A Gallop Through Production Sites: ML at Nornickel

Alexey Klokov27 May 2023

Problem

Nornickel, the largest non-ferrous metals producer, runs a chain of five stages: mining, crushing and grinding, flotation, thickening and smelting. At each stage technologists and operators watch the process by eye and by routine checks. This is costly, and a missed event can stop a workshop. The speaker's team builds "little helpers" that take this monitoring off people, mostly as prototypes.

Example from the article: a tooth breaks off an excavator bucket, rides unnoticed into the crusher and the mill, and production may stop for a week. Before, the driver checked about 12 teeth after every loading cycle.

Solution in brief

Put a camera at each stage and let a model show the technologist a signal: detection of bucket teeth, segmentation of rocks, froth bubbles and water turbidity. Where control is needed, use reinforcement learning (RL) in a simulation (mill load, blending), or a linear functional of controllable features that is minimized directly (flotation).

Questions this case answers
  • How to take routine visual monitoring of a heavy industrial process off operators with cameras and a vision model that raises alerts
  • How to optimize a process setpoint without building a reinforcement learning simulator, by minimizing a learned linear function of controllable features
  • How to take routine equipment monitoring off operators in a mining and ore-processing plant with cameras and ML
  • How to detect a missing bucket tooth on a mining excavator before it reaches the crusher

Tools and methods

  • Object detection — missing bucket teeth in video
  • Segmentation — foreign objects and rock sizes on conveyors, froth bubbles, water turbidity
  • Bubble tracking — speed of froth removal
  • RL agent in a simulation — mill load and blending of concentrates
  • Linear functional with a minimizer — flotation setpoints
Continue reading — it's free

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