Research on foreign object detection and sorting strategy in raw coal based on an improved YOLACT

Foreign objects mixed in raw coal at coal preparation plants not only compromise coal quality but can also damage hardware equipment during transportation. Efficient detection and removal of such impurities in the production process has long been a key management challenge. This study focuses on the manual sorting belt in the screening and crushing workshop and proposes a foreign object detection method and sorting strategy for raw coal based on an improved YOLACT model. The approach employs a single-stage YOLACT instance segmentation network as the baseline algorithm. To address issues such as missed detections and redundant detections of coal, gangue, and impurities on the conveyor belt, the model is enhanced through loss function optimization, improved non-maximum suppression, and attention mechanism integration. Additionally, a sorting strategy is developed based on the segmentation results to guide robotic manipulation. Experimental results demonstrate that the improved YOLACT network achieves a segmentation mAP of 87.8% at an IoU threshold of 0.5, with a processing speed exceeding 50 FPS, meeting the real-time demands of industrial applications. Furthermore, using the foreign object centroids extracted from the segmentation output, the proposed “sliding window grasping” and “rotational grasping” strategies significantly improve the success rate of robotic foreign object removal. These findings provide effective theoretical and technical support for the implementation of robotic sorting systems in coal preparation plants.