Abstract
The size distribution of coke is important in order to decide the burden layer structure and the burden porosity in the shaft of a blast furnace (BF), which fluctuates daily and can be determined by several parameters. It is measured two or three times per shift by screening the raw material. However, the screening method used is random and takes a lot of time and manpower, resulting in the susceptive size distribution of the raw material and delayed operation of the BF. Therefore, in this paper, a new online approach used to measure the size distribution of particles was selected through comparison. Four common algorithms were used to detect the coke particles from images, including the Marker-based Watershed (MW), Histogram of Oriented Gradient + Support Vector Machine (HOG + SVM), Faster Region-based Convolutional Neural Networks (Faster R-CNN), and You Only Look Once (YOLOv3). The results show that the MW and HOG + SVM were not suitable for coke image detection. The average mean average precisions (mAPs) of Faster R-CNN and YOLOv3 were 93.391% and 91.348%, respectively, which meet the requirements of coke particle recognition. However, the YOLOv3 (5.419 fps) was selected as the final coke particle image detection algorithm, which is about 4.3 times faster than the average detection speed of Faster RCNN (1.269 fps). After this, the YOLOv3 and screening were used to detect 100 coke images and to generate particle size distribution statistics. The results show that the two methods are basically consistent. YOLOv3 can be used in the online measurement of BF coke. This research, which is of important value, provides a basis for the online measurement of the particle size distribution of raw material in a BF.