Page 24 - Acoustics Today Summer 2011

ly A
Fig. 5. High level diagram of the Blind Source Separation Algorithm.
ˆ gearbox surface). This forces the independent sources S into
separate channels. The weight matrix W, once learned, is ideal-
-1
ery such as bearings using measurements made at the gearbox (or machine) surface and can in theory separate out all of the unwanted signals that are present in the mixture of signals at the measurement points. This allows an analysis of the “clean” signatures alone. This is the reason that ICA/BSS algorithms have seen wide application from voice processing to the isola- tion of sounds in the human chest cavity.
Acoustical signal processing methods for localizing gunshots
To an unalerted listener, the sound of gunfire represents an aural event in the soundscape that evokes a reaction of surprise tempered by curiosity. However, in engineering acoustics it represents a tran- sient acoustic signal generated by the dis- charge of a firearm from which information can be extracted such as the location of the point of fire22 (localization) and the type of firearm (classification). The sound pulse generated by the discharge of a bullet from a firearm is referred to as the muzzle blast or in military parlance—the report. The acoustic energy propagates at the speed of sound travel in air and expands as a spheri- cal wavefront (of constant phase) centered on the point of fire. Because the propagation of the sound through the atmosphere is omnidirectional, it can be heard from any direction, even behind the firer. If the lis- tener is positioned forward (towards the front) of the firer and the bullet travels at supersonic speed, then a second transient signal, which is referred to as the (ballistic) shock wave is heard. To the listener, the ori- gin of the shock wave occurs at a point Pn (referred to as the detach point), which is located at a distance xn along the trajectory of the bullet—see Fig. 7. When the listener is near the trajectory, the sound pulse (or sonic boom) is extremely loud. Similar to the muzzle blast wave, the shock wavefront
. By observing changes in the trace of the weight matrix W, one can observe points where the algorithm diverges (i.e., where there are abrupt statistical changes in the signals). By plotting the trace of W as a function of the gear tooth locations, the modulation of the meshing signal can be detected. Figure 6 gives a plot of the trace of W versus gear tooth number for a healthy gear and for the same gear at a later time as it begins to fail. The point of the future failure is very visible in the lower plot. This approach provides the possibility to non-invasively monitor the health of gears or other types of rotating machin-
Fig. 6. Result of Blind Source Separation Processing for a healthy gear (top) and for a gear that is beginning
to break (bottom).
20 Acoustics Today, July 2011