multi-head masked attention mechanisms and the Transformer architecture. Initially, an equipment-formation coupling model is established based on contact mechanics and rock mechanics to determine the foundation for formation classification. Furthermore, a data collection strategy is proposed to ensure the real-time data collection at construction sites, thereby guaranteeing the authenticity and reliability of the data. To improve prediction accuracy, an advanced data cleaning method is also developed to eliminate invalid data. Additionally, the LSTM neural network combines multi-head attention mechanisms and the Transformer architecture, further enhancing the efficiency and accuracy of formation prediction. The Transformer architecture consists of an encoder and a decoder, where the encoder uses multiple Transformer encoder layers to learn time-series features based on time-window shifts, and leverages multi-head masked attention mechanisms to better capture key features in the input sequence. Subsequently, the decoder combines the time-series features extracted from each Transformer encoder layer using a one-dimensional Convolutional Neural Network (1D-CNN) and inputs them into the LSTM neural network for formation prediction. The validation results show that the proposed LSTM model achieves up to 98% accuracy in complex formation recognition tasks, outperforming other machine learning models such as Random Forest and Support Vector Machine (SVM). The model excels in identifying complex formation boundaries.