Zhiming Zhao - Department of Petroleum Engineering, Panjin Vocational and Technical College, Panjin, Liaoning, China
Wenyang Gao - Department of Petroleum Engineering, Panjin Vocational and Technical College, Panjin, Liaoning, China
Jiabiao Chang - CCDC Changqing General Drilling Company, Xi’an, Shaanxi, China
Yiming Chen - Department of Oil and Gas Engineering, Liaoning Petrochemical University, Fushun, Liaoning, China\ Faculty of Engineering and Applied Science, University of Regina, Regina, Canada
Qiushi Zhang - Department of Oil and Gas Engineering, Liaoning Petrochemical University, Fushun, Liaoning, China
Bo Wang - Faculty of Engineering and Applied Science, University of Regina, Regina, Canada

Traditional mud logging cuttings identification relies on professionals to carry out visual identification and analysis based on experience. The workload is large and subject to the influence of subjectivity, which is likely to cause errors in information extraction and result analysis. Based on applying deep learning theory in image processing technology, ResNet, DenseNet, and SqeezeNet deep neural network models were built according to the classification of cuttings images. The deep neural network models were used to identify the pictures of cuttings subdivision classification. The evaluation indexes, such as stability, robustness, and recognition effect of different models, were compared and analyzed, and the three models were selected according to the best. The results showed that under the Top-۲ standard, the deep neural network model was more accurate in recognizing composite cuttings images. In contrast, the SqeezeNet ۱_۰ model had the best performance in identifying cuttings after synthesizing different evaluation indicators. The final recognition rate of the optimized SqeezeNet ۱_۰ model reaches ۹۹.۴۸%. In addition, the obtained SqeezeNet ۱_۰ network model can effectively identify sandstone, mudstone, and conglomerate cuttings on-site and can be extended to the daily identification of composite cuttings.

Deep Learning, deep neural network models, image recognition, cuttings images, mud logging