Motion-Robust Optical Granulometry for Comminution Circuits: A Multi-Stream Temporal Deep Learning Framework

Accurate near-real-time granulometry is critical for optimizing energy efficiency in comminution circuits, yet standard optical monitoring systems fail under the harsh conditions of industrial mining. High conveyor velocities induce motion blur, while heavy dust loads occlude particle boundaries, leading to severe under-counting of fine fractions and an overestimation of product coarseness. This paper proposes a novel tracking framework to overcome these limitations. The method integrates a five-frame temporal rolling buffer with a multi-stream augmentation architecture that decouples motion and texture features. Experimental validation on the feed of a secondary conical crusher at an operational facility demonstrates that the proposed framework recovers significant particle data lost by baseline models, correcting the measured D80 and D50 values of the crushed material. This correction reveals a finer true load, enabling more precise closed-loop control of crusher settings and reducing unnecessary energy consumption.