What the mill shell is trying to tell you.

Every tumbling mill is already a sensor. The shell is a steel diaphragm that gets struck thousands of times a minute by balls, ore, and liners, and each impact leaves a signature in the vibration spectrum. The problem has never been that the signal isn't there — it is that the signal is loud, overlapping, and alive on timescales that classical plant instruments were not built for.

Mill fill, cataract-vs-cascade transition, liner wear state, grind-out risk — the physics for each of these is in the shell signal. Reading them in real time is where the engineering lives.

The problem with reading the shell today

Most installed vibration kits were designed as condition-monitoring tools: a scalar RMS value sampled slowly, fine for watching bearing health, far too slow to tell the controller anything useful about what is happening inside the drum right now. The information that the operator needs — am I under-filled, am I at the overload cliff, is a liner about to let go — sits in the impact-rate structure of the signal, not its average amplitude.

Extracting it demands wide-bandwidth sampling and a demodulation pipeline that runs continuously, not an alarm limit on an RMS reading. That is a different class of instrument.

What we are building

Our in-flight prototype — the Fill Analyzer — runs an envelope-demodulation DSP pipeline close to the digitizer, so the heavy lifting happens before the data ever leaves the mill. The output the controller sees is not raw vibration; it is a small set of physically interpretable signals: live impact rate on the shell, estimated fill fraction, and a liner-health trend.

• Wide-bandwidth acquisition at the shell, DSP at the digitizer
• Envelope demodulation to isolate impact-rate structure from steady vibration
• Interpretable outputs for the controller — fill, impact rate, liner-health trend
• No new sensors on the mill beyond the vibration pickup itself

Why this matters for the circuit

The SAG runs scared because the operator cannot see the overload cliff until they are on it. A live fill signal with real latency budget moves the cliff from "a surprise" to "a known distance away" — the margin the plant has been paying for in unmilled tonnes becomes a margin the controller can safely close. Liner-health as a trend, not a shutdown-inspection snapshot, is the same story on a longer time axis.

This is still research in progress. The validation path is digitizer-to-server, not the HTML prototype we used to sanity-check the DSP math. When we have numbers worth standing behind, they will appear here.