Intelligent lithology identification while drilling using BiLSTM integrating attention mechanism

Objective Vibration signals while drilling reflect the physical properties and microstructural characteristics of rocks, providing a valuable basis for real-time lithology identification. However, existing models suffer from limited accuracy and robustness due to the low dimensionality of signal features and the class imbalance induced by variations in the rock layer thickness.

Methods To overcome these limitations, this study proposed a novel lithology identification method based on a bidirectional long short-term memory (BiLSTM) network that integrates attention mechanism (also referred to as the BiLSTM-Attention). This approach incorporates automated feature extraction and class weight adjustment strategies to enhance both the classification performance and generalizability of the model. First, triaxial vibration acceleration signals were acquired using a custom-designed measurement sub. Subsequently, a wide range of statistical, time-domain, and frequency-domain features were automatically extracted utilizing Tsfresh—a Python package. Key features were then selected through feature importance evaluation, followed by the construction of a high-quality feature space. Afterward, temporal dependencies inherent in the time-series data were modeled using the BiLSTM network. Meanwhile, attention mechanism was introduced to enhance critical features through weighting, thus improving the sensitivity to lithological variations of the resulting model. In addition, a class weight calculation mechanism was adopted to reduce the adverse impact of class imbalance, thereby enhancing the ability of the model to accurately identify a minority of lithologies.

Results and Conclusions  The experimental results indicate that for the dataset treated with feature engineering using Tsfresh, the proposed BiLSTM-Attention model yielded a training accuracy of 99.59% and a low loss value of 0.0125, which increased by 12.35% and decreased by 0.3915, respectively compared to raw data without feature engineering. The proposed model yielded precision ranging from 98.31% to 100.00%, recall from 98.99% to 100.00%, and F₁ scores from 99.15% to 99.73%, increasing by 4.23% to 6.89%, 4.46% to 5.13%, and 4.7% to 5.38%, respectively compared to those of the GRU, LSTM, BiLSTM, and LSTM-Attention models. Furthermore, the proposed method outperformed traditional models in terms of feature representation and classification accuracy. Notably, it exhibited enhanced robustness and stability in identifying a minority of lithologies. Overall, this study provides a novel approach for intelligent lithology identification while drilling.