See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/235743820

An investigation on gearbox fault detection using vibration analysis

techniques: A review *

Article in Australian Journal of Mechanical Engineering · November 2012

DOI: 10.7158/M11-830.2012.10.2

CITATIONS READS

56 9,517

1 author:

Amit Aherwar
Madhav Institute of Technology & Science Gwalior
61 PUBLICATIONS 953 CITATIONS

SEE PROFILE

All content following this page was uploaded by Amit Aherwar on 18 June 2017.

The user has requested enhancement of the downloaded file.

 1

An investigation on gearbox fault detection

using vibration analysis techniques: A review*

A Aherwar†
Department of Mechanical Engineering, Madhav Institute of Technology & Science, Gwalior, India

ABSTRACT: Gears are critical element in a variety of industrial applications such as machine
tool and gearboxes. An unexpected failure of the gear may cause significant economic losses. For
that reason, fault diagnosis in gears has been the subject of intensive research. Vibration analysis
has been used as a predictive maintenance procedure and as a support for machinery maintenance
decisions. As a general rule, machines do not breakdown or fail without some form of warning, which
is indicated by an increased vibration level. By measuring and analysing the machine’s vibration, it
is possible to determine both the nature and severity of the defect, and hence predict the machine’s
failure. The vibration signal of a gearbox carries the signature of the fault in the gears, and early
fault detection of the gearbox is possible by analysing the vibration signal using different signal
processing techniques. This paper presents a review of a variety of diagnosis techniques that have
had demonstrated success when applied to rotating machinery, and highlights fault detection and
identification techniques based mainly on vibration analysis approaches. The paper concludes with
a brief description of a new approach to diagnosis using neural networks, fuzzy sets, expert system
and fault diagnosis based on hybrid artificial intelligence techniques.

KEYWORDS: Condition monitoring; vibration analysis; fault diagnosis; gearbox.

REFERENCE: Aherwar, A. 2012, “An investigation on gearbox fault detection using

vibration analysis techniques: A review”, Australian Journal of Mechanical Engineering, Vol.
10, No. 2.

1 INTRODUCTION tooth faces and the time varying support stiffness

of geared system, change the ratio, ie. vibration
The ubiquitous of gears in rotating machinery has which characteristics can reflect symptoms of a lot
made the study of vibration a more interesting of faults or defects. The benefit of using vibration
subject. One study found that 65% of gear box analysis for their monitoring and diagnosis has been
damage is due to faults in the gears (Allianz demonstrated to be successful since the early time
Versicherungs-AG, 1978). In engineering, they show because of the ease of measurement. The approaches
a considerable number of different forms of damage. of gear vibration analysis are mainly subdivided
The common types of gear damage mainly consist of into three categories according to analysis domains.
pitting, scuffing, spalling, cracking and wear (Michel They are time domain, frequency domain and
& Miller, 1983). One of the major reasons for gear time-frequency domain. The time domain methods
faults is excessive vibration. Vibration can be thought include time synchronous average and statistical
of as a ratio of the forces acting on the gear to its analysis. The former is a signal averaging process
dynamic stiffness. The backlash, the error of the gear over a large number of cycles, synchronous with the
running speed of a specific shaft in the gearbox. It
transmission, the unbalanced inertia mass, the time
can remove not only the background noise but also
varying mesh stiffness of tooth, the friction between
periodic events that are not exactly synchronous
* Paper M11-830 submitted 13/11/11; accepted for publication
with the gear being monitored (McFadden, 1989;
after review and revision 5/09/12. 1991). The latter consists of many descriptive
† Corresponding author Amit Aherwar can be contacted at statistics such as sample skewness, kurtosis and so
[email protected]. on. The frequency domain methods include spectral

© Institution of Engineers Australia, 2012 Australian Journal of Mechanical Engineering, Vol 10 No 2

2 “An investigation on gearbox fault detection using vibration analysis techniques: ...” – Aherwar

analysis based ones such as power spectral density mode and methods of collecting and quantifying
and Cepstrum analysis (Randal, 1982), and higher- the evidence is needed. Although many faults may
order statistics and spectra (Zhang et al, 1999). The be easily detectable by physical examination of a
time-frequency domain methods are composed of the component, using techniques such as microscopy,
short-time Fourier transform (STFT), Wigner-Ville, x-ray, dye penetrates, magnetic rubber, etc., these
Cohen distribution and wavelet analysis (Wang & methods usually cannot be performed without
McFadden, 1993; Lai et al, 2003). Feature extraction removal of, and in some cases physical damage
methods play an important role in machine condition to, the component. While physical examination
monitoring and fault diagnosis, from which the techniques still play a critical role during manufacture,
diagnostic information can be obtained. Through assembly and overhaul, they are impractical in an
gear vibration analysis, a lot of features are acquired, operational large transmission system and other
and the next step is optimisation and classification. (non-intrusive) fault detection methods need to be
In the present work the authors present a review of employed for routine monitoring purposes. Most
a variety of diagnosis techniques for gearbox fault modern techniques for gear diagnostics are based
identification with particular regard to vibration on the analysis of vibration signals picked up from
analysis. The vibration techniques were developed the gearbox casing. The common target is to detect
with two main purposes. The first purpose is to the presence and the type of fault at an early stage
separate the gearbox related signal from other of development and to monitor its evolution, in
components and to minimise the noise that may order to estimate the machine’s residual life and
mask the gearbox signal, especially in the early stages choose an adequate plan of maintenance. It is well
of the fault. The second purpose is to identify the known that the most important components in gear
status of the gearbox, to distinguish the good and the vibration spectra are the gear meshing frequency
faulty gear and to indicate the defective components. and its harmonics, together with sidebands due to
Nowadays the demands for condition monitoring modulation phenomena. The increment in the number
and vibration analysis are no more limited trying to and amplitude of such sidebands may indicate a fault
minimise the consequences of machine failures, but condition. Moreover, the spacing of the sidebands is
to utilise existing resources more effectively. related to their source. source identification and fault
detection from vibration signals associated with items
2 FAULT DETECTION AND DIAGNOSIS which involve rotational motion such as gears, rotors
FROM VIBRATION ANALYSIS and shafts, rolling element bearings, journal bearings,
flexible couplings, and electrical machines depend
Diagnostics is understood as identification of a upon several factors: (i) the rotational speed of the
machine’s condition/faults on the basis of symptoms. items; (ii) the background noise and/or vibration
Diagnosis requires a skill in identifying machine’s level; (iii) the location of the monitoring transducer;
condition from symptoms. The term diagnosis (iv) the load sharing characteristics of the item; and
is understood here similarly as in medicine. It is (v) the dynamic interaction between the item and other
generally thought that vibration is a symptom of a items in contact with it. The main causes of mechanical
gearbox condition. Vibration generated by gearboxes vibration are unbalance, misalignment, looseness and
is complicated in its structure but gives a lot of distortion, defective bearings, gearing and coupling in
information. We may say that vibration is a signal accuracies, critical speeds, various form of resonance,
of a gearbox condition. To understand information bad drive belts, reciprocating forces, aerodynamic or
carried by vibration one have to be conscious/aware hydrodynamic forces, oil whirl, friction whirl, rotor/
of a relation between factors having influence to stator misalignments, bent rotor shafts, defective
vibration and a vibration signal. In order to detect rotor bars, and so on. Some of the most common
(and diagnosis) an impending failure, a good faults that can be detected using vibration analysis
understanding of the evidence relating to the failure are summarised in table 1.

Table 1: Some typical faults and defects that can be detected with vibration analysis.

Item Fault
Tooth messing faults, misalignment, cracked and/or worm teeth, eccentric
Gears
gear
Unbalance, bent shaft, misalignment, eccentric journals, loose components,
Rotors and shaft
rubs, critical speed, cracked shaft, blade loss, blade resonance
Rolling element bearings Pitting of race and ball/roller, spalling, other rolling elements defect
Flexible coupling Misalignment, unbalance
Unbalanced magnetic pulls, broken/damaged rotor bars, air gap geometry
Electrical machines variations, structural and foundation faults, structural resonance, piping
resonance, vortex shedding

Australian Journal of Mechanical Engineering Vol 10 No 2

 “An investigation on gearbox fault detection using vibration analysis techniques: ...” – Aherwar 3

Ebersbach et al (2005) investigated the effectiveness error. Design errors may be due to causes like
of combining both vibration analysis and wear improper gear geometry, use of wrong materials,
debris analysis is an integrated machine condition quality, lubrication and other specifications.
monitoring maintenance program. Decker (2002) Application errors can be due to problems like
proposed two new detection techniques. The time vibration, mounting and installation, cooling, and
synchronous averaging (TSA) concept was extended maintenance, while manufacturing errors can be
from revolution-based to tooth engagement-based. in the form of mistakes in machining or problems
The detection techniques are based on statistical in heat treating. Summary of safety critical failure
comparisons among the averages for the individual modes are presented in table 2. Several researchers
teeth. These techniques were applied to a series have worked on the subject of gearbox defect
of three seeded fault crack propagation tests. detection and diagnosis through vibration analysis.
Polyshchuk et al (2002) presented the development Time domain, frequency domain, time frequency
of a novel method in gear damage detection using domain based on STFT, wavelet transform and
a new gear fault detection parameter based on the advanced signal processing techniques have been
energy change in the joint time-frequency analysis implemented and tested.
of the vibration analysis of the vibration signal.
Choy et al (2003) demonstrated the use of vibration 3.1 Time domain analysis
signature analysis procedures for health monitoring
and diagnostics of a gear transmission system. Lin
& Zuo (2003) introduced an adaptive wavelet filter The time domain methods try to analyse the
based on the Morlet wavelet; the parameters in the amplitude and phase information of the vibration
Morlet wavelet function are optimised based on the time signal to detect the fault of gear-rotor-bearing
kurtosis maximisation principle. The wavelet used system. The time domain is a perceptive that feels
is adaptive because the parameters are not fixed. The natural, and provides physical insight into the
adaptive wavelet filter is found to be very effective vibration (Forrester, 1996). It is particularly useful in
in detection of symptoms from vibration signals of analysing impulsive signals from bearing and gear
a gearbox with early fatigue tooth crack. defects with non-steady and short transient impulses
(McFadden, 1987). Time domain vibration signals,
if understood properly, can yield enormous amount
3 GEARBOX FAILURE AND ITS of information. Identify some characteristics which
VIBRATION ANALYSIS TECHNIQUES are not readily observed can be highlighted by this
technique. Various time-domain techniques can be
The principle causes for gear failure are: (i) error of used in machine condition monitoring and these are
design; (ii) application error; and (iii) manufacturing as follows.

Table 2: Safety critical failure modes.

Failure Failure mode Cause Contributing factors

Unbalance
Fatigue Misalignment Coupling, bearing failure
Shaft
Bent shaft
fracture
Interference Incorrect assembly, bearing failure
Overload
Operational
Life limit exceeded
Gear Fatigue
Surface damage
fracture
Resonance Design
Life limit exceeded
Bending fatigue Surface damage Process related
Tooth Thin tooth Excessive wear, destructive scoring
fracture Random fracture Surface damage Process related, foreign object, pitting/spalling
Interference Incorrect assembly, bearing failure
Overload
Operational
Insufficient oil
Over- Lubrication
Loss of oil Oil line failure, filter bowl failure
heating
Insufficient cooling Cooling fan failure Shaft/gear fracture

Australian Journal of Mechanical Engineering Vol 10 No 2

 4 “An investigation on gearbox fault detection using vibration analysis techniques: ...” – Aherwar

3.1.1 Waveform analysis and sufficient load is applied to maintained full

contact. At higher operating speed, both the peak
Waveform analysis consists of recording the time and RMS values increase proportionally, giving a
history of the event on a storage oscilloscope or a real relatively constant crest factor. In the absence of
time analyser. Apart from an obvious fundamental significant impulsiveness, the reliability of the crest
appreciation of the signal, it is useful in the study of factor technique to detect gear damage breaks down.
non-steady conditions and short transient impulses.
Waveform analysis can also be useful in identify Peak value = Amax (1)
vibrations that are non-synchronous with shaft speed
N
and in machine coast down analysis waveform can 2

indicate the occurrence of resonance. In machine cost

¦ [A(n)]
RMS n 1
(2)
down analysis waveform can indicate the occurrence N
of resonance. The digitised data of 8202 samples is
used to transfer in time domain signature with the Here Amax = maximum amplitude value in the time
help of MATLAB algorithms of 1.934 seconds time domain; A(n) = amplitude of the nth digitised point
length. Figure 1 shows the waveform (sample time in the time domain; and N = number of points in the
record) of vibration signals of defected gearbox. time domain.
The simplest approach to measuring defects in the
3.1.2 Indices time domain is using the RMS approach. However,
the RMS level may not show appreciable changes in
Indices are also used in vibration analysis (Braun, the early stages of gear and bearing damage. A better
1986; Forrester, 1996). The peak value, root-mean- measure is to use “crest factor”, which is defined as
square (RMS) level and their ratio crest factor are the ratio of the peak level of the input signal to the
often used to quantify the time signal. The peak RMS level. Therefore, peaks in the time series signal
level is not a statistical quantity and hence may
will result in an increase in the crest factor value.
not be reliable in detecting damage continuously
This feature is used to detect changes in the signal
operating systems. The RMS value, however, is
pattern due to impulsive vibration sources such as
more-satisfactory for steady-state applications. The
tooth breakage on a gear or a defect on the outer race
crest factor, defined as the ratio of the peak value
of a bearing. The crest factor feature is not considered
to RMS level, has been proposed as a trending
a very sensitive technique. Below is the equation for
parameter as it includes both parameters. Crest
the crest factor:
factors are reliable only in the presence of significant
impulsiveness. Typical values of crest factors for Crest factor = Peak level/RMS level (3)
gear in a good condition range from 3.5 to 4.0, and
values for gear with impulsive defects are higher,
ranging up to ~10.0. Generally crest factors higher 3.13 Statistical methods
than 4.5 are indicative of damage. Crest factor
values of a vibration signal are relatively insensitive Statistical analysis can also be carried out on
to operating speed and gear load, provided that time domain data. The probability density is the
sufficient speed is maintained to generate a gear probability of finding instantaneous values within
vibration which is above the background noise level, a certain amplitude interval, divided by the size of
the interval. All signals will have a characteristic
probability density curve. These curves if derived
from machinery vibration signals can subsequently
50
be used in machine condition monitoring.
40

30
3.1.3.1 Probability density moments
Ac c eleration [g]

20 The shape of the probability density curve can be

10
described by a series of single-number indices. These
are the moments of the curve and are analogues to
0 mechanical moment about the centroid of a plane.
-10
3.1.3.2 Kurtosis (K)
-20

-30 Kurtosis, as explained in equation (4), is defined as

0 1000 2000 3000 4000 5000 6000 7000 8000
Tim e [Sec ]
the fourth moment of the distribution and measures
the relative peakedness or flatness of a distribution
Figure 1: A typical waveform of vibration as compared to a normal distribution (Braun, 1986).
for duration of 1.934 seconds of Kurtosis provides a measure of the size of the tail of
defected gearbox. distribution and is used as an indicator of major peaks

Australian Journal of Mechanical Engineering Vol 10 No 2

 “An investigation on gearbox fault detection using vibration analysis techniques: ...” – Aherwar 5

f
in a set of data. As a gear wears and breaks this feature j 2 Sft
should indicate an error due to the increased level of x(t) ³ X( f )e df (7)
f
vibration. The equation for kurtosis is given by:
N
4
In the above equations, t stands for time, f stands for
¦ ª¬ y n  P º¼
n 1
frequency and x denotes the signal at hand. Note that
K (4) x denotes the signal in time domain and the X denotes
NV 4 the signal in frequency domain. This conversion is
used to distinguish the two representations of the
Here y(n) = data (n = 1, 2, 3, ..., N); N = total number of
signal. Equation (6) is called the Fourier transform
data samples;  = mean; and  = standard deviation.
of x(t), and equation (7) is called the inverse Fourier
transform of X(t), which is x(t).
3.1.3.3 Skewness (S)
The signal x(t) is multiplied with an exponential term
Skewness is a measure of symmetry, or more e–2ft, at some certain frequency f and then integrated
precisely, the lack of symmetry. A distribution of over all times.
data set is symmetric if it looks the same to the left The exponential term e–2ft can also be written as
and right of the centre point. Equation (5) is used to equation (8):
calculate the values of skewness.
N cos(2ft) + jsin(2ft) (8)
3
¦ ¬ª y
n 1
n  P ¼º
The above expression has a real part of cosine
S (5)
NV 3 of frequency f, and an imaginary part of sine of
frequency f. In equation (6), multiplication of the
original signal with a complex expression, which
3.2 Frequency domain analysis has sins and cosines of frequency f, is done. Then
this product is integrated. In other words, addition
The frequency domain methods include fast Fourier of all the points in this product is done. If the result
transform (FFT), Hilbert Transform Method and of this integration (which is nothing but some sort of
Power Cepstrum Analysis. They use the difference infinite summation) is a large value then, the signal
of power spectral density of the signal due to the x(t), has a dominant spectral component at frequency
fault of gear and/or bearing to identify the damage f. This means that, a major portion of this signal is
of elements (Braun, 1986). Any real world signal can composed of frequency f. If the integration result is
be broken down into a combination of unique sine a small value, than this means that the signal does
waves. Every sine wave separated from the signal not have a major frequency component of f in it. If
appears as a vertical line in the frequency domain. this integration result is zero, then the signal does
Its height represents its amplitude and its position not contain the frequency f at all.
represents the frequency. The frequency domain
representation of the signal is called the signal. The 3.2.1.1 Discrete Fourier transform
frequency domain completely defines the vibration.
Frequency domain analysis not only detects the faults If the signal (function) is in discrete time form, the
in rotating machinery but also indicates the cause of discrete Fourier transform (DFT) process can be used
the defect (Forrester, 1996). to analyse it. DFT can be expressed by equation (9):
1 N 1
3.2.1 Fourier transform X(m) ¦ x(n)e  j 2Smn/N
Nn0
(9)

Fourier transform is a mathematical approach used to

express any deterministic periodic and non-periodic For m = 0, 1, 2, ..., N–1. This is the N-point DFT for
function by an infinite series of sum of periodic the time series x(n) for n = 0, 1, 2, ..., N–1, where N is
exponential functions. The purpose of using Fourier the total number of sample in a given time record,
transform is to obtain the frequency components of x(n) is the discrete time function and X(m) is the DFT
the vibration signal and to present it in frequency of the function.
domain. The Fourier transform can be expressed by The DFT algorithm is the basis of digital signal
equation (6): analysis with N2 complex multiplications required to
f establish a single N-point transform. If averaging is
 j 2 Sft
X( f ) ³ x(t)e dt (6) required over M time signals, then MN2 calculations
f are required. The FFT algorithm significantly reduces
the number of computations that are required; it is
where x(t) is the original function in time domain; X(f)
essentially a more efficient procedure for evaluating a
is the Fourier transform of the function x(t); j is the
DFT. While the DFT transform above can be applied
square root of –1; and e denotes the natural exponent.
to any complex valued series, in practice for a large
The inverse transform is: series it can take considerable time to compute, the

Australian Journal of Mechanical Engineering Vol 10 No 2

6 “An investigation on gearbox fault detection using vibration analysis techniques: ...” – Aherwar

time being proportional to the square of the number 2500

of points in the series. The much faster FFT algorithm (a)

was developed by Cooley & Tukey (1965). The only 2000

Amplitude [Relative Unit]

requirement of the most popular implementation
Fm
of this algorithm (Radix-2 Cooley-Tukey) is that 1500

the number of points in the series be a power of 2.

The computing time for the FFT is proportional to 1000

Nlog2(N). So, for example, a transform on 1024 points

using the DFT takes about 100 times longer than 500

using FFT, a significant speed increase.

0
The vibration characteristics of any rotating machine 0 200 400 600 800 1000 1200 1400 1600
Frequenc y [Hz ]
are to some extent unique, due to the various transfer
characteristics of the machine. In the FFT plot, various
large and small peaks are presented corresponding to 2Frg (b)
2000
characteristic frequencies shows the origin of defects;

Amplitude [Relative Unit]

or we can say FFT shows the frequencies in terms of
1500
shaft harmonics. For gear problems, special attention 2Frp
must be given to the FFT spectrum’s bearing defect 1000 Frg 3Frg
frequencies. The spectra of FFT may produce peaks at
identified fault frequencies. These peaks may or may 500

not represent the indicated fault. One must look for

harmonics to determine if the identified frequencies 0
0 50 100 150 200 250 300
were generated from the indicated fault: Frequenc y [Hz ]

• If peak appears at the fundamental fault

Figure 2: (a) A typical FFT spectrum of defected
frequency and another peak appears at two times
gear vibration signal. (b) Zoom view
the fundamental frequency, it is a very strong of a typical FFT spectrum of defected
indication that the fault is real. gear vibration signal.
• If no peak appears at the fundamental fault
frequency, but peaks are present at two, three of components are anticipated known as discrete
and maybe four times the fundamental fault frequency monitoring. The frequency of interest
frequency, then this also represents a strong in the detection of bearing faults is the modulating
indication that the indicated fault is valid. frequencies because the resonant frequency of
Figure 2(a) shows the spectrum of defected gear housing is high and the number of spectral lines
vibration signal. In this figure, the peaks are found typically limited, it is difficult if not possible to
at Fm and its second multiple frequencies, but there resolve sidebands. Pass band technique allow one
are some other peaks due to modulation effect of the to get a direct measure of the frequency of the
signal. In figure 2(b) the peaks are found at Frg and modulating signal. The purpose of band pass filtering
Frp, and its second and third multiple frequency, but is to reject the low frequency high amplitude signals
there are some other peaks due to modulation effect associated with unbalance, misalignment, looseness
of signal. These peaks show a strong indication that and to eliminate random noise. The filter design can
be design by defining the transfer function. Filtering
the gear is in faulty condition and the nature of fault
in the frequency domain is an operation in which
is chipped tooth pitch to the top at 2% thickness.
the frequency components to be further analysed is
FFT for determination of the severity of the fault: multiplied by unity and other frequency components
• One way to determine the fault’s severity is to are multiplied by zero. Mathematically, we can define
compare its amplitude with the past readings filter function as follows by equation (10):
taken under consistent conditions. ­1 , Z1  Z  Z2
H (Z ) ® (10)
• Another way is to compare the amplitude to the ¯0 , otherwise
other readings obtained by similar machines
running under same conditions. A higher than Therefore band pass filtering technique is employed
normal reading indicates a problem. in this section to extract the features of the defected
gearbox for further fault categorisation. There are
3.2.2 Frequency band analysis various types of filters are available, such as Bessel,
Butterworth, Chebyshev and Elliptic. The Elliptic
Pass band analysis is yet another technique of filters offer steeper roll-off characteristics than then
reducing the quantity of data made available in a Butterworth or Chebyshev filters, but are equi-ripple
spectrum to manageable proportions. The technique in both the pass and stop bands. In general, Elliptic
involves monitoring only a band of frequencies filters meet given performance specifications with
either broad or narrow, in which defect frequencies the lowest order of any filter type.

Australian Journal of Mechanical Engineering Vol 10 No 2

 “An investigation on gearbox fault detection using vibration analysis techniques: ...” – Aherwar 7

3.2.3 Spectral analysis 3.3 Order analysis

The simple spectral analysis is generally unable to Order analysis is a technique for analysing noise
detect gear failures at an early stage; for this reason, and vibration signals in rotating or reciprocating
many researchers have proposed the application machinery. Some examples of rotating or reciprocating
of other vibration analysis techniques for the machinery include aircraft and automotive engines,
early detection of fault symptoms. Spectral (or compressors, turbines, and pumps. Such machinery
frequency) analysis is a term used to describe the typically has a variety of mechanical parts such as
analysis of the frequency domain representation of a shaft, bearing, gearbox, blade, coupling and belt.
a signal. Spectral analysis is the most commonly Each mechanical part generates unique noise and
used vibration analysis technique for condition vibration patterns as the machine operates. Each
monitoring in geared transmission systems and mechanical part contributes a unique component
has proved a valuable tool for detection and basic to the overall machine noise and vibration. This
diagnosis of faults in simple rotating machinery technique is suitable for analysing noise and
(Dalpiaz et al, n.d.). Whereas the overall vibration vibration signals when the rotational speed changes
level is a measure of the vibration produced over over time. Order is defined as the normalisation of
a broad band of frequencies, the spectrum is a the rotational speed. The first order is the rotational
measure of the vibrations over a large number speed and order n is n-times the rotational speed.
of discrete contiguous narrow frequency bands. Order components are the vibration harmonics of the
The fundamental process common to all spectral rotational speed. In the case of the PC fan, the shaft
analysis techniques is the conversion of a time vibration is the first order vibration. The coil and
domain representation of the vibration signal into blade vibrations are the fourth and seventh order
a frequency domain representation. This can be vibrations, respectively. In general, order analysis
achieved by the use of narrow band filters or, more techniques relate noise and vibration signals the
commonly in recent times, using the DFT of digitised rotational speed. Order analysis techniques also
data. The vibration level at each “frequency” reduce these signals into characteristic components,
represents the vibration over a narrow frequency associate these components to mechanical parts,
band centred at the designated “frequency”, with and provide repeatable noise and vibration
a bandwidth determined by the conversion process measurements. You can obtain information about
employed. For machines operating at a known individual mechanical parts as well as the entire
constant speed, the frequencies of the vibrations machine with order analysis. When performing
produced by the various machine components vibration analysis many sound and vibration
can be estimated therefore, a change in vibration signal features are directly related to the running
level within a particular frequency band can speed of a motor or machine such as imbalance,
usually be associated with a particular machine misalignment, gear mesh, and bearing defects. Order
component. Analysis of the relative vibration analysis is a type of analysis geared specifically
levels at different frequency bands can often give towards the analysis of rotating machinery and
an indication of the nature of a fault, providing how frequencies change as the rotational speed of
some diagnostic capabilities. The frequency domain the machine changes. It resamples raw signals from
spectrum of the vibration signal reveals frequency the time domain into the angular domain, aligning
characteristics of vibrations if the frequencies of the signal with the angular position of the machine.
the impulse occurrence are close to one of the gear This negates the effect of changing frequencies on
characteristic frequencies, such as gear frequency, the FFT algorithm, which normally cannot handle
pinion frequency, gear mesh frequency, as shown in such phenomena.
equations (11) to (13). Then it may indicate a defect Rotating or reciprocating machines are comprised
related fault in the gearbox. of a variety of mechanical parts like shaft, bearing,
The gear frequency (Frg) is given by: gearbox, coupling etc. The conditions of the
mechanical parts mainly determine the overall
Frg = Rg/60 [Hz] (11) machine conditions. If the mechanical parts of the
machine are well tuned, ideally the machine will
The pinion frequency (Frp) is given by: generate very low vibrations. On the contrary, the
existing flaws in mechanical parts will usually cause
Frp = Rp/60 [Hz] (12)
the machine to vibrate significantly. Therefore, the
The tooth mesh frequency (Frm) is given by: machine vibrations can be used as good indicators
of the mechanical faults. Common mechanical faults
Frm = FrpNp [Hz] or FrgNg [Hz] (13) like imbalance, misalignment, mechanical looseness,
bearing fault, resonance, gearbox defects have
where Rg is the speed of gear [rpm], Rp is the speed of specific vibration characteristics. Table 3 summarises
pinion [rpm], Np is the number of teeth on the pinion, the relationships between the common mechanical
and Ng is the number of teeth on the gear. faults and vibration components.

Australian Journal of Mechanical Engineering Vol 10 No 2

8 “An investigation on gearbox fault detection using vibration analysis techniques: ...” – Aherwar

Table 3: Noise or vibration characteristics of mechanical faults.

S. No. Mechanical faults Vibration components

1 Imbalance 1x component
2 Misalignment 1x and 2x components.
3 Mechanical looseness Harmonics of 1x and 0.5x components
High vibration amplitude and large phase change at certain
4 Resonance
speed range
Gear mesh nx components (n is the number of gear teeth),
5 Gear defect
usually modulated by rotational speed components
Non-synchronous vibration components, usually
6 Rolling-element bearing defect
modulated by rotational speed components
Note: 1x means first-order component and nx means nth order component.

Figure 3: Common order analysis application process.

The noise and vibration components of the possible if it were a time domain vibration signal for one
mechanical faults are usually the most important revolution of an individual, isolated shaft with
information of the machine. The key tasks in noise attached gears. TSA is a fundamentally different
and vibration analysis in most applications are to process than the usual spectrum averaging that is
extract the patterns from the noise and vibrations generally used in FFT analysis. While the concept
signals and evaluate the condition of the mechanical is similar, TSA results in a time domain signal with
parts with the patterns. The order analysis can help lower noise than would result with a single sample.
you to get those characteristic components. An FFT can then be computed from the averaged
A common order analysis application is usually time signal.
comprised of five steps: The fundamental principle behind synchronous
1. acquire noise or vibration signals and tachometer signal averaging is that all vibration related to
signal a shaft, and the gears on that shaft, will repeat
2. pre-process the noise or vibration signals periodically with the shaft rotation. By dividing the
vibration signal into contiguous segments, each being
3. process the tachometer signal to get the rotational
exactly one shaft period in length, and ensemble
speed profile
averaging a sufficiently large number of segments,
4. perform order analysis with the noise or vibration the vibration which is periodic with shaft rotation
signals and speed profile will be reinforced and vibrations which are not
5. display the analysis results in different formats. periodic with the shaft rotation will tend to cancel
out; leaving a signal which represents the average
3.4 Time synchronous averaging vibration for one revolution of the shaft. Figure 4
illustrates how this process might be performed on
Stewart (1977) showed that with TSA the complex a continuous time signal from a gearbox, using a
time-domain vibration signal from a transmission tacho multiplier to calculate each rotational period of
could be reduced to estimates of the vibration for the shaft. The process illustrated in figure 4 assumes
individual shafts and their associated gears. The the vibration signal being averaged is a continuous
synchronous average for a shaft is then treated as time signal. In practice, the signal averaging process

Australian Journal of Mechanical Engineering Vol 10 No 2

 “An investigation on gearbox fault detection using vibration analysis techniques: ...” – Aherwar 9

Figure 4: Synchronous signals averaging.

usually takes place on a discrete sampled signal The signal is sampled using a trigger that is
(eg. via an analogue-to-digital converter in a PC) synchronised with the signal. The averaging process
and, in addition to defining the start and end points gradually eliminates random noise because the
of the shaft rotation, some mechanism is needed to random noise is not coherent with the trigger. Only
ensure that the sample points are at equally spaced the signal that is synchronous and coherent with the
angular increments of the shaft and that these are trigger will persist in the averaged calculation, as
at the same angular position for each revolution of shown below. Traditional spectrum-based averaging
the shaft. That is, the sampling must be coherent records a frame of data in the time domain, computes
with the rotation of the shaft. Originally, Stewart the FFT and then adds the FFT spectrum to the
(1977) used a phase-locked frequency multiplier, averaged spectrum. The time signal is discarded
however, McFadden (1986c) showed that far greater and then the process is repeated until the averaging
accuracy and flexibility could be achieved using number is complete. The result is a spectrum with very
digit resampling of time-sampled vibration based low noise, but if you examine each time record that
on a reference derived from a simultaneously time is used to compute the FFT spectra, each time record
sampled tacho signal. will include the signal of interest plus random noise
because the averaging is performed in the frequency
The principle of synchronising the averaging with
domain, not the time domain. Another important
some other process (in this case the rotational
application of TSA is in the waveform analysis of
frequency of a shaft) is fundamental to the technique;
machine vibration, especially in the case of gear drives.
whether it is per formed on continuous or discrete
In this case, the trigger is derived from a tachometer
signals. As was seen above, when the process is
that provides one pulse per revolution of a gear in the
performed on discrete signals the sampling must
machine. This way, the time samples are synchronised
be coherent with the rotation of the shaft (hence the
in that they all begin at the same exact point related
term “coherent rotational signal averaging” used
to the angular position of the gear. After performing
by Swansson et al). Note that the process can be
a sufficient number of averages, spectrum peaks that
performed in the time or frequency domain (as long
are harmonics of the gear rotating speed will remain
as the frequency domain averaging is performed while non-synchronous peaks will be averaged out
on the complex frequency domain representation). from the spectrum. Two kinds of time synchronous
The term “time domain averaging” (McFadden, average: time synchronous linear average and time
1986b; Braun, 1975) was used to distinguish the synchronous exponential average.
technique from that of averaging of amplitude or
power spectra to reduce variance in spectral analysis For time synchronous linear average, the spectrum
(Randall, 1987). To properly describe the process will stop updating when the average number is
when applied to discrete signals, it should probably reached. The nth frame of the spectrum is calculated
be referred to as “rotationally coherent synchronous from An.
signal averaging”. However, this is quite clumsy and An = (An–1(n – 1) + Tn)/n (14)
therefore the technique will normally be referred to
as “synchronous signal averaging” (the rotational for n = 1 ~ N and A1 = T1, where An = nth average of
coherency being implied when the technique is the time block signal, Tn = nth frame of the time block
applied to discrete data). signal, and N = average number given.

Australian Journal of Mechanical Engineering Vol 10 No 2

10 “An investigation on gearbox fault detection using vibration analysis techniques: ...” – Aherwar

When the average number N is reached, the averaged FM4 is non-dimensional and designed to have a
time block signal is: nominal value of 3 if d is purely Gaussian. When
higher-order sidebands appear in the vibration
AN = (AN–1(N – 1) + TN)/N signal, FM4 will deviate from this value.
(15)
= (A1 + A2 + A3 + … + AN–1 + AN)/N

The averaged spectrum is calculated from AN. 3.4.3 NA4

For time synchronous exponential average, the The parameter NA4 was developed by Zakrajsek
spectrum keeps updating and never stops. The (1994) as a general fault indicator that reacts not
averaged time block signal is: only to the onset of damage as FM4 does, but also
to the continuing growth of the fault. The residual
Acur = (1 – )Apre + Tcur (16) signal r, given in equation (19), is first constructed.
The quasi-normalised kurtosis of the residual signal
The averaged spectrum is calculated from Acur, which
is then computed by dividing the fourth moment
is the current average of the time block signal.  =
of the residual signal by the square of its run time
1/N = inverse of the average number N, Tcur = current
averaged variance. The run time averaged variance
frame of the time block signal, and Apre = previous
is the average of the residual signal over each time
average of the time block signal.
signal in the run ensemble up to the point at which
Stewart (1977) developed a number of non- NA4 is currently being calculated. NA4 is given as:
dimensional parameters based on the synchronous N
N ¦ i 1 (riM  rM )4
signal average, which he termed “Figures of Merit” NA 4( M ) 2 (19)
(McFadden, 1986b). These were originally defined as ­1 M N 2 ½
® ¦ j 1 ª¬ ¦ i 1 (rij  rj ) º¼¾
a hierarchical group, with which Stewart described ¯M ¿
a procedure for the detection and partial diagnosis
of faults. where r is the mean of the residual signal, N is the
total number of data points in the time signal, M is
the number of the current time signal, and j is the
3.4.1 FM0
index of the time signal in the run ensemble. Like
FM4, NA4 is non-dimensional and designed to have
The parameter FM0 was developed by Stewart (1977)
a nominal value of 3 if r is purely Gaussian.
as a robust indicator of major faults in a gear mesh.
Major changes in the meshing pattern are detected by
comparing the maximum peak-to-peak amplitude of 3.4.4 M6A
the signal to the sum of the amplitudes of the mesh The parameter M6A was proposed by Martin (1989)
frequencies and their harmonics. FM0 is given as: as an indicator of surface damage on machinery
PPx components. The underlying theory is the same as
FM 0 H (17) that of FM4. However, it is expected that M6A will
¦ n 0
Pn
be more sensitive to peaks in the difference signal
due to the use of the sixth moment. M6A is given as:
where PPx is the maximum peak-to-peak amplitude
N
of the signal x; Pn is the amplitude of the nth harmonic; N 2 ¦ i 1 (di  d )6
and H is the total number of harmonics in the M6 A 3 (20)
ª ¦ N (di  d )2 º
frequency range. Notice that in cases where PPx ¬ i1 ¼
increases while Pn remains relatively constant, FM0
increases. Also, if Pn decreases while PPx remains Note that in this case, the moment is normalised by
constant, FM0 also increases. the cube of the variance.

3.4.2 FM4 3.4.5 M8A

The parameter M8A, also proposed by Martin (1989),

Developed by Stewart (1977), the parameter FM4 is designed to be yet more sensitive than M6A to
was designed to complement FM0 by detecting faults peaks in the difference signal. M8A uses the eighth
isolated to only a limited number of teeth. This is moment normalised by the variance to the fourth
accomplished by first constructing the difference power and is given as:
signal, d, given in equation (18). The normalised N
kurtosis of d is then computed. FM4 is given as: N 3 ¦ i 1 (di  d )8
M8 A 4 (21)
N ª ¦ N (di  d )2 º
N ¦ i 1 (di  d )4 ¬ i1 ¼
FM 4 2 (18)
ª ¦ N (di  d )2 º
¬ i1 ¼ 3.4.6 NB4

where d is the mean of the difference signal and N The parameter NB4 was developed by Zakrajsek
is the total number of data points in the time signal. (1994) as an indicator of localised gear tooth damage.

Australian Journal of Mechanical Engineering Vol 10 No 2

 “An investigation on gearbox fault detection using vibration analysis techniques: ...” – Aherwar 11

The theory behind NB4 is that damage on just a few separate approximations of the amplitude and phase
teeth will cause transient load fluctuations different modulation functions and that these approximations
from those load fluctuations caused by healthy could subsequently be inspected to find early
teeth, and that this can be seen in the envelope of the indications of gear damage (McFadden, 1986a;
signal. As with NA4, NB4 uses the quasi-normalised Cempel & Staszewski, 1992). This work was further
kurtosis. However, instead of the difference signal, refined by Blunt & Forrester (1985) to produce a
NB4 uses the envelope of the signal band-pass useful damage indicator referred to as a bulls-eye
filtered about the mesh frequency. The envelope, plot which indicates both amplitude and phase
s, is computed using the Hilbert transform and is demodulations simultaneously.
given by:

s(t) ª¬b(t)  i > H(b(t))@º¼ (22) 4 ADVANCED SIGNAL PROCESSING

TECHNIQUES IN VIBRATION ANALYSIS
where b(t) is the signal band-pass filtered about the
mesh frequency, H(b(t)) is the Hilbert transform of An overall schema for intellectual diagnostics is
b(t), and i is the sample index. presented in figure 5. Intelligent diagnosis begins
with the act of data collection which is followed
by feature extraction usually employing the
3.4.7 NA4* frequency spectra. Feature extraction techniques
are widespread and can range from statistical to
The parameter NA4* was developed in by Decker
model-based techniques and comprises a variety of
et al (1994) as an enhancement to NA4. In this
signal processing algorithms which includes wavelet
case, the denominator of NA4 is statistically
transforms. Fault detection and identification is
modified, ie. when the variance of the residual signal
a subsequent step and is further classified in this
exceeds a certain statistically determined value, the
review into the four categories shown in the figure.
averaging stops and the denominator is locked. This
These will now be treated separately.
modification was made based on the observation that
as damage progresses from localised to distributed, With the development of soft computing techniques
the variance of the signal increases significantly, such as artificial neural network (ANN) and fuzzy
causing the kurtosis to settle back to nominal values logic, there is a growing interest in applying these
after the initial indication of the onset of damage. approaches to the different areas of engineering.
By normalising the fourth moment by the variance ANNs have become the outstanding method in the
of a baseline signal from the transmission operating recent decades exploiting their non-linear pattern
under nominal conditions, NA4* is provided with classification properties, offering advantages for
enhanced trending capabilities. Since it was observed automatic detection and identification of gearbox
that the variance of a damaged transmission signal failure conditions, whereas they do not require
is greater than that of a healthy transmission signal, an in-depth knowledge of the behaviour of the
the decision to lock the denominator is made based system. Recent systems have relied on artificial
on an upper limit, L, given by: intelligence techniques to strengthen the robustness
of diagnostics systems. Four artificial techniques
Z have been widely applied as expert system, neural
L v V (23)
N networks, fuzzy logic, and model-based systems
(McFadden, 1986a). Different kinds of artificial
where v is the mean value of previous variances, intelligence method have become common in fault
Z is the probability coefficient usually chosen for diagnosis and condition monitoring. For example,
a normal distribution,  is the standard deviation fuzzy logic and neural networks have been used
of the previous variances, and N is the number of in modelling and decision making in diagnostics
samples. For a normal distribution, Z can be found in schemes. Neural networks-based classifications
any introductory statistics text. However, the actual are used in diagnosis of gearbox. Rafiee et al (2007)
choice of Z should be made based on experimentation proposed fault detection and identification of
as too small a value could lead to an overabundance gearboxes using a new feature vector extracted from
of false alarms.

3.4.8 Demodulation Data Feature Fault detection

Collection Extraction and identification
The original observation made by Stewart (1977)
that gear tooth damage causes an increase in the
amplitude of the sidebands about the regular
meshing components led to further investigations Neural Fuzzy sets Expert Hybrid AI
into the nature of the amplitude and phase Networks Systems Technique
modulation functions. It was proposed that the
vibration signal could be demodulated to obtain Figure 5: Intelligent fault diagnosis.

Australian Journal of Mechanical Engineering Vol 10 No 2

12 “An investigation on gearbox fault detection using vibration analysis techniques: ...” – Aherwar

standard deviation of wavelet packet coefficients the classical modelling and control methods (Liu &
of vibration signals of various faultless and faulty Shi, 2001).
conditions of a gearbox using ANN. Over and above
the structure of ANN, an appropriate feature vector
plays a vital role in training a high performance 5 CONCLUSIONS
ANN. Ultimately a multi-layer perceptron network
with a 16:20:5 structure has been used that not only This paper has presented a brief review of some
is small in size but also with a 100% perfect accuracy current vibration-based techniques used for condition
and performance to identify gear failures and detect monitoring in geared transmission systems. After
bearing defects (Rafiee et al, 2007). the review of literature on gear fault analysis, the
following points are concluded:
ANN-based research to carry out the task can • Gearbox vibration signals are usually periodic
be categorised into two distinct groups: fault and noisy. Time-frequency domain average
identification systems with low efficiency, which was technique successfully removes the noise from the
presented by Kazlas et al (1993) to recognise gears signal and captures the dynamics of one period
and bearings failures of a helicopter gearbox; and of the signals.
fault detection systems with high efficiency, which
was illustrated by Samanta & Al-Balushi (2003) to • Time domain techniques for vibration signal
detect roller-bearing elements defects. Precisely analysis as waveform generation, indices
speaking, fault identification proves effective in (RMS value, peak level value and crest factor)
the case of particular fault classification systems, and overall vibration level do not provide any
whereas this may be in conflict with a situation diagnostic information, but may have limited
that there is a requirement to a comprehensive fault application in fault detection in simple safety
detection system to provide accordingly precision critical accessory components. The statistical
and promptness. The objective of this research was to moment as kurtosis is capable to identify the fault
develop an ANN-based system with high efficiency condition but skewness trend has not shown any
and the lowest erroneous outcome to identify faulty effective fault categorisation ability in this present
gears and detect faulty bearing of a gearbox, which gear fault condition.
has a lot of applications for preventing from fatal • Spectral analysis may be useful in the detection
breakdowns in rotary machineries. Zhenya et al and diagnosis of shaft faults.
(1992) proposed a multilayer feed-forward network- • In frequency domain, FFT was able to show the
based machine state identification method. They impulses at fault characteristics frequencies and
represented certain fuzzy relationships between the its multiple frequencies but other peaks are also
fault symptoms and causes, with highly non-linearity presents due to signal modulation effect. By
between the input and the output of the network this technique identification of fault categories
(Zhenya et al, 1992). Fuzzy logic-based fault diagnosis is difficult.
methods have the advantages of embedded linguistic • In band pass analysis of gear vibration signals it
knowledge and approximate reasoning capability. is found that the technique is feasible for feature
The fuzzy logic proposed by Zadeh (1965) performs extraction for fault diagnostics. It has been
well at qualitative description of knowledge. concluded that RMS value of filtered signal in
However, the design of such a system depends three frequency bands can be valuable feature to
heavily on the intuitive experience acquired from develop intelligent system with the use of TSA.
practicing operators thus resulting in subjectivity of
• Synchronous signal averaging has the potential of
diagnosed faults. The fuzzy membership function
greatly simplifying the diagnosis of shaft and gear
and fuzzy rules cannot be guaranteed to be optimal
faults (ie. the safety critical failures) by providing
in any case. Furthermore, fuzzy logic systems lack
significant attenuation of non-synchronous
the ability of self-learning, which is compulsory in
vibrations and signals on which ideal filtering
some highly demanding real-time fault diagnosis
can be used. Further development needs to done
cases (Hu et al, 2000). Rough set-based intelligence
on the implementation of synchronous averaging
diagnostic systems have been constructed and used
techniques and the analysis of results.
in diagnosing valves in three-cylinder reciprocating
pumps (Liu & Shi, 2001) and turbo generators (Hu • Expert system based on ANN and fuzzy logic can
et al, 2000). be developed for robust fault categorisation with
the use of extracted features from vibration signal.
Intelligent systems cover a wide range of techniques
• The results further show that the waveform
related to hard science, such as modelling and
generation in case of multiple faults at gear
control theory, and soft science, such as the artificial
contact surfaces is only useful to find the healthy
intelligence. Intellectual systems, including neural
or faulty condition but not capable to identify the
networks, fuzzy logic, and hybrid techniques,
categories of fault.
utilise the concepts of biological systems and human
cognitive capabilities. These three systems have been These conclusions motivate further research to
recognised as a robust and alternative to some of incorporate other parameters and symptoms with

Australian Journal of Mechanical Engineering Vol 10 No 2

 “An investigation on gearbox fault detection using vibration analysis techniques: ...” – Aherwar 13

vibration features to develop more robust expert Decker, H. J., Handschuh, R. F. & Zakrajsek, J. J.
systems for diagnose the problem of gear faults 1994, “An enhancement to the NA4 gear vibration
signature analysis. It has been shown that using diagnostic parameter”, Technical Report NASA
these ways of vibration signal analysis there are TM-106553, ARL-TR-389, NASA and the US Army
possibilities to detect signal faults and distributed Research Laboratory, July.
faults in gearboxes. A signal fault is caused by a tooth
crack/fracture and breakage, a spall in a gearing or in Ebersbach et al, 2005, “The investigation of the
an inner or outer race of a bearing, a spall on a rolling condition and faults of a spur gearbox using vibration
element of a bearing; distributed faults are caused by and wear debris analysis techniques”, International
uneven wear (pitting, scuffing, abrasion, erosion). Conference on Wear of Materials, Vol. 260, pp. 16-24.

Forrester, B. D. 1996, “Advanced Vibration Analysis

ACKNOWLEDGEMENTS Techniques for Fault Detection and Diagnosis
in Geared Transmission Systems”, PhD Thesis,
The author is pleased to acknowledge the support of Swinburne University of Technology, Australia.
Madhav Institute of Technology and Science (MITS),
Gwalior, India, for providing the facility of literature Gao, X. Z. & Ovaska, S. J. 2001, “Soft computing
review. A special thanks to Prof Pratesh Jayaswal, methods in motor fault diagnosis”, Applied Soft
Reader, Mechanical Engineering Department, MITS Computing, Vol. 1, No. 1, pp. 73-81.
Gwalior, India, for the motivation and guidance
throughout the study. Hu, T., Lu, B.C. & Chen, G. J. 2000, “A Rotary
Machinery Fault Diagnosis Approach Based on
Rough Set Theory”, 3rd World Congress on Intelligent
REFERENCES Control and Automation, Hefei, China, pp. 589-685.
Allianz Versicherungs-AG, 1978, Handbook of Loss Kazlas, P. T., Monsen, P. T. & LeBlanc, M. J. 1993,
Prevention, Springer-Verlag, Berlin. “Neural network-based helicopter gearbox health
monitoring system”, Neural Networks for Signal
Blunt, D. M. & Forrester, B. D. 1995, “Health Processing, Proceedings of the IEEE-SP Workshop, 6-9
monitoring of blackhawk and seahawk main September, pp. 431-440.
transmissions using vibration analysis”, Proceedings
of the Sixth Australian Aeronautical Conference/Second Lai, W., Xuan, J., Shi, T. & Yang, S. 2003, “Research
Pacific International Conference on Aerospace Science and of Vigner-Ville Time Frequency and Application in
Technology, Melbourne, Australia. Detecting Gear Pinion Fault”, Journal of Vibration
Engineering, Vol. 16, pp. 247-250.
Braun, S. 1975, “The Extraction of periodic waveforms
by Time domain Averaging”, Acustica, Vol. 32, pp. Lin, J. & Zuo, M. J. 2003, “Gear box fault diagnosis
69-77. using adaptive wavelet filter”, Mechanical Systems
and Signal Processing, Vol. 17, No. 6, pp. 1259-1269.
Braun, S. (editor), 1986, Mechanical Signature Analysis
– Theory and applications, Academic Press Inc., Liu, S. & Shi, W. 2001, “Rough set based intelligence
London, UK. diagnostic system for valves in reciprocating
pumps”, IEEE International Conference on Systems,
Cempel, C. & Staszewski, W. J. 1992, “Signal Man, and Cybernetics, Dept. of Mech. Eng., Daqing
demodulation techniques in vibroacoustical Pet. Inst., Heilongjiang, China, Vol. 1, pp. 353-358.
diagnostics of machinery”, Machine Dynamics
Problems, Vol. 5, pp. 161-173. Martin, H. R. 1989, “Statistical moment analysis as
a means of surface damage detection”, Proceedings
Choy, F. K. et al 2003, “Damage Identification of a of the Seventh International Modal Analysis Conference,
Gear Transmission Using Vibration Signatures”, Society for Experimental Mechanics, Schenectady,
Journal of Mechanical Design, Vol. 125, pp. 394-403. NY, pp. 1016-1021.

Cooley, J. W. & Tukey, J. W. 1965, “An Algorithm for McFadden, P. D. 1986a, “Detecting fatigue cracks in
the Machine Calculation of Complex Fourier Series”, gears by amplitude and phase demodulation of the
Mathematics of Computing, Vol. 19, pp. 297-301. meshing vibration”, Journal of Vibration, Acoustics,
Stress, and Reliability in Design, Vol. 108, pp. 165-170.
Dalpiaz, G. et al n.d., “Gear fault monitoring
comparison of vibration analysis techniques”, Italian McFadden, P. D. 1986b, “A model for the extraction
National Research Council, pp. 623-632. of periodic waveforms by time domain averaging”,
Aero Propulsion Technical Memorandum 437,
Decker, H. J. 2002, “Gear crack detection using tooth Department of Defence, Aeronautical Research
analysis at NASA Research Center”. Laboratory, March.

Australian Journal of Mechanical Engineering Vol 10 No 2

14 “An investigation on gearbox fault detection using vibration analysis techniques: ...” – Aherwar

McFadden, P. D. 1986c, “Interpolation techniques for Randall, R. B. 1987, Frequency Analysis, 3rd edition,
the time domain averaging of vibration data with Brüel and Kjær.
application to helicopter gearbox monitoring”, Aero
Propulsion Technical Memorandum 437, Department of Samanta, B. & Al-Balushi, K. R. 2003, “Artificial
Defence, Aeronautical Research Laboratory, September. neural network based fault diagnostics of rolling
element bearings using time-domain features”,
McFadden, P. D. 1987, “A revised model for the Mech. Systems and Signal Processing, Vol. 17, No. 2,
extraction of periodic waveforms by time domain pp. 317-328.
averaging”, Mechanical Systems and Signal Processing,
Vol. 1, No. 1, pp. 83-95. Stewart, R. M. 1977, “Some useful analysis
techniques for gearbox diagnostics”, Technical
McFadden, P. D. 1989, “Interpolation Techniques Report MHM/R/10/77, Machine Health Monitoring
for Time Domain Averaging of Gear Vibration”, Group, Institute of Sound and Vibration Research,
Mechanical Systems and Signal Processing, Vol. 3, pp. University of Southampton, July.
87-97.
Wang, W. J. & McFadden, P. D. 1993, “Early Detection
McFadden, P. D. 1991, “A Technique for Calculating of Gear Failure by Vibration Analysis I. Calculation
the Time Domain Averages of the Vibration of the of the Time-Frequency Distribution”, Mechanical
Individual Planet Gears and Sun Gear in an Epicyclic Systems and Signal Processing, Vol. 7, pp. 82-192.
Gearbox”, Journal of Sound and Vibration, Vol. 144,
pp. 163-172. Zadeh, L. A. 1965, “Fuzzy sets”, Information and
Control, Vol. 8, pp. 338-353.
Michel, A. N. & Miller, R. K. 1983, Gears and Their
Vibration, Marcel Dekker, Inc.
Zakrajsek, J. J. 1994, “A review of transmission
Oppenheim, A. V. & Schafer, R. W. 1975, Digital Signal diagnostics research at NASA Lewis Research
Processing, Prentice Hall International. Center”, Technical Report NASA TM-106746,
ARL-TR-599, NASA and the US Army Research
Polyshchuk, V. V. et al 2002, “Gear fault detection with Laboratory, December.
time frequency based parameter NP4”, International
Journal of Rotating Machinery, Vol. 8, No. 1, pp. 57-70. Zhang, G., Shi, T., Xuan, J. & Yang, S. 1999,
“Classification of Gear Faults Using RBF Network
Rafiee, J. et al 2007, “Intelligent condition monitoring Combined with Higher-order Statistics”, China
of a gearbox using artificial neural network”, Mechanical Engineering, Vol. 11, pp. 1250-1252.
Mechanical Systems and Signal Processing, Vol. 21, pp.
1746-1754. Zhenya, H., Meng, W. & Bi, G. 1992, “Neural network
and its application on machinery fault diagnosis”,
Randal, R. B. 1982, “A New Method of Modeling Proceedings of the IEEE International Conference on
Gear Faults”, Journal of Mechanical Design, Vol. 104, Systems Engineering, Kobe, Japan, September, pp.
pp. 259-267. 576-579.

Australian Journal of Mechanical Engineering Vol 10 No 2

 “An investigation on gearbox fault detection using vibration analysis techniques: ...” – Aherwar 15

AMIT AHERWAR

Amit Aherwar received his BE degree in Mechanical Engineering and M-Tech

degree in Production Engineering from the Rajiv Gandhi Technical University,
Bhopal, Madhya Pradesh, India, in 2006 and 2010, respectively. He is currently
an Assistant Professor in the Department of Mechanical Engineering, Anand
Engineering College (SGI) Agra (UP), India. His current research interests are
condition monitoring, machine fault signature analysis, and application of AI
and fuzzy techniques in mechanical systems.

Australian Journal of Mechanical Engineering Vol 10 No 2

 16 “An investigation on gearbox fault detection using vibration analysis techniques: ...” – Aherwar

Australian Journal of Mechanical Engineering Vol 10 No 2

View publication stats