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

net/publication/254478369

USING SHORT-TIME FOURIER TRANSFORM IN MACHINERY DIAGNOSIS

Article · January 2005

CITATIONS READS

19 308

3 authors, including:

A. A. Lakis Marc Thomas

Polytechnique Montréal École de Technologie Supérieure
174 PUBLICATIONS   1,743 CITATIONS    187 PUBLICATIONS   1,822 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Developing brain-computer interaction - Optimize the multi-focal studies on human emotions in relation to machines View project

Motor current signal analysis View project

All content following this page was uploaded by Marc Thomas on 13 May 2014.

The user has requested enhancement of the downloaded file.

 USING SHORT-TIME FOURIER TRANSFORM
IN MACHINERY DIAGNOSIS

M.S. Safizadeh, A.A. Lakis and M. Thomas

Department of Mechanical Engineering, École polytechnique de Montréal.
C.P. 6079, Succ. Centre-Ville, Montréal, Québec, Canada H3C 3A7

Abstract value the alarm rings and the machine stops.

Diagnostic of machine using the time- But to day, our objective is no only to protect
frequency representation of vibration is the machine but also to detect and identify
becoming widely used by many companies defaults in the first step to have the necessary
and experts. In recent years, the preventive time to schedule repairs with minimum
maintenance is no more desirable and in fact it disruption to operations and production [1].
has evolved into the prediction maintenance. Then, the key factor of the prediction
The current methods of machine diagnosis can maintenance is diagnostic. A diagnosis is not
not provide detailed diagnostics and condition an assumption; it is a conclusion reached after
prediction. a logical evaluation of the observed
This paper proposes the application of Short- symptoms.
Time Fourier Transform (STFT) as a time- In practice, different strategies of maintenance
frequency method, which can provide more can be observed:
information about a signal in time and in a) Conditional preventive maintenance: an
frequency and gives a better representation of inspection is performed if the surveillance
the signal than the conventional methods in vibration values go beyond the limit values.
machinery diagnosis. b) Systematic preventive maintenance:
In this paper, we review the machine diagnosis inspection is periodically performed but the
techniques based on the verification of inspection date depends on the old of machine
classical vibration parameters. Then the and the life length of the elements of
necessity of using time-frequency analysis in machines.
machinery diagnostics is discussed. Finally, c) Prediction maintenance: an inspection
the theory of the Short-Time Fourier would have been performed if a defect has
Transform is briefly studied and its advantages already been detected.
are shown by some practical examples. There are several conventional methods,
which have been applied for a long time to
1. Introduction fault detection and identification. Some of
The objective of diagnostic of machine by these methods provide a representation of
vibration analysis has been considerably signals in time domain and others provide a
changed. The initial objective was the security representation of signals in frequency domain.
of machine against the important damages. If In all of these methods, it is assumed that
the vibration amplitude (displacement, signals are stationary but this assumption is
velocity or acceleration) reaches to the limit not always true. In certain machines when

1
defects begin, vibration signals become non- indicator values of the signal spectrum in large
stationary and in this case, the conventional bands and in narrow bands is very unreliable.
methods are not applicable. In recent years, a One reason for this is that it is necessary to
number of new analysis methods have been define a large number of indicators
developed in the field of signal processing corresponding to a small number of defects.
called joint time-frequency analysis methods. In addition, we need to take into consideration
But there is a little tendency to use these not only the increase in the power of the
methods in the field of machinery diagnostics. signal, but also the development of its form.
Mechanical engineers are not well familiar An alternative techniques have been also
with these methods and some works are applied to verifying the variation in the form
necessary to show the true potential of these of a signal such as Cepstrum [5] (the inverse
new methods. The research into the Fourier transform of the logarithmic spectrum
development of the theory of joint time- of the signal) and the envelop method [3][6]
frequency methods and other non stationary (Hilbert transform) of the narrow band of the
signal processing methods are fast progressing signal.
but it seems that some works are needed to
motivate the industrial people and show them 3. Is the time-frequency analysis necessary?
the capability of these new methods in As mentioned in the last section, each of the
condition monitoring of mechanical systems. conventional vibration methods used in fault
The objective of this work is on the one hand, detection and identification for steady speed
to express the limitations of the conventional machines has several limitations. The
methods and on the other hand, to demonstrate assumption of constant speed in the above
the rapidity and the accuracy which can be methods results in stationary and pseudo-
obtained by using the joint time-frequency stationary vibration signals. However, even if
analysis methods in field of machinery we take this assumption into account, the
diagnostics. limitations of the above methods reduce their
performance. On the other hand, there are
2. Conventional Vibration Analysis presently several types of varying and variable
Techniques speed rotating machinery for which the
Overall level measurement [2] is the most stationary or pseudo-stationary vibration
common vibration measurement in use. It is a signals cannot be assumed. The time-
simple and inexpensive type of measurement frequency analysis not only enables us to
to undertake. There are charts available which represent the signal in three dimensions (time-
indicate the levels deemed acceptable, for frequency-amplitude) but also permit us:
example VDI 2056. The greatest limitation is
the lack of sensitivity and information a) To detect and follow the development of
available in the data. the defects which generate weak vibration
Great many indicators have been also power. However, the weak vibration power
developed for machine condition monitoring can modify the form of the signal to a
and fault detection, such as crest factor [3] considerable extent, as happens when defects
and Kurtosis [4]. Machine monitoring and produce the amplitude modulation or
diagnosis on the basis of variations in the
frequency modulation of certain characteristic components. Examples of this are the journal

2
 2
bearing of a shaft with a slow or very slow 1 ∞
P(t , ω ) = S t (ω ) = e − jωτ s (τ )h(τ − t )dτ
2
( 2)
rotational velocity, a rotating oven, dryer 2π −∞

cylinders, the press sections of a paper

machine, etc.; For each different time we get a different
b) To supervise the installations in which the spectrum and the ensemble of these spectra
normal functional process produces high provide the time-frequency distribution
amplitude periodic shocks (piston or screw P(t , ω ) .
compressor, reciprocating machinery and cam Resolutions in time and frequency will be
mechanisms, ...) which may mask the faulty determined by the width of window h(τ ) . A
frequency producing the impulsive forces.
large window width is chosen when we need
(fault in a bearing, coupling, ...).
greater accuracy in frequency and a small
The time-frequency methods are regarded as
window width when we want to have greater
advanced diagnostic techniques, which offer
accuracy in time. However, the STFT depends
high sensitivity to faults and a good diagnostic
greatly on the width of the window and by
capability.
varying the window used, one can exchange
resolution in time for resolution in frequency.
4. Short-Time Fourier Transform
The short-time Fourier transform (STFT) was
5. Industrial application of STFT in
the first time-frequency method, which was
mechanical systems
applied by Gabor [7] in 1946 to speech
In the last few decades, many methods of
communication.
time-frequency analysis have been applied to
The STFT may be considered a method that
various areas of physics and engineering, such
breaks down the non-stationary signal into
as speech processing and image processing. In
many small segments, which can be assumed
the field of machinery diagnostics, Forrester
to be locally stationary, and applies the
[8] has used time-frequency methods in the
conventional FFT to these segments.
detection of damaged gears in helicopter
The STFT of a signal s t (τ ) is achieved by
gearboxes. He has shown that, with the signal
multiplying the signal by a window function, enhancement techniques (conventional
h(τ ) , centered at At@, to produce a modified
methods) offered by Stewart and McFadden, it
signal. Since the modified signal emphasises is difficult to distinguish one type of fault, e.g.
the signal around time At@, Fourier tooth-cracking or pitting, from another; but the
Transforms will reflect the distribution of time-frequency methods can more accurately
frequency around that time. reveal the type of defect. Wang and McFadden
1 ∞
S t (ω ) = e − jωτ s (τ )h(τ − t )dτ (1) [9] have also studied the application of time-
2π −∞
frequency analysis to the detection of gear
damage. They have proven that application of
We may consider S t (ω ) as the sum of the the Spectrogram (STFT) for the early
Fourier base functions but the base functions detection of damage in gears has some
are a modulated version of the window advantages over the application of other time-
function. frequency methods.
The energy density spectrum at time At@ may In another work, Rohrbaugh [10] has applied
be written as follows: time-frequency analysis to several sets of
marine machinery. He compared the

3
Spectrogram (STFT) with cone-kernel time- easiest and the fastest method.
frequency representation from Cohen class This work is an attempt to present the
distributions. He showed that, while the limitations of conventional methods of
Spectrogram can reveal the general time- vibration analysis in machine diagnosis and to
varying characteristics of a vibration signal, emphasise the application of the STFT to fault
for more information about the signal we must detection and identification.
consider other time-frequency methods. Some A user-friendly software has also been
applications of time-frequency analysis to the developed to facilitate the use of time-
monitoring of machining processes, such as frequency methods by engineers whether or
drilling and grinding operations, have been not they are familiar with time-frequency
presented by Loughlin, Atlas, Bernard and analysis.
Pitton [11]. They showed that, although the Now, the efficiency of STFT for an industrial
Spectrogram (STFT) is an efficient method for case without any prediction of defects is
the demonstration of the time-varying demonstrated. This case comes from the
characteristics of a process, sometimes newer defective gear-train of a hoist drum in a large
time-frequency methods can provide more shovel operating at an open-pit iron mine. The
detail on the signal. They concluded that the data are measured by International
newer methods of time-frequency analysis Measurement Solutions company in order to
might assist in the early detection of problems. find the problem in the machine.
Atlas, Bernard and Narayanan [12] Gears generate a mesh frequency equal to the
summarized some applications of time- number of teeth on the gear multiplied by the
frequency analysis in different domains of rotational speed of the shaft driving it. A high
machinery diagnostics. They emphasised the vibration level at the mesh frequency is often
importance of using time-frequency analysis caused by tooth error, wear of the meshing
in manufacturing and monitoring applications. surfaces, or any other problem that would
The papers cited above give some examples of cause the profiles of meshing teeth to deviate
the application of different time-frequency from their ideal geometry. Sidebands at the
methods, including the STFT, to condition mesh frequency, on the other hand, are
monitoring of mechanical systems. Each of typically due to a failure of mating teeth.
these papers shows the way in which a new Imagine a cracked tooth, which is not yet
time-frequency method can reveal certain broken, and will consequently not be noticed
information about the signal that can not be by the operating personnel. However, it will,
obtained by traditional methods. due to its weakened mechanical condition,
Today, one of the most important factors deflect more under load than the other
limiting the progress of machine diagnostic (healthy) teeth when it goes into mesh. This
techniques is the lack of familiarity of results in a signal with amplitude modulation.
mechanical engineers with new signal Thus, an increasing level in the sidebands
processing methods. The complicated theory spaced with rotation speed in the frequency
of time-frequency analysis and the absence of spectrum results from the cracked tooth.
operational software for time-frequency A minimum length of time is required to
analysis restrict engineers from using these perform an FFT analysis of each process. The
Among the various time-frequency methods, time resolution required will depend on the
the Short-Time Fourier Transform is the period of each tooth mesh and the desired

4
level of accuracy. Sometimes, it is not gives the best result, in particular when the
possible to measure the signal for long enough precise type of fault is not known.
to provide the periodicity of shock in the FFT The Short-time Fourier Transform is an
spectrum. effective method of time-frequency analysis
In this particular case, the process did not even and a powerful tool in machine condition
last one revolution of the driven gear. The monitoring. The short-time spectrum gives a
case was investigated by time-frequency clear representation of the time-frequency
distribution precisely because it is known that plane and a simple interpretation of the energy
time-frequency methods do not need as much variation due to damage. There is,
time signal as the FFT spectrum. unfortunately, a fundamental problem with
Figure 1 shows respectively the signal and its this approach: high resolution cannot be
spectrum and its short-time Fourier transform. obtained simultaneously in the time domain
The spectrum of the signal indicates some and the frequency domain. Although this
large peaks around 200 Hz and some other method gives the time-frequency information
smaller peaks in the vicinity of 400 Hz, 800 with limited precision, but it is better than the
Hz and 1200 Hz. However, it is very difficult conventional methods of machine diagnosis
to assume or confirm any defects at this point. and may be used to present the general time-
On the other hand, The Short-Time Fourier varying characteristic of a signal.
transform clearly displays time-frequency The development of a user-friendly software
representation of the signal. There are a gear package facilitates the use of time-frequency
meshing frequency at approximately 220 Hz techniques in machine diagnosis. The time-
and some pulses at approximately 440 Hz. It frequency methods, including the short-time
has known that pluses are appeared in the spectrum, have been implemented on a
vibration signal of a gearbox if there is a computer and used, along with conventional
broken tooth. methods, in the analysis of vibration signals.
Then, we may conclude that there is a broken The advantages of the short-time spectrum
tooth in this gearbox. It is also possible to find have been demonstrated by using this method
the location of the fault in gearbox. The on vibration signals from an industrial
frequency of repetition of the pulses gearbox.
1
( = 1.5 Hz) determines the rotating
≈ 2.65 sec 7. References
speed of the shaft with faulty gear in gearbox. [1] R.G. Smiley, Rotating Machinery:
As well as, in frequency domain the frequency Monitoring and Fault Diagnosis, Sound and
of pulses ( ≈ 440 Hz) is equal to the meshing Vibration, 17(9) (1983) 26.
frequency of faulty gear.
[2] R. Archambault, Getting More Out of
6. Conclusion Vibration Signals: using the logarithmic scale,
It has been shown that, although the majority Proceedings of the 1st International Machinery
of conventional methods may give good Monitoring & Diagnostic Conference &
results when detecting a single fault in various Exhibit, (1989) 567.
simple elements of machines, no single
technique can provide all the answers for all [3] A. Barkov, N. Barkova and J.S. Mitchell,
cases. It is difficult to decide which method Condition Assessment and Life Prediction of

5
Rolling Element Bearing - Part 1, Sound And [9] W.J. Wang and P.D. McFadden, Early
Vibration, 29(6) (1995) 10. Detection of Gear Failure by Vibration
Analysis - I. Calculation of the Time-
[4] M.J. Brennan, M.H. Chen and A.G. Frequency Distribution, Mechanical Systems
Reynolds, Use of Vibration Measurements to and Signal Processing, 7(3) (1993) 193.
Detect Local Tooth Defects in Gears, Sound
and Vibration, 31(11) (1997) 12. [10] R. Rohrbaugh, Application of Time-
Frequency Analysis to Machinery Condition
[5] Angelo Martin, Vibration Monitoring of Assessment, Soc. Proceedings of the 27th
Machines, Brueel & Kjaer Technical Review, Asilomar Conference on Signal, Systems &
No. 1, PP 1-36, 1987 Computer, 2 (1993) 1455.

[6] R.M. Jones, Enveloping for Bearing [11] P. Loughlin, L. Atlas, G. Bernard and J.
Analysis, Sound and Vibration, 30(2) (1996) Pitton, Application of Time-Frequency
10. Analysis to the Monitoring of Machining, Life
Extension of Aging Machinery and Structures:
[7] D. Gabor, Theory of Communication, J. Proc. of 49th Mtg. of MFPT Soc., Vibration
IEEE (London), 93 (1946) 429. Inst.: Virginia Beach, 49(4) (1995) 305.

[8] B.D. Forrester, Use of Wigner-Ville [12] L. Atlas, G. Bernard and S.B. Narayanan,
Distribution in Helicopter Fault Detection, Applications of Time-Frequency Analysis to
Proceeding of Australian Symposium on Signal From Manufacturing and Machine
Signal Processing and Applications - ASSPA Monitoring Sensors, Proceedings of the IEEE,
89, Adelaide, PP. 78-82, April 1989. 84(9) (1996) 1319.

6
 Figure 1: The signal measured on a defective gearbox, its spectrum and its short-time Fourier transform.

View publication stats