Convolutional neural networks for froth imaging in a flotation machine with electromagnetic regrinding impeller

This study investigates a vision-based supervisory framework for a prototype flotation machine equipped with an in-line electromagnetic regrinding impeller, a configuration intended to enhance mineral liberation and recovery for fine particles and tailings. Building on recent advances in machine vision for process industries, convolutional neural networks (CNNs) were trained to infer supervisory variables directly from froth images: (i) bubble population metrics (bubble count/density) and (ii) overall concentrate yield. An image corpus of > 1000 froth frames was collected across three trials under deliberately varied operating conditions (airflow, electromagnetic mill frequency, and feed mass), yielding substantial covariate shift. The results provide preliminary evidence that CNN-based froth imaging can supply actionable supervisory signals in electromagnetic-assisted flotation. Future work will expand datasets, harden illumination robustness, incorporate spatiotemporal modeling and sensor fusion, and evaluate closed-loop control in prospective trials.