← All articles SAG Mill · Our products

Sound Analyzer: Listening to the SAG Mill, with Fill Analyzer and Ball Counter

What the mill's sound says about fill, impacts and media, and how three instruments turn it into numbers a controller can use

Our products · 7 min read · 28 October 2026

Key takeaways

Operators have always listened to the SAG mill. Sound responds in seconds to where the charge is landing, while bearing pressure and power respond in minutes and can mislead: power reads the same on both sides of its peak, and bearing pressure drifts with temperature and liner wear (Pax et al., MMG Century).
Acoustics alone has a mixed record as a control signal - plants report good alarms but unreliable control from single 'electronic ears'. The lesson from the literature is to use sound as one input in a model, not as a stand-alone setpoint (control.com practitioners; La Rosa et al., Cannington).
The Aeterna Sound Analyzer turns shell noise into control-ready tags five times a second: a weighted sound level, 31 third-octave bands and an 'impact rhythm' that checks whether media lifts and lands once per lifter pass (expected rate = rpm/60 x lifter rows).
The Fill Analyzer fuses the sound (or bearing pressure), ball counts and measured power in a two-state Kalman filter built on the Morrell power model, estimating total fill Jt and ball fill Jb continuously. Published estimators of the same family reach 79-96 % accuracy on ball level (UdeC thesis) and are the basis of Hatch's 2023 SAG work.
Ball Counter keeps the ball-fill side honest by counting balls into (and out of) the mill and tracking split balls, so ball charge no longer waits for a crash stop. Media condition matters: removing worn scats raised Navachab throughput by ~11 %.
Together they answer the two questions an operator asks every shift - how full is the mill, and where is the charge landing - and feed both to the MPC.

Why listen to a SAG mill

A SAG mill is a tumbling mill where rock and steel balls are both the grinding media. Its performance depends on three things the operator cannot see directly: how full the mill is (total fill Jt, % of mill volume), how much of that is steel (ball fill Jb), and where the falling charge lands - on the toe of the charge, where it breaks rock, or on bare shell liner above it, where it breaks liners and balls.

The conventional signals are bearing pressure (or load cells) and motor power. Both are useful and both are slow and ambiguous. Bearing pressure weighs everything on the bearings, so liner wear and relines shift its zero, and at MMG Century it was described as noisy, temperature-sensitive and lacking useful sensitivity (Pax et al., 2014). Power peaks at an intermediate filling, so the same kilowatts can mean the mill is under-filled or overloading.

Sound arrives in seconds. When balls strike bare shell the mill rings; as the charge builds and slurry cushions impacts, the sound becomes muffled. Experienced operators use exactly this - and early 'electronic ear' systems tried to automate it.

Figure 1 - How fast each signal sees a change inside the mill (indicative)

What the literature says about acoustic monitoring

Early systems used one or two microphones; one patent describes the two-point method as crude because it samples intensity at only two fixed points.
Practitioner experience is mixed: 'good for alarm but not for control' is a common view, with others reporting that well-maintained ears indicate load level well.
MMG Century replaced failing load cells with a 12-microphone array as a load proxy; at 1,100 t/h the cyclic acoustic signal was used to stabilise the mill through feed water, after a >25 % throughput push had exposed new instabilities (Pax et al., MetPlant 2015).
At Cannington, Metso's SmartEar acoustics and an online model-based charge estimator were compared with bearing pressure; the authors concluded that redundant load estimates give more confidence (La Rosa et al., MetPlant 2008).
Shell-mounted vibration sensors (e.g. Molycop MillSlicer) report impact angle and toe angle; an IMPC 2024 case held impact angle around 135 degrees through a liner campaign by trimming mill speed.
Model-based estimators using the Morrell power model and a Kalman filter can estimate both Jt and Jb (Toor et al., SAG 2023; Iglesias et al., IMPC 2014; Herbst et al., 1990).

The Sound Analyzer

The Sound Analyzer listens to a microphone near the mill shell and publishes, five times a second, the numbers a metallurgist can reason with. The audio is windowed and transformed into a spectrum (about 6 Hz resolution), weighted with a standard A, C or Z curve, and integrated into a sound level in dB over a chosen frequency band. The same spectrum is split into 31 third-octave bands from 20 Hz to 20 kHz, so changes in character - ringing versus muffled - are visible, not just changes in loudness.

The second output is the impact rhythm. Each lifter row throws media once per pass, so impacts repeat at (mill rpm / 60) x number of lifter rows; a 10 rpm mill with 60 lifter rows gives 10 impacts a second. The analyzer demodulates the 500-5,000 Hz band into an envelope and looks for that repetition. A strong, narrow peak means media is lifting and landing as designed; a weak or smeared peak points to overload, pooling or slip. The search can lock to the kinematic rate or find the peak automatically.

Tags go to the DCS and historian like any other instrument, and a microphone can also contribute its impact position to the Fill Analyzer's trajectory check.

Reading the impact rhythm

Figure 3 shows the idea with illustrative signals. In the operating window the envelope spectrum has a sharp peak at the expected impact rate and its harmonic. As the mill fills and slurry cushions the charge, the peak drops and broadens. A falling sound level with a smearing rhythm is an earlier, clearer overload warning than power flattening a few minutes later.

Figure 3 - Impact rhythm, healthy versus overloading (illustrative)

Fill Analyzer: what is inside the mill

No single signal measures both total fill and ball fill. The Fill Analyzer combines several. A fill indicator - bearing pressure or the Sound Analyzer level, calibrated by regression - gives a first estimate of Jt. Ball Counter's counts, less a wear model between deliveries, give a first estimate of Jb. The Morrell power model then predicts the power those two fillings should draw for the mill's geometry, speed, liner wear and densities.

A two-state extended Kalman filter compares predicted with measured power each cycle and corrects Jt and Jb. Engineers can choose which estimate to trust when they disagree (for example, trust the ball counter and let total fill absorb the correction), and calibrate against crash stops, grind-outs and liner scans when they happen. Outputs include Jt, Jb, rock fill (Jt - Jb), charge mass and predicted versus measured power - a continuous health check on the whole estimate.

This is the same family of method described in the literature: Kalman filters with a power model (Herbst, 1990; IFAC 2008 comparison favouring the extended Kalman filter) and Morrell-based Jb/Jt estimation (Iglesias, 2014; Toor et al., 2023).

Ball Counter: the ball-fill half of the picture

Ball charge is usually measured by a crash stop followed by a grind-out: the mill is stopped with feed cut, then run without ore for 10-20 minutes so balls can be measured. It is accurate and it costs production every time. SAG mills typically carry 8-21 % ball charge, and the right level moves with ore and liners.

Ball Counter uses a camera and an instance-segmentation network to count balls as they pass, classifying each as whole or split. Split fraction is an early sign of media degradation, and a normalised count turns fragments into whole-ball equivalents. Counting balls in (and, where fitted, out) gives the Fill Analyzer a live Jb prior between grind-outs.

Media condition has a measurable effect: at Navachab, removing 30 t of worn scats raised throughput by about 11 % (Powell & Smit). Knowing the ball charge continuously lets the plant act before the next shutdown.

Signals compared

Signal	What it measures	How fast it responds	Main weakness	Aeterna instrument
Bearing pressure / load cells	Total mass on the bearings (charge + liners + shell)	Minutes	Noisy, drifts with temperature; liner wear shifts the zero	Input to Fill Analyzer
Motor power	Torque needed to lift the charge	Minutes	Same kW on both sides of the power peak, so it cannot tell under- from over-filled	Cross-check inside Fill Analyzer
Sound level	Energy of impacts on the shell	Seconds	Needs care with mic position, background noise and calibration	Sound Analyzer
Impact rhythm (envelope spectrum)	Whether media lifts and lands once per lifter pass	Seconds	Interpretation is site-specific; needs mill speed and lifter count	Sound Analyzer
Fill and charge angles	Total fill Jt, ball fill Jb, toe / impact position	Seconds to minutes	Model needs geometry, liner wear and periodic calibration	Fill Analyzer
Ball additions and condition	Balls in (and out), split fraction, size	Continuous	Camera view must be kept clean	Ball Counter
Crash stop + grind-out	Direct measure of Jt and Jb	A few times a year	Lost production at every measurement	Used to calibrate

How it fits with control

The three instruments do not replace bearing pressure or power; they add the fast and the missing signals. With Jt and impact position measured continuously, the MPC can hold the load closer to its productive limit and trim speed to keep media landing on the toe as liners wear. See the companion pack R-038 (SAG APC) for how sound, bearing pressure and power are used together as constraints.

Common misconceptions

'Loud means overloaded.' Usually the opposite: a ringing mill is often under-filled with balls hitting bare shell; overload muffles the sound.
'Sound can replace bearing pressure.' Single microphones are better alarms than controllers; the value comes from fusing sound with load and power.
'Ball charge only needs checking at shutdowns.' Media wear and splitting change the charge every day; a continuous count closes the gap between grind-outs.
'Power tells you how full the mill is.' Power peaks at an intermediate filling, so on its own it cannot tell which side of the peak the mill is on.

Uncertainties and limits

Published papers on acoustic SAG control (Pax, La Rosa) give qualitative results in their abstracts; quantitative accuracy comparisons sit behind paywalls and were not verified here.
Microphone placement, neighbouring equipment noise and calibration drift affect sound level; site commissioning and periodic calibration are part of the product, not an extra.
Fill Analyzer accuracy depends on geometry, liner wear model and calibration points; published Jb estimators range from 79 % to 96 % accuracy across mills.
Aeterna's published outcome range (+1-3 % throughput, 3-8 % lower kWh/t) refers to the full MPC plus instruments, and the site lists case figures as sample numbers - treat them as targets until a site case is published.

Curious how full your SAG mill really is between grind-outs? Ask us about a Fill Analyzer trial.

Talk to an engineer

Sources

Implementation of acoustic arrays for semi-autogenous grinding mill operation, Pax, Wynn, Mann, Cornish - 12th Mill Operators' Conference (AusIMM), 2014
A novel measurement for the internal operation of a SAG mill using acoustic sensors - a case study, MetPlant 2015 (AusIMM), 2015
Understanding size effects of SAG mill operation (MMG Century), Pax & Cornish - 13th Mill Operators' Conference (AusIMM), 2016
The use of acoustics to improve load estimation in the Cannington AG mill, La Rosa, Cantarutti, Wortley, Ozkocak - MetPlant 2008, 2008
Electronic ear in SAG mill (practitioner forum), control.com
On-line sensors for measuring the total ball and charge level in tumbling mills, University of Cape Town (thesis)
MillSlicer: SAG mill, Molycop
Impact angle control of a 32 ft SAG mill (IMPC 2024 proceedings), XXXI IMPC 2024 (SME), 2024
Real-time estimation of SAG mill charge characteristics for process optimization, Toor, Valery, Morrell, Duffy (Hatch) - SAG 2023, 2023
Development of a dynamic Jb/Jc estimation model in a single stage SAG mill, Iglesias et al. - IMPC 2014, 2014
Continuous measurement and control of charge volume in tumbling mills, Herbst et al. - SME, 1990
State estimation for a SAG mill (EKF vs particle filter vs MHE), IFAC World Congress 2008, 2008
Online ball level estimation using a power model, Universidad de Concepcion (thesis), 2024
Online SAG mill pulse measurement and optimization (crash-stop and grind-out practice), US DOE (EERE) project report
Startling effect of ball scats removal on SAG mill performance, M. Powell (UCT) & I. Smit (AngloGold)
System for monitoring mechanical waves from a moving machine (patent), US Patent 6,874,364
Fill Analyzer; Ball Counter (product pages), Aeterna Automation, 2026
Smart Instruments (BallManager), Aeterna Automation, 2026