Abstract
The use of machine learning in the mining industry has become increasingly prevalent in recent years, as these algorithms allow for analyzing substantial amounts of data. In the mine safety space, although many interesting studies have been conducted in recent years, applications related to machine learning in underground operations remain among the more challenging topics. In this study, we employ computer vision and machine learning technologies to identify risks in underground working areas, capturing data with low-cost hardware such as smartphones. Specifically, we focus on the problem of roof fall identification. We created a dataset of images captured underground and divided them into four distinct categories: “Hazard” (associated with images of potential roof fall), “Maintenance” (intervention is required to prevent future hazards), "No-Hazard" (appropriate ground control measures are in place and/or there are no hazards), and “Non-Usable Image” (alert user that hazard identification cannot be made). The resulting dataset contains 2,000 images so far. The images were processed as a classification problem with the aid of three machine learning models: support vector machines (SVMs), random forests (RF), and decision trees (DT). Each model performance was evaluated using a confusion matrix and metrics such as accuracy, recall, and F score (F-1). The results indicate that even though all 3 models are effective at identifying potential roof fall hazards, SVM outperforms the other 2, providing a promising approach for a mobile-device-based real-time hazard detection tool. Our investigation shows that the use of machine learning algorithms is a valuable tool for enhancing safety in the mining industry, helping mine operators to identify potential hazards before they occur and to gather valuable information for maintenance and ground control in a collaborative way.