The Fifth International Symposium on Mine Safety Science and Engineering (ISMSSE2021)

57 / 2021-09-08 06:30:01

Using digital images to monitor the concentration of mine dust—a preliminary investigation

digital images; mine dust; monitoring; machine learning algorithm

Mining operations release large amounts of dust particles into the air. The produced dust in the form of suspended particulate matter can cause serious safety and health problems such as respiratory diseases result from respirable dust and safety concerns caused by the reduced visibility. Therefore, it is of great significance to accurately monitor the real-time dust emission to help dust management at mine sites. However, the monitoring of dust emission can be difficult on some occasions using conventional monitoring instruments. Research is needed to develop new monitoring techniques for these occasions.



Thus, this paper aims to conduct a preliminary study for the development of a dust monitoring technique using digital images. To achieve this goal, a machine learning algorithm was adopted to correlate the dust concentrations and image features. First, a batch of photos was accessed from the weather station at the University of Alberta, and the corresponding values of PM2.5 were collected from the Edmonton Central weather station. Then, 14 image features were extracted through Python Coding. In total, 203 data points were collected for modelling. After that, a machine learning algorithm—support vector regression (SVR)— was proposed to correlate the dust concentration with the image features. The SVR model was optimized with the Bayesian algorithm and validated using the k-fold (k=5) cross-validation technique. To evaluate the performance of the proposed SVR model, an artificial neural network (ANN) was constructed for comparison. The accuracy of models was assessed by two indicators: the coefficient of determination R² and the mean absolute error (MAE). The results suggested that the proposed SVR model showed a better performance than that of ANN, and it can predict the dust concentration in high accuracy with R² of larger than 90% and MAE of smaller than 10 ppm. This proposed approach provides a potential method to monitor dust concentration remotely for locations where conventional monitoring devices are difficult to be installed.



Despite the high accuracy of the proposed approach, the current study has limitations that need further investigation. Future studies are needed to validate the model using data from mine sites.