Ball Counter.

The Ball Counter watches grinding balls as they roll past the mill discharge. A two-class instance-segmentation network classifies every detection as a whole ball or a fractured fragment, and a tracker matches detections across frames so each unique ball is counted exactly once as it crosses the inspection zone.

What it measures

• Whole ball count & split ball count — per minute, per shift, per day
• Normalised population — fragments converted to whole-ball equivalents using fragment-area weighting, so the metric tracks effective charge mass, not raw detection noise
• Split fraction — the percentage of detections that are fractured, the leading indicator of media degradation
• Ball diameter distribution — percentiles P10 through P100 for the population
• Zone flux — cumulative counts with configurable daily or shift-level resets

How it changes the circuit

Ball wear used to be a manual count, performed when something already felt wrong. With a continuous, normalised population signal, the operator and the controller see media degradation building before it costs throughput, and ball-addition decisions become a data conversation instead of a hunch.

Deployment

RTMDet-style instance segmentation on edge GPU or CPU. Polygon ROI defines the inspection zone. Automatic reset scheduler for shift-level operational tracking.