Abstract: Traditional lithology identification methods are based mainly on drilling cores, curves, and rock images. These identification methods have high requirements on geological conditions and cannot meet real-time identification while drilling. For this reason, a new intelligent recognition method of lithology classification based on acoustic signals is proposed. This method first solves the problem of data sparseness through data enhancement technology based on the audio data obtained by the collision between the collected drill bit and the rock formation. Then, the deep learning model based on CGRU （CNN+GRU,convolutional neural networks+gated recurrent unit）  is used to conduct deep learning and training on the collected three types of rock audio data. Finally, to improve the model’s recognition ability in a complex background, the attention model is introduced for optimization. The attention model can realize key learning of lithology information in a complex environment and improve the recognition performance of the model. The experimental results show that, compared with GMM (Gaussian mixture model)-SVM (support vector machine), CNN and CGRU models, the accuracy of CGRU-AttGRU (attention mechanism model+GRU) hybrid model is about 81.17%, which is 13.31%, 9.99% and 5.93% higher than GMM-SVM, CNN and CGRU models respectively.

Key words: lithology recognition, feature extraction, data enhancement, CGRU-AttGRU model