An automatic pixel-level crack identification method for coals experiencing SHPB impact tests

Beijing Xie,Dihao Ai,Yu Yang

DOI: https://doi.org/10.1093/jge/gxz007

2019-04-01

Journal of Geophysics and Engineering

Abstract:Split Hopkinson Pressure Bar (SHPB) devices have been widely applied in investigating the dynamic mechanical properties and destruction processes of geomaterials. To precisely identify and quantitatively calculate the surface cracks of coals during SHPB impact loading, a framework, including dataset processing, model building and selection of evaluation indicators, has been presented in the study. Twelve groups of coal specimens were first processed by the SHPB system under different impact velocities. A high-frame-rate camera with 50,000 FPS was then used to capture the damage process of the specimen by employing a manual annotation method. A dataset of SHPB damage images, including a total of 350 original images and 350 corresponding cracked annotations, were obtained. Afterward, a deep convolution network model, namely CrackCoal, with 440,000 parameters was established based on a deep learning algorithm. This model is capable of automatically identifying the crack on the coal surface in each frame image. Finally, to accurately evaluate the merits of the surface crack identification algorithms on the coal mass surface, three indices, namely, precision, recall and F1-score, were proposed. Cracks can be quantitatively analyzed by the proposed model, and the propagation rate of cracks can also be calculated. The following results were obtained: (1) the CrackCoal model, established on the basis of deep learning, achieved 85.9, 72.1 and 77.1% accuracy in terms of Precision, Recall and F1-score, respectively; (2) three proposed evaluation indices fully reflected the performance of the algorithm in identifying cracks and helped promote the rapid development of the coal surface crack identification algorithm and (3) the quantitative analysis results showed that the growth trend of a crack after the initiation of cracking satisfied a linear relationship. The proposed automatic pixel-level crack identification method provided a new technique for describing the failure process of coals under high-speed impact damage.

geochemistry & geophysics