NING Houyin, LU Weiwen, YU Zhengwei, CHEN Liangjun, WANG Guangying, LONG Hongming
Coke powder, as the primary fuel in sintering production, has a particle size distribution that significantly affects both product quality indicators and energy consumption levels. The current method for measuring coke powder particle size typically involves manual sampling, drying, and vibrating sieving, which is complex, time-consuming, and unsuitable for timely control of crushing and proportioning systems. Although image recognition technologies for online detection of moving material particle sizes are developing rapidly, the fine particle size of coke powder and the harsh conditions in crushing and proportioning processes lead to severe challenges for image acquisition and size recognition. To address these issues, an image acquisition system for complex industrial environments was developed. The system comprises an image acquisition chamber and a multi-stage dust removal pipeline, designed to minimize the effects of lighting, temperature, and dust on image capture. Considering the fine particle characteristics of coke powder, an improved neural network was used for multiple training rounds to optimize the particle size recognition model, enabling the identification of surface particle size distribution on the conveyor belt. A predictive model was then constructed using machine learning algorithms, combining image-recognized surface particle size distribution data with manually sieved results to train a model capable of predicting the overall particle size distribution, thereby enhancing recognition accuracy. This image acquisition and particle size recognition system has been put into application in the coke powder crushing workshop of a domestic iron and steel enterprise. The application results showed that the recognition errors of the system for the particle size distribution ratio of coke powder in four intervals of (0, 0.5), [0.5, 3), [3, 5) and [5, ∞) mm were all less than 3%.
