We piloted a camera-driven air/level/mass-pull loop on a Cu rougher last quarter and saw +0.6% recovery with 8% less collector at about 1,200 t/d, but night-shift lighting and model drift were persistent issues; anyone running full closed-loop dosing off froth vision long-term? Looking for specifics on lighting setups, retrain cadence, and how you gate the control to avoid chasing noise during ore-type step changes.
Ran closed-loop on Cu roughers at about 1,200–1,500 t/d by fixing light, not the model: 5000K fixed‑lux LED bars with cross‑polars at 45°, exposure locked, and a white reference tile in-frame for a nightly gain check; retrain monthly only if a 7‑day drift flag trips. > Looking for specifics on lighting setups, retrain cadence, and how you gate the control to avoid — we only enable dosing when vision confidence >0.8 for 10 min, cap collector moves to 0.2 L/t per 5 min, and hard‑freeze on >5% mill power or pH jumps; is night‑shift exposure locked on your cameras?
At night, strobe the LEDs and sync to the camera; a hooded ring light plus a simple lux/glare score lets you gate control so it doesn’t chase ghosts at 2am… We only let dosing move when image health and froth velocity stay in-bounds for 5 min, otherwise it ‘holds last good dose’ with a hard rate limit; retrain is quick labeled top-ups every 2 weeks and a quarterly full refresh. @mason_b92’s fixed light point is solid — have you tried strobed bars to kill shift-to-shift drift?