Metso, a provider of equipment and solutions to the mining, mineral and metal industry, has integrated automated digital solutions into its products and seeks further innovation. This thesis explores development of a foreign object detection system using Metso’s 3D scanning-based rock imaging product and a deep learning model to identify unwanted objects (e.g., wood, metals, plastics) among crushed ores on a conveyor belt before milling.
The primary objective was to develop a continuous online monitoring system with three key goals: fast model execution, high accuracy with minimal false alarms, and the ability to identify unseen object types. The study focused on deep learning viability through transfer learning, comparing image classification and object detection for optimal performance and execution speed.
A data set of 214 images was created using a custom rig simulating a moving conveyor belt with ore load. Three image formats were obtained from the camera attached to the rig: 1) a depth/range map, 2) an intensity image, and 3) a multichannel combined image from these two formats. Image classification models were trained on multichannel images, while object detection models used reflectance images. The models were evaluated on several performance metrics and inference speed. A novel metric was introduced to compare image classification and object detection models by converting object detection outputs into classification outputs.
Comparative analysis of pre-trained models showed that image classification models outperformed object detection models, especially in detecting unseen object types. The top-performing image classification models also had faster inference speeds. Among the evaluated models, the YOLOv8s-cls model was the most effective, delivering superior performance across multiple metrics with acceptable inference speed.
The research validates the viability of deep learning for foreign object detection in industrial settings, laying the groundwork for future implementation, although real-world site evaluations are needed for comprehensive validation.
