Envelope and Cepstrum Analyses for

Machinery Fault Identification

Hans Konstantin-Hansen and Henrik Herlufsen, Brüel and Kjær, Nærum, Denmark

All rotating machinery parts wear, as does ancient wisdom.

Envelope and cepstrum analyses tend to be overlooked, but they
are still useful tools to identify local machinery faults. The intent
of this article is to introduce these two tools in a pragmatic way
and to demonstrate their properties by using a couple of practi-
cal examples.

The spectrum of a random or random-like signal is a continuous

spectrum. The spectrum of a sine signal is a single discrete line
in the spectrum. The spectrum of a repeated impulse, repetition
period = T, is a discrete line spectrum throughout the frequency
range of interest with a line spacing of 1/T Hz.
A local fault in a machine could be a damaged tooth in a gear, a
faulty turbine blade, a crack or spalling in a rolling element bear-
ing, etc. Each time a local fault is engaged, the machine structure
is exposed to an impact force, an impulse, and the corresponding
vibration spectrum measured on the structure will contain all the
harmonics of the repeated impact event. So, just as a sine signal
is periodic in the time domain, the spectrum of a local fault is
periodic in the frequency domain, which is the property exploited
in envelope and cepstrum analyses.
As noted, the response of a local fault is measured somewhere
on the structure and not where it originates. This means that the
response is the repeated impact seen though a certain, unknown
transmission path.
In mathematical terms, the response spectrum, Y(f), equals Figure 1. (a) Baseband spectrum; (b) Envelope spectrum; (c) Zoom spec-
the product of the excitation spectrum, X(f), and the spectrum of trum.
transfer path, H(f): Y(f) = H(f) X(f). carrier, but there is no carrier to modulate. Just like the carrier in
Envelope analysis benefits from the inevitable resonances found radio communication is the transmission facilitator, the structural
in any mechanical structure. At resonances, H(f) amplifies X(f), resonance facilitates the transfer of the impact information. In
the excitation, making it possible to identify local faults at a very short, the resonance is the (imaginary) carrier, and the impact in-
early stage. formation can be retrieved by amplitude demodulation/envelope
Cepstrum analysis bypasses the transmission path and reveals detection around the resonance. The result in the time domain
the fault directly. This is important for trend analyses in cases and a subsequent Fourier transformation together constitute the
where the transmission path changes, because the mechanical envelope analysis and gives the spectrum of the local fault. The
structure changes, is modified, or the response accelerometer result is shown in Figure 1b, the first eight harmonics of a local
is not mounted in exactly the same way and position for every fault with an occurrence of 42 Hz.
measurement. The zoom spectrum around the resonance, Figure 1c, already
holds the required information, but the result is most often not
Envelope Analysis operational. If the first harmonic varies, say 1 Hz due to RPM
Envelope analysis, alias amplitude demodulation, dates back variation, the Nth harmonic will vary by N Hz. The harmonics
to the dawn of radio communication. In AM transmission, the will smear, the higher harmonics will merge and the result is not
information signal is amplitude modulated onto the carrier, which intelligible – not to mention the problems to deal with if more local
in turn is transmitted via the antenna. The purpose of the carrier is faults at different occurrences are in play.
only to carry the information and facilitate the transmission. The
receiver tunes in on the carrier frequency and demodulates the Example: Local Faults in Roller Bearings
signal to retrieve the information signal. The demodulation is band Rollers, or balls rolling over a fault in a bearing, produce a series
pass filtering around the carrier frequency followed by detection; of force impacts. The rate of the impacts is determined by the RPM
that is, the creation of the envelope of the filtered signal, which is of the shaft involved and the geometry of the bearing. The repetition
now the information signal that originally modulated the carrier. rates are denoted bearing frequencies and they are as follows:
How does this relate to identification of local faults in rotating • BPFO: ball passing frequency outer race, local fault on outer
machinery? As previously mentioned, the local fault is manifested race
by the harmonics of the impact frequency. The response signal is • BPFI: ball passing frequency inner race, local fault on inner
the harmonics weighted by the transmission path, so the obvious race
frequency range in which to look for the harmonics is where the • BFF: ball fault frequency = 2 ¥ BSF, local fault on rolling ele-
harmonics are predominant, which is around the structural reso- ment
nances where the harmonics are amplified. • BSF: ball spin frequency
Figure 1a shows the response spectrum from a machine with and • FTF: fundamental train frequency, fault on the cage or mechani-
without a local fault. The green curve shows the vibration response cal looseness
to a single impact with an ordinary hammer while the machine is Most bearing suppliers offer a bearing frequency calculator,
at rest. It shows a structural resonance at 1.9 kHz. The red curve where the inputs are the bearing type number and the RPM of the
shows the response of the local fault. The highlighted (yellow) part shaft involved and the output is the bearing frequencies; that is,
of the spectrum looks like the spectrum of an amplitude-modulated the potential fault frequencies.

10  SOUND & VIBRATION/MAY 2010 www.SandV.com

 tion signal definitely confirms the presence of a local fault. The
propulsion train comprises some 10 (different) bearings and a
reduction gear, meaning there are at least 50 different local fault
candidates.
Figure 3b shows the baseband spectrum of the vibration signal.
Unbalances, misalignments and gear-tooth meshing frequency
show up at the lower frequencies, and the complete spectrum looks
somewhat peaky, again suggesting a local fault. The peak at 2.6 kHz
indicated by the yellow range looks like a structural resonance and
invites envelope analysis of the indicated range.
Figure 3c shows the envelope of the time signal that clearly en-
hances the impact events. And Figure 3d shows the final spectrum
of the envelope, from which the fault can be identified.
The fault signature is that of a clean inner race crack. It shows
three modulation groups around the three first harmonics of 35
Hz and modulation sidebands within the groups of 2.95 Hz. The
propeller shaft is running at 177 RPM = 2.95 Hz, and at this speed,
Figure 2. Second modulation. one of the bearings has a BPFI of 35 Hz. Together, this doubles
the evidence for an inner race fault in a particular bearing, which
proved to be correct when the bearing was replaced.
The real power of the method is that it is able to identify multiple
local faults with different fault rates. The different faults will show
as different harmonic families in the envelope spectrum, which in
turn can be identified using the harmonic cursor.

Cepstrum Analysis
The cepstrum is a relative of the well-known autocorrelation,
RAA(t), which can be derived from the autospectrum as:
RAA (t ) = F -1 (SAA ( f )) (1)
where F-1is the inverse Fourier transform and SAA(f) is the au-
tospectrum. RAA(t) is in the time domain (more properly the lag-
time domain. The autocorrelation reveals the delay information
inherent in a periodic autospectrum. One example is identification
of acoustic reflections.
The cepstrum is derived similarly; it is the inverse Fourier
transform of the logarithm of the autospectrum:

{
C AA (t ) = F -1 log (SAA ( f )) } (2)
This has two implications:
• The logarithm boosts lower levels of the autospectrum. This
means that a low-level part of the spectrum that holds periodic
information that would not show in the autocorrelation will
show in the cepstrum.
• As noted previously, the response spectrum is the product of
the excitation spectrum and the spectrum of the transfer path:
Y(f)=H(f) X(f) or in power terms:
SYY ( f ) = H ( f ) SXX ( f )
2
(3)
The cepstrum of the response spectrum becomes:

{ } { (
CYY (t ) = F -1 log (SYY ( f )) = F -1 log H ( f )
2
)} + F {log (S
-1
XX ( f ))} (4)

which means that the effect of the transfer path on the excitation
Figure 3. (a) Baseband time signal; (b) Baseband spectrum; (c) Envelope of
filtered time signal; (d) Envelope spectrum. is additive in the cepstrum, and the two can be separated in the
analysis, as will be shown later.
Second Modulation. As stated, the engagement of a local fault The term cepstrum looks like a spelling error and is often cor-
produces an impact force. If the engagements are subjected to the rected by proof readers. It is not an error; the term has been chosen
same (radial) load, the resulting pulse train will be uniform as to give the function a scent of “spectrum” because it is advanta-
shown in Figure 2. The fault in the outer, stationary race of the geous to consider the cepstrum as the spectrum of (the logarithm
bearing is subjected to the same load for each over roll of the ball. of) the autospectrum.
A fault on the rotating inner race, on the other hand, is subjected To support this perception, all terms related to the cepstrum are
to a varying load, where the frequency of the load variation is the warped versions of the corresponding spectrum terms:
RPM of the inner race. Consequently, the envelope signal will be • Spectrum: cepstrum
amplitude modulated as shown in Figure 2. This is a real ampli- • Frequency: quefrency
tude modulation that will show in the envelope spectrum. This • Harmonics: rahmonics
modulation should not be confused with the (pseudo) modulation • Low-pass filter: short-pass lifter
of the structural resonances that is demodulated by the envelope • High-pass filter: long-pass lifter
analysis. The lifters are used to edit the cepstrum prior to a Fourier trans-
Ship Propeller Shaft. Figure 3a shows the vibration signal form back to a spectrum, the frequency domain:
measured on a bearing of a ship propeller shaft. The signal clearly
shows some impact events, and an audio playback of the vibra- Liftered Spectrum: log (SLL ( f )) = F {Lifter (t ) Cepstrum(t )} (5)

www.SandV.com SOUND & VIBRATION/MAY 2010  11

 The cepstrum is useful for all periodic manifests in a response
spectrum from a rotating machine like amplitude modulation of
gear-tooth meshing frequency due to load variation, gear eccen-
tricity or general wear of gears and, of course, for identifying local
faults, which is the aim of this article.

Example – Local Fault in a 5:3 Gear

The input gear spins with 3000 RPM = 50 Hz, which means that
a fault on a tooth will generate a repeated impact with the repeti-
tion time t = 1/50 Hz = 20 ms. Figure 4. Original and liftered vibration spectra.
The output gear spins with 3000 ¥ 3/5 RPM = 1800 RPM = 30
Hz. A faulty tooth on this gear will give a repeated impact every
t = 1/30 Hz = 33.3 ms.
The noise from the gear suggests a local fault, and the vibration
was measured on the housing of the gear. The green curve in Figure
4 shows the vibration spectrum. The spectrum appears peaky; the
peaks may include one (or more) periods due to local fault(s)
The cepstrum of the vibration spectrum will reveal a possible
periodicity in the spectrum. The green curve in Figure 5 is the
cepstrum, and it reveals two rahmonic families indicating two
local faults with a different fault rate: Figure 5. Vibration cepstrum.
• A rahmonic family with the fundamental rahmonic = 20 ms
(50 Hz), which means a local fault on the input gear, a faulty encompasses the rahmonics caused by the local faults, and the
tooth. short-pass lifter covers what is left, that is, the nonperiodic part
• A rahmonic family with the fundamental rahmonic = 30 ms of the original spectrum.
(33.3) Hz, indicating a faulty tooth on the output gear. The transformations back to the frequency domain are shown
The reason for the rahmonic component at quefrency 60 ms being in Figure 4. Recalling the relation Y(f) = H(f) X(f), the chosen lift-
higher than the other components is that both local faults contribute ers have effectively separated the H(f) term, the unknown transfer
to this component: The 3rd rahmonic of the fault on the input gear path (blue) and X(f) term, the spectrum of the local fault at the
and the 2nd rahmonic of the fault on the output gear. origin of the fault (red).
At quefrency 120 ms, the two faults should again combine, but A claim was that the Cepstrum bypasses the unknown transfer
the rahmonics seem to have drifted apart. The sixth rahmonic of path and reveals the local fault directly. As shown previously, the
the input fault comes out at 120 ms, but the fourth rahmonic output unknown transfer path wraps up at the short quefrencies in the
fault comes out at 121 ms. cepstrum, and the local faults are manifest at the long quefrencies.
The reason is that the gear ratio is not exactly 5:3 but rather 47:31, This qualifies cepstrum analysis for trending local faults, because
given by the number of teeth on the gears. So the effective gearing the result is insensitive to changes of the structure and to the
is 5:3.29. The output RPM is correctly 3000 ¥ 3.29/5 = 1974 RPM mounting position of the accelerometer.
= 32.9 Hz, which means that the fourth rahmonic of the output The cepstrum is a pure calculation of a power spectrum, mean-
fault comes out at quefrency 1/32.9 ¥ 4 ms = 121 ms. ing that the cepstrum can be used in other contexts than those
This leads to a discussion of the resolution of the analysis. It discussed here. It is applicable on an order-power spectrum
is not the intention of this article to go into operational details; it when addressing a local fault on a machine operating at varying
merely presents an understanding of the power of the cepstrum RPM. It can also be used on top of an envelope spectrum to reveal
analysis if resolution is selected appropriately. It might be worth- multiple faults that have been dug out of background noise by the
while to mention that inspection showed a faulty tooth on both envelope analysis.
gears.
Liftered Spectrum. The Fourier transformation of the Cepstrum References
is the original logarithmic spectrum: 1. R. B. Randall, Frequency Analysis, Third Edition, Brüel & Kjær, 1987.
2. R.B. Randall and J. Hee, “Cepstrum Analysis,” Brüel & Kjær Technical
{ {log (S
F {C AA (t )} = F F -1
AA ( f ))}} = log (SAA ( f )) (6) Review, No.3, 1981.
3. H. Konstantin-Hansen, “Envelope Analysis for Diagnostics of Local
By applying a weighting function to the cepstrum prior to the Faults in Rolling Element Bearings,” Brüel & Kjær Application Note,
transformation, it is possible to get spectra corresponding to differ- BO 0501, 2003.
4. www.bksv.com/Library, Library search: Cepstrum or Envelope.
ent parts of the cepstrum. In Figure 5, two weighting functions are
5. PULSE Knowledge Library; Part of PULSE Analyzer Platform.
shown. In cepstrum terminology, the red lifter is called a long-pass
lifter, and the blue lifter a short-pass lifter. The long-pass lifter The author can be reached at: hans.konstantin-hansen@bksv.com.

12  SOUND & VIBRATION/MAY 2010 www.SandV.com