Flotation Monitoring: Rapid Optimization Support

Ivan Zaporozhets25 May 2024

Problem

A small data science team at a Norilsk Nickel site runs 20 flotation machine optimizers (OFM), auto-control services that set process actions every 15 minutes. The process swings with ore quality and reagent stocks, so each model needs constant retraining and rule changes. Technologists can override or cancel any action. When a service works badly, it disappoints both the team and the site. Updates were slow: one request could take about a month from incident to launch.

Example from the article: a request to update an algorithm or model took about 31 days; uptime on medium-maturity optimizers was only 30%.

Solution in brief

The team added two tracks: daily duty-shift checks of home-made dashboards, and a fast request-handling loop. The loop combines dashboards that technologists can read via screenshots, a refactored optimizer core, and a chat with the technologists.

Questions this case answers
  • How to keep an automatic process optimizer reliable when conditions drift constantly and operators can override its actions
  • How to shorten the cycle from an operator complaint to a model or logic update in a production control service
  • How to keep a production ML auto-control service running when process conditions drift all the time
  • How to monitor an optimizer that technologists can override, with uptime and task-completion metrics

Tools and methods

  • Weighted compromise loss with grid search — picks the action with the smallest forecast error sum.
  • Duty-shift dashboards — uptime, task completion, technical service metrics.
  • State snapshots and state detail tables — shared view of the situation for DS and technologists.
  • Python, Dash, Plotly, Waitress, apscheduler — dashboard stack.
  • Core and Custom Part classes — DataProvider, DataContext, Processor, Optimizer.
Continue reading — it's free

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