Machine Learning for Flotation Optimization

Yuriy Katser25 May 2024

Problem

A copper flotation circuit has 19 flotation machines in four operations and processes about 700 t/h. Operators set air flow, pulp level and reagent dosage by hand, and the ore keeps changing. Metal lost to the tailings cannot be recovered, and the concentrate must still meet the shipment quality. Pure data-science controllers are black boxes that technologists may refuse to use, while classic advanced process control (APC) needs frequent retuning.

Example from the article: a copper circuit with 19 flotation machines and about 700 t/h, where the task is to raise recovery at the same concentrate quality, with fewer manual decisions.

Solution in brief

The control has three layers: fuzzy-logic rules written with the plant staff, prediction models over a horizon of tens of minutes, and MPC on top. Fuzzy logic stays as the fallback when sensors or models fail. Computer vision on the froth adds a virtual sensor that limits the controller or feeds MPC.

Questions this case answers
  • How to build a controller for a continuous multi-stage process that raises yield while keeping output quality, combining rules, predictions and model predictive control
  • How to keep a learned controller running when sensors fail or the process drifts, with a rule-based fallback that operators trust
  • How to build model predictive control (MPC) for a flotation circuit that raises copper recovery while keeping concentrate quality
  • How to make an industrial ML controller that plant technologists will accept and trust

Tools and methods

  • Fuzzy logic rules — interpretable base layer built from dependency analysis and expert knowledge
  • Prediction models, 10–40 min horizon — forecast the state of the circuit
  • MPC — picks the optimal trajectory of air, level and reagent setpoints
  • Froth computer vision — bubble size, speed, stability, luster, color
  • X-ray fluorescence analyzers and flowmeters — online quality and flow signals for MPC
Continue reading — it's free

Google or a one-time email code. No password.