Machine learning segmentation and volume analysis of 3D-point clouds

Sammanfattning

The first step of processing ore is to break the ore into smaller fragments, typically using a mill or a crusher. For some mills, steel balls are used as the grinding media to break the ore. Over time, steel balls are worn due to various factors, affecting their size. This leads to a reduced milling efficiency of the ore. Therefore, it is relevant to know the size distribution of steel balls and ore. In this thesis, we investigate whether machine-learning-based segmentation of point clouds can be used to identify and separate individual steel balls and rocks (ore) in mill rock beds, regardless of rock bed composition. Further, it examines whether the model predictions can be used to estimate the sizes and spatial distributions in various rock beds. Results show that it is possible to identify and separate steel balls and rocks in point clouds of rock beds, where only samples from data annotation are present. It is also possible to use the labels from the model predictions to estimate the sizes of steel balls and rocks. However, test results show an accuracy of 54.2%, an average precision of 40.1%, and an average recall of 52.6%. This means the resulting size distributions from model predictions are unreliable. Performing size estimation directly on annotated point cloud data could replace manual measurements in mills, possibly reducing maintenance downtime. Further work could improve segmentation results, thereby providing more reliable size distributions.