Best detector so far
Aligned multimodal, phase-aware is the current benchmark at 2.40s MAE on the reference run, with all four fan changes landing inside 5 seconds.
Roast logs, chamber-aligned computer vision, and audio combine into a machine-readable description of fluidization, lift, packing, and phase behavior in a fluid-bed roaster — not just “a fan event happened here.”
Not just “motion happened here” — chamber-relative measurements of how the bean mass is actually behaving.
Aligned multimodal, phase-aware is the current benchmark at 2.40s MAE on the reference run, with all four fan changes landing inside 5 seconds.
Movement mean rose to 0.052 on the latest run once the region-of-interest followed the chamber instead of a fixed crop. Chaos and bed-state numbers shifted too — the old measurements weren’t wrong, they were just unaligned.
On the stress-test run, the phase-aware aligned pass cut MAE from 30.40s to 6.84s by stopping the picker from treating power changes like fan changes.
Fan timing is still just a proxy. The stronger result is cleaner chamber-relative measurements for bed height, fluidization, coherence, and lift — usable across roasts and roasters.
A deliberate pipeline. Each step is a choice about what to measure and what to ignore.
Each roast is registered to the roaster, not to raw frame pixels.
Motion becomes bed height, fluidization, coherence, and lift.
Bean behavior is lined up with bean temp, RoR, fan, power, and roast events.
Phase-aware detection highlights moments that look operationally important.
Fan timing is a proxy check. The bean-state readings are the real product.
Vertical reach of the moving bed.
A direct read on whether the coffee is settled, lofted, or riding too high.
How freely the beans circulate.
Healthy fluidization usually sits in the middle, not at either extreme.
How tightly the bean mass moves together.
Higher coherence means a denser, more packed bed with less separation.
How much activity reaches the upper chamber.
Useful for spotting airflow changes that are subtle in the raw video.
Aligned, chamber-relative readings — movement and chaos shift with the workflow, not just the bean.