Rock blasting crack network recognition based on faster RCNN-ZOA-DELM model

Abstract

Identifying cracks in rock blasting provides an accurate representation of the crack network that occurs during the blasting process. It serves as a crucial tool for the precise evaluation of the dynamic response characteristics of rocks. However, most crack characterizations rely on manual measurements, which are often inaccurate, prone to significant errors, and are both time-consuming and costly. Therefore, this study compiled a database of 1,000 images of rock blasting fractures. The images were divided into foreground and background images by Faster RCNN. Five parameters were selected as the input variables, with the optimal image threshold set as the prediction target. A deep extreme learning machine (DELM) was optimized using swarm intelligence algorithms to develop eight hybrid models. The performances of these prediction models were comprehensively evaluated using four metrics. The results indicate that the proposed DELM-based hybrid model can consistently provide accurate predictions of the optimal image threshold. The DELM model using the zebra optimization algorithm performed best, with a root mean square error (RMSE) of 0.027 and a mean absolute percentage error (MAPE) of 4.58%. Finally, the proposed calculation method could quickly and accurately extract crack characteristics, including the crack network area, crack length, crack twist angle, and maximum crack width. The research results of this study could provide an effective way to identify the crack network characteristics.