Optimal Control of Iron Ore Flotation with Soft Sensors for Grade and Recovery


Procemin-Geomet 2015, Santiago, Chile, 21-23 October, 2015 

For Brazilian iron ore producers, reverse cationic flotation is preferred to obtain high quality concentrates from silicate ores. The SiO2 (silica) particles are rendered hydrophobic by amine (cationic collector) in alkaline conditions, while iron is depressed by the addition of starch. The injected air forms bubbles to which the silica attaches due to the amine action, leaving iron rich slurry at the bottom. Although this is undoubtedly an efficient process, there is clear scope for optimization as the concentrate’s SiO2 and Fe compositions must be traded off for maximizing mass recovery. This is essential for on-specification production at the lowest possible cost.

Step testing flotation banks for the purpose of Model-Predictive Control (MPC) implementation can be difficult when dynamics are slow and instrumentation is unreliable. It is harder when laboratory assays are necessary for the grade because some sampling points may be unreachable, staff may be insufficient and other reasons. Moreover, machine vision for froth analysis is not common on reverse flotation and, if available, it is susceptible to dirt, vibrations, weather, illumination and froth stability. Therefore, there is strong incentive for more comprehensive modelling of the underlying phenomena to deploy virtual sensors for grade. Neural networks have proven to be a key component of the model solution.

This article reports a nonlinear MPC solution on iron oxide flotation delivered for Vale at its Brucutu site. The design is based on a Quality and Recovery Controller (QRC) featuring virtual inline instrumentation for all of its controlled variables. This sends level setpoints to a slave Level Stabilization Controller (LSC) to optimize grade and recovery. Preliminary results show production of the most valuable concentrate type (%SiO2 ≤ 0.87) increased from 46% to 73% with 100 gr less amine used per feed SiO2 ton. Vale is planning replicating the solution at another site.


Manoel Morales (a)
Cássio Costa (b)
Tiago Caixeta (b) 
Carlos Quintero (c), corresponding author

(a) Rockwell Software, Brazil, (b) Vale S/A, Brazil and (c) Rockwell Software, Chile


To purchase this paper, select the paper and register with Gecamin Digital Publications here