MSE Graduate Seminar: The Application of Machine Learning to Multi-Sensor Based Ore Sorting

Supervisor: Prof. Erin Bobicki

ABSTRACT: The mineral processing industry is currently facing major challenges such as high energy consumption in comminution and the exponential growth of produced tailings (waste) caused by decreasing ore grades and increasing global demand for metals. One technique to mitigate these issues is ore sorting, which is the act of separating valuable and invaluable (gangue) material before comminution and downstream processing. This reduces the amount of material being processed, directly decreasing tailings generation. Current ore sorters use a single sensor and algorithms to detect whether a rock is valuable or not. The drawbacks are that they are dependent on one sensor and algorithms are based on a fixed set of rocks. Considering that the characteristics of the deposit can change over time, this can make a pre-set algorithm unreliable. This makes the traditional method for creating ore sorting algorithms limited in their development, accuracy, and implementation. One solution to this problem is the use of machine learning (ML). ML is not restricted by the number of input variables, it is flexible to changing conditions, and can find previously undetected patterns and distinctions in the ore. The objective of this project is to apply ML to generate more accurate sorting decisions using a multi-sensor algorithm. The algorithm is composed of two parts: (i) object detection and segmentation and (ii) rock analysis and sorting decision. Two ores were studied: a Ni-Cu ore and a Zn-Pb ore. The two sensing methods being used are x-ray transmission and microwave infrared.