A new weakly supervised learning approach for real-time iron ore feed load estimation

• First attempt using image-based methods for iron ore load estimation. The problem is modelled as a weakly supervised learning problem. A new two-stage modelling training neural networks are proposed. Empirical evidence of experiments shows good performance and economic boost. Iron feed-load control one the most critical settings in mineral grinding process. It has direct impact on quality final products. setting feed heavily replies characteristics pellets. However, such challenging to acquire many production environments, requiring speical equipments complicated process with high cost. To provide an low-cost easier-to-implement solution, this paper, we present our work deep models estimation from pellet images. address challenges caused by large size images shortage accurately annotated data, proposed use apporach model algorithm two network architectures developed. experiment results show competitive performance, trained can be used real-time grind optimisation.