The oldest and still most widely used scuffing criterion is the flash temperature hypothesis of Blok (1937), under which failure occurs when the sum of the bulk surface temperature and the transient temperature rise in the contact reaches a critical value. Later work has emphasised the collapse of the elastohydrodynamic film, desorption or exhaustion of boundary films, and the removal of protective oxide as contributing steps (Bowman and Stachowiak, 1996). Disc experiments have shown that surface finish strongly affects the load at which scuffing occurs at high sliding speed (Patching et al., 1995). The common element is that scuffing begins at isolated asperity contacts and spreads, so that a period of localised, intermittent adhesion precedes the general failure.

AE in a sliding contact arises from the elastic and plastic deformation of asperities, from fracture of asperities and surface films, and from the formation and rupture of adhesive junctions, and it occupies a band from tens of kilohertz to beyond one megahertz (Baranov et al., 2007). Early experimental studies related the AE count rate and root-mean-square (RMS) voltage to sliding speed, load and wear rate (Lingard and Ng, 1989), and distinguished adhesive from abrasive wear by the character of the signal (Boness and McBride, 1991). Fan et al. (2010) modelled the AE from sliding friction as the sum of elastic energy released at asperity contacts, which links the signal level to load, speed and surface topography. Hase et al. (2012) reported that adhesive and abrasive wear produce AE with different frequency content, an observation that suggests spectral features may help to identify the mechanism as well as its intensity.

In machinery diagnostics AE has been applied to bearings and gearboxes for many years, with the repeated finding that it responds to changes in asperity contact and lubrication regime earlier than vibration does (Mba and Rao, 2006). Tan and Mba (2005) identified asperity contact under rolling and sliding as the main AE source in a spur gearbox. Closer to the present subject, Saeidi et al. (2016) described automatic detection of scuffing from AE in lubricated sliding tests, and Shevchik et al. (2017) used time-frequency features with a random forest classifier to recognise the state preceding failure. These studies support the premise that scuffing has an acoustic precursor. They used particular rigs, sensors and classifiers, and they leave open how a laboratory should configure the measurement on its own machine.

The test conditions determine what the AE signal can be expected to show, so they are fixed first. The record should state the disc materials, hardness and surface finish, the Hertzian contact pressure at each load stage, entrainment and sliding speeds, oil supply temperature and flow, and the calculated specific film thickness. A stepped-load sequence with stages of constant duration, analogous to that used in standard gear scuffing tests, gives a clear structure for baselines: each stage has its own steady state against which departures can be judged.

A wideband piezoelectric sensor is preferred to a resonant one because the frequency content of the signal is itself of interest (Hase et al., 2012). The sensor cannot be placed on the rotating discs without slip rings or telemetry, so it is mounted on the nearest stationary component with a short, solid transmission path to the contact, normally a bearing housing or the loading arm. Every interface in that path attenuates and filters the wave. The path should cross as few joints as possible, the mounting face should be flat and clean, a couplant should be used, and the clamping force should be repeatable.

Before each test series the transmission path is checked with a repeatable artificial source. The pencil lead break described by Hsu and Breckenridge (1981) is the accepted source, and ASTM E976 describes its use for checking the reproducibility of sensor response (ASTM International, 2015). Leads are broken at a marked point close to the contact and the peak amplitude and spectrum are recorded, which gives a reference for comparing later tests. The noise floor is then recorded in three conditions: machine stationary, shafts rotating with the discs separated, and discs in light contact under full-film lubrication. These recordings show which frequency bands are occupied by motor, bearing and electrical noise.

Scuffing precursors are expected to appear as a rise in continuous emission with superimposed bursts, so continuous streaming is more suitable than threshold-triggered hit recording, which discards the signal between hits. The sampling rate must exceed twice the highest frequency retained, with an analogue anti-aliasing filter ahead of the converter. A high-pass filter set above the band occupied by mechanical noise, as identified in Step 3, removes most machine-borne vibration. A once-per-revolution pulse from each shaft should be recorded on the same clock so that features can be related to angular position.

The raw stream is reduced to a small number of features computed over windows synchronised with disc rotation. Table 1 lists the features proposed and the reason for each. They are deliberately few and physically interpretable. The RMS level tracks the overall rate of energy release at the contact (Fan et al., 2010); a measure of impulsiveness responds to discrete adhesive events; and band energies capture a shift in spectral content. Spectral kurtosis, developed for detecting transients in rotating machinery signals (Antoni, 2006), may be used to choose the band in which bursts stand out most clearly from the background.

Freshly finished surfaces emit strongly while asperities are truncated during running-in, and the level then settles. Detection must not begin until this has happened. For each load stage the procedure therefore discards an initial interval and takes the mean and variability of each feature over the following steady period as the baseline for that stage. Normalising by the stage baseline removes the step changes in level that accompany each increase in load or speed and leaves the drift and excursions that are of interest.

An onset rule turns the feature histories into a decision. A fixed multiple of the baseline is simple but reacts either too readily to single outliers or too slowly to gradual change. A cumulative sum scheme of the kind introduced by Page (1954) for industrial process control accumulates small persistent departures and is well suited to this task. The proposal is to run such a scheme on the normalised RMS level and on kurtosis in parallel and to declare an indication when either exceeds its limit for a set number of consecutive revolutions. The limits are set initially from the baseline variability and are treated as provisional.

An AE indication is a hypothesis about the surface. On each indication the test is stopped, or in early trials allowed to continue to conventional failure with the time of the indication logged. Friction torque, disc bulk temperature and, where fitted, electrical contact resistance are examined over the same interval, and the discs are inspected by optical microscopy and profilometry for transferred material and local roughening. Only indications matched by physical evidence of incipient adhesion count as confirmed events. The interval between an indication and conventional failure, and the number of indications with no corresponding surface change, are the two quantities by which the thresholds are then revised.