Neural Cell Supervisor Control

Application

Flotation cells are deceptively difficult to control well. Airflow, pulp level, concentrate flow and feed-tank levels all interact, and the froth responds to changes with delays and non-linearities that frustrate fixed setpoints and manual intervention. Operators tend to run conservatively to avoid instability, leaving recovery and throughput on the table.

Minealytics Neural Cell Supervisor Control is an integrated supervisory-control architecture for flotation cells and banks. It can fuse froth-vision metrics, such as bubble size, froth velocity and stability, with conventional process-control signals such as airflow, pulp levels, concentrate flow and feed-tank levels.

The architecture is deliberately layered so that fast, safe regulatory control is never compromised by the slower, exploratory work of optimisation.

Solution: A Three-Layer Architecture

Layer 1 — PLC (“the muscles”). Fast regulatory PID loops hold the physical setpoints for airflow and pulp level. This layer retains all of the plant’s existing interlocks and hard permissives. Nothing above it can override a safety limit; the PLC remains the final authority on what the equipment is allowed to do.

Layer 1.5 — Stabilisers and conditioners. Sitting between the supervisor and the PLC, these are deterministic, bounded, rate-limited modules that translate supervisory targets into safe physical setpoints. They clip values to allowable ranges, rate-limit how quickly setpoints can move, freeze outputs when conditions warrant, and protect against runaway recycle, level and air behaviour. This layer performs conditioning, not optimisation — its job is to guarantee that whatever the optimiser asks for is delivered to the plant in a safe, smooth form.

Layer 2 — Supervisor optimiser (“the brain”). An AI optimiser searches slowly over internal supervisory targets — air splits and pulp-level targets per cell group — evaluating an objective function over stable windows of operation. By adjusting targets gently and observing the response, it learns which direction improves performance without disturbing the circuit.

The Objective Function

The optimiser balances competing goals through a single objective that:

• Rewards stable concentrate flow and high froth velocity, which serves as a proxy for recovery.
• Penalises froth collapse, poor bubble health, excessive movement of setpoints, proximity to air and level limits, and recycle instability.

This balance keeps the optimiser pushing toward better recovery while actively discouraging the behaviours that lead to instability, so gains are durable rather than transient.

Safety and Scaling

The supervisor is safety-first by construction. It never violates the PLC’s hard limits or the circuit’s mass-balance rules, and it pauses or overrides its own actions when the plant enters an unstable or constrained state — handing authority back to the conditioning layer and the PLC until conditions recover.

The same architecture scales from dual-cell control up to full-bank coordination. With feed-forward cross-cell biasing, an air change made upstream is anticipated downstream, so later cells do not have to overcompensate for adjustments made earlier in the bank. Where cameras are installed the supervisor is vision-informed; where stages are only poorly observed — such as some recycle streams — it treats them conservatively rather than acting on incomplete information.