Acoustic Signature Based Early Fault

Detection in Rolling Element Bearings

Amir Najafi Amin, Kris McKee, Ilyas Mazhar, Arne Bredin,

Ben Mullins and Ian Howard

Abstract Early fault detection in rotary machines can reduce the maintenance cost
and avoid unexpected failure in the production line. Vibration analysis can diagnose
some of the common faults inside the rolling element bearings; however, the
vibration measurement should be taken from a transducer that is located on the
bearing or very close to the supporting structure, which is sometimes not feasible.
This study compares acoustic and vibration signature-based methods for detecting
faults inside the bearings. It uses both time and frequency based fault indicators (i.e.
RMS, Kurtosis and envelope analysis) for investigating the condition of the system.
Experiments were carried out on a belt-drive system with three different bearing
conditions (normal, corroded and outer race fault). The experimental results show
acoustic signature-based methods can detect the system’s fault from close distances,
and even for relatively far distance, some bearing conditions are still detectable.

A. Najafi Amin (&)
K. McKee
I. Mazhar
A. Bredin
I. Howard

Department of Mechanical Engineering, Curtin University, Kent Street,
Bentley, WA 6102, Australia
e-mail: amir.najafiamin@curtin.edu.au
K. McKee
e-mail: k.mckee@curtin.edu.au
I. Mazhar
e-mail: i.mazhar@curtin.edu.au
A. Bredin
e-mail: a.bredin@curtin.edu.au
I. Howard
e-mail: i.howard@curtin.edu.au
B. Mullins
Department of Health, Safety and Environment, Curtin University, Kent Street,
Bentley, WA 6102, Australia
e-mail: b.mullins@curtin.edu.au

© Springer Nature Switzerland AG 2019 415

J. Mathew et al. (eds.), Asset Intelligence through Integration and Interoperability
and Contemporary Vibration Engineering Technologies, Lecture Notes
in Mechanical Engineering, https://doi.org/10.1007/978-3-319-95711-1_41
416 A. Najafi Amin et al.

1 Introduction

Fault detection in rotary machines is vital, and early fault detection can reduce the
maintenance cost and avoid unexpected failure in the production line. This study
focuses on fault detection in bearings. According to bearing manufacturers some of
the faults such a ball fault requires immediate action, and some like inner race and
outer race faults may give few weeks to few months till failure [11]; therefore it is
crucial to monitor and detect bearing faults at the early stage.
Root-Mean-Square (RMS), Crest factor and Kurtosis values can be used for fault
detection. Detecting faults through just using one of these values is not usually
possible and combinations of them can be required. This also helps to find the stage
of the fault in the system [2]. Principal Component Analysis (PCA) on historical
data can show unusual variability and is useful for pre-processing of multi-fault
diagnosis [7, 10]. Furthermore, neural networks and fault classification can be
implemented to facilitate the fault detection [5, 18, 20]. Spectral Kurtosis (SK) is
also a powerful tool which does not required historical data for detecting faults [15].
Localised faults in bearings generate periodic impulse forces each time the ball
passes over the defect [6]. The frequencies of these impulses depend on the
geometry of the bearing and the rotating speed. There are four common fault
frequencies for a bearing. The Fundamental Train Frequency (FTF), which indi-
cates the rotation rate of the cage or ball retainer. The Ball Spin Frequency (BSF),
which shows the rotation rate of the balls or rollers. The Ball Pass Frequency Outer
Race (BPFO) and the Ball Pass Frequency Inner Race (BPFI) show the rate at
which a ball passes over a fault in outer or inner race. When the outer race is
stationary, and the inner race is rotating with the shaft, these frequencies derive
from:


1 Bd
FTF ¼ 1  cos h x; ð1Þ
2 Pd

2 !
Pd Bd 2
BSF ¼ 1 ðcos hÞ x; ð2Þ
2Bd Pd


n Bd
BPFO ¼ 1  cos h x; ð3Þ
2 Pd


n Bd
BPFI ¼ 1þ cos h x; ð4Þ
2 Pd

where n is number of bearing balls, Bd is ball diameter, Pd is bearing pitch diameter,

h is contact angle of the ball on the race and x is the speed of shaft rotation [12].
For these mathematical models, it is assumed that there is no sliding between
elements; however, by having sliding, the theoretical value for the inner race fault
Acoustic Signature Based Early Fault Detection … 417

will be slightly lower than the actual value and for the outer race fault will be
higher [12].
Since there are a large number of frequency components in the raw frequency
data, detecting fault frequencies from raw frequency components is extremely
difficult [3]. The vibration response usually contains extraneous noise; however
adaptive filtering can be used for cancelling some of these noises [18]. Wavelet
transforms can also serve as a tool to filter the signal and extract the fault vibration
signals and detect the fault size [17, 20]. The fault frequencies can also be extracted
through the use of the envelope demodulation technique of high frequency reso-
nance responses [8, 15, 17]. For this procedure first the proper frequency band that
contains the impulsive components excited by the defect in the bearing should be
detected [15, 17]. Although SK initially was used for detecting randomly occurring
signals [3], it can also be used for detecting the proper frequency band for envelope
analysis [15, 16, 19].
Stress waves (SW) and acoustic emissions (AE) signals are also useful for
condition monitoring of rotary machines [9, 13, 14]. The AE sensors have also been
found to be more effective than accelerometers for early fault detection [4]. Still due
to attenuation of the signal, for getting rich data from the measurement, the
transducer needs to be located on the bearing or very close to the supporting
structure [1, 4], which is sometimes not feasible. This study uses acoustic data and
tries to detect faults inside a bearing. This allows the use of non-contact measured
acoustic data from reasonable distances from the source of the noise for the fault
detection.

2 Experimental Setup

Experiments in this research were carried out on a belt-driven fin fan test rig at
Curtin University. As Fig. 1 shows, the test rig consists of an electric motor which
drives a fan’s shaft using a V-belt. Two bearings support this shaft and the fan
rotates along a horizontal axis. Three accelerometers measure the vibration of the
motor, the driven bearing and the fan bearing. In addition, two high quality
microphones at the same position measure acoustic noise from two opposite
directions (driven side and fan side).
A National Instrument Compact DAQ system captures and digitizes data from
the accelerometers and microphones, and transfers the data to MATLAB. During
the test, data from all channels were digitized with 51,200 samples per second for
each channel for a duration of thirty seconds.
The focus of this study was on detecting faults in bearings. For this purpose,
different faults were separately induced to the bearing at the fan side. This bearing
has pitch diameter of 34.5 mm and 8 balls. Each ball has 7.92 mm diameter and
contact angle of 0 degree. Three different bearing conditions were investigated,
healthy bearing, bearing with outer race fault and corroded bearing. For introducing
the outer race fault, a hole was drilled in the outer race.
418 A. Najafi Amin et al.

Fig. 1 Test rig (left) and faulty bearings (right)

3 Experimental Results

Fault detection for bearings usually focuses on localized faults which provide
periodic impulses. This research investigated both localized and non-localized fault
detection. First, the system with the healthy bearing was tested to find the base line
for the RMS spectra and frequency spectra. In this test, the motor rotated at 13 Hz
(780 rpm) and the fan’s shaft rotated at 2.5 Hz.
For investigating the non-localized fault, the bearing at the fan side was replaced
with an artificially corroded bearing. Figure 2 compares the RMS spectra of the
healthy bearing and corroded bearing tests in 1/3rd Octave bands. This figure shows
there to be distinguished amplification in both acceleration and acoustic RMS
signature for frequencies between 200 and 1000 Hz.
For investigating the localized fault, the bearing at the fan side was replaced with
a bearing with an outer race fault (Fig. 1). For this bearing at three different shaft
speeds, the acceleration and acoustic data were captured. Equation 3 shows how the

Fig. 2 RMS comparison of normal and corroded bearing

Acoustic Signature Based Early Fault Detection … 419

Table 1 Tests speeds Speed of Speed of Ball pass frequency of the

motor (Hz) shaft (Hz) outer race fault (Hz)
13 2.5 7.6
17 3.2 10.0
20 3.8 11.7

outer race fault frequency can be derived from the geometry of the bearing. Table 1
shows the speed of the shaft for these tests, and the expected fault frequencies.
Figure 3 shows the acoustic SK for a healthy bearing and bearing with outer race
fault. As this figure shows, the acoustic frequency band of 200–500 Hz has high
Kurtosis which has been chosen for envelope analysis filtration. Figure 4 shows the
result of the acoustic envelope analysis for the three shaft speeds. In these figures,
the related frequency of outer race fault for each shaft speed and its harmonics are
clearly visible. These results show how the acoustic measurement at relatively close
distance can be used for detecting localized bearing fault.
The acoustic data in these tests was captured with a microphone very close to the
source of noise (near field); therefore, it could be assumed that the acoustic waves
come directly from the source to the microphone. Furthermore, acoustic measure-
ments were made for 8 different angles of orientation of the test rig, at distances of 2
and 4 m (Fig. 5), to investigate the effect of distance on acoustic fault detection.
The RMS results for the corroded bearing for these measurements still showed
distinguished amplification for frequencies between 200 and 1000 Hz; however, for
the bearing with the outer race fault, the fault frequencies were not detectable any
more. Since for these tests, the acoustic measurement was placed in the far field, the
acoustic reflection inside the lab could easily interfere with the measurement. As a
result, the periodic signals were faded out due to high noise levels.

Fig. 3 Acoustic Spectral Kurtosis for a healthy bearing based on acoustic measurement
420 A. Najafi Amin et al.

Fig. 4 Envelope analysis result for bearing with outer race

Fig. 5 Microphones
positions in respect to the test
rig at distance of 2 and 4 m

4 Discussion

For fault detection and condition monitoring of rotary machines, acceleration, SW

and AE signals are commonly used [9, 13, 14]. This study uses the common fault
detection techniques on acoustic measurement data and shows the measured sound
from close distance can be used for detecting some localized and non-localized
faults in bearings. In contrast to vibration measurement, where the transducer
should be mounted on the bearing or on the supporting structure, for the acoustic
measurement, the microphone can have more distance from the source of noise. In
addition, the accelerometers must be rigidly mounted while the microphone must
not be mounted and should be free of vibration input.
For extracting the fault frequency and its harmonics using envelope analysis, the
filtering usually starts at several KHz [15, 18]; however, the result of this study
shows with acoustic measurement, the whole frequency band for filtering can be
below 1 kHz. Therefore, for digitizing and capturing data, lower sample rates could
be sufficient and cheaper devices can be used.
This study investigates a bearing with outer race fault and shows that the fault
frequency related to this condition and also its harmonics are easily visible with
envelope analysis on both acceleration and acoustic data. Tests with different shaft
Acoustic Signature Based Early Fault Detection … 421

speeds show that the frequency band for envelope analysis for these speeds was not
very sensitive to the shaft speed. Furthermore, tests on different fan blade pitch
angles showed that there were no significant changes in RMS spectra or result of
envelope analysis as well.
This study only investigated one type of localized fault in the bearing, and the
rest of other known localized faults should be investigated; however, due to the
common nature of these faults to generate harmonic impulses, similar results are
expected to be achieved.
This study first used high quality microphones and showed that acoustic mea-
surement at close distance can be used for detecting bearing faults. It then repeated
the tests with low quality and cheap microphones. The comparison result between
these two types of microphone showed that only for low frequencies (below
40 Hz), there was a major difference in result. Since the selected frequency band for
RMS spectra and envelope analysis started at 200 Hz, these low-quality micro-
phones were found to be very suitable for acoustic measurement and bearing fault
detection.

5 Conclusion

This study investigated three bearing conditions and the following conclusions can
be drawn from the current research:
• For a corroded bearing which has non-localized fault, the RMS spectra in 1/3rd
Octave band showed major changes in both acceleration and acoustic signatures
for frequencies between 200 and 1000 Hz. The result from acoustic measure-
ment at both near and far distance to the source, confirmed this change in
acoustic RMS signature.
• For a bearing with outer race fault, the SK showed high kurtosis values for
frequencies between 200 and 500 Hz. Therefore, this frequency band was used
for envelope analysis. For this bearing condition, the envelope analysis result on
both acceleration and acoustic data at close distance showed the fault frequency
and its harmonics. The tests on different shaft speeds and different fan blade
pitch angles also confirmed that acoustic measurement at close distance can be
used for detecting localized faults in bearings.
• For the far distance measurement, the fault frequency of the bearing with outer
race fault was no longer visible due to interference of the acoustic reflection
inside the room with the original signal.

Acknowledgements This work has been supported by Woodside Energy, CISCO and Curtin
University.
422 A. Najafi Amin et al.

References

1. Amini A, Entezami M, Papaelias M (2016) Onboard detection of railway axle bearing defects
using envelope analysis of high frequency acoustic emission signals. Case Stud Nondestr Test
Eval 6:8–16
2. Bhende AR, Awari GK, Untawale SP (2014) Comprehensive bearing condition monitoring
algorithm for incipient fault detection using acoustic emission. J Tribol 2:1–30
3. Dwyer RF (1984) Use of the Kurtosis statistic in the frequency domain as an aid in detecting
random signals. IEEE J Oceanic Eng 9
4. Gu DS, Choi BK (2011) Machinery faults detection using acoustic emission signal—from
microdevices to helioseismology. INTECH Open Access Publisher
5. Hariharan V, Srinivasan P (2009) New approach of classification of rolling element bearing
fault using artificial neural network. J Mech Eng ME 40:119–130
6. Hemmati F, Orfali W, Gadala MS (2016) Roller bearing acoustic signature extraction by
wavelet packet transform, applications in fault detection and size estimation. Appl Acoust
104:101–118
7. Kamiel B, Forbes G, Entwistle R, Mazhar I, Howard I (2005) Impeller fault detection for a
centrifugal pump using principal component analysis of time domain vibration features.
Department of Mechanical Engineering, Universitiey Muhammadiyah Yogyakarta, Indonesia
8. Kamiel B, Mckee K, Entwistle R, Mazhar I, Howard I (2005) Multi fault diagnosis of the
centrifugal pump using the wavelet transform and principal component analysis. In:
Proceedings of the 9th IFToMM international conference on rotor dynamics. Springer,
pp 555–566
9. Mba D, Bannister RH, Findlay GE (1999) Condition monitoring of low-speed rotating
machinery using stress waves Part 1 & 2. Proc Inst Mech Eng Part E: J Process Mech Eng
213:153–185
10. Mckee KK, Forbes G, Mazhar I, Entwistle R, Hodkiewicz M, Howard I (2015) Cavitation
sensitivity parameter analysis for centrifugal pumps based on spectral methods. Engineering
Asset Management-Systems, Professional Practices and Certification. Springer
11. Quistgaard O (2012) Using selective envelope detection (SED) for the early detection of faults
in rolling-element bearings. In: Kjær B (ed) BPT0046-EN-11. Brüel & Kjær
12. Roque AA, Silva TAN, Calado JMF, Dias JCQ (2009) An approach to fault diagnosis of
rolling bearings. WSEAS Trans Syst Control 4:188–197
13. Sandoval HMU, Ramirez CAP, Méndez JEQ (2013) Acoustic emission-based early fault
detection in tapered roller bearings. Ingeniería E Investigación 33:5–10
14. Sato I (1990) Rotating machinery diagnosis with acoustic emission techniques. Electr Eng
Japan 110:115–127
15. Sawalhi N, Randall RB (2004) The application of spectral kurtosis to bearing diagnostics. In:
Proceeding of ACOUSTICS, November 2004 Gold Coast, Australia, pp 393–398
16. Sawalhi N, Randall RB (2013) Localized fault detection and diagnosis in rolling element
bearings: a collection of the state of art processing algorithms. In: 15th Australian
international aerospace congress
17. Shi DF, Wang WJ, Qu LS (2004) Defect detection for bearings using envelope spectra of
wavelet transform. J Vib Acoust 126:567–573
18. Wang H, Cheni P (2008) Fault diagnosis for a rolling bearing used in a reciprocating machine
by adaptive filtering technique and fuzzy neural network. WSEAS Trans Syst Control 7:1–6
19. Wang Y, Xiang J, Markert R, Liang M (2016) Spectral kurtosis for fault detection, diagnosis
and prognostics of rotating machines: a review with applications. Mech Syst Signal Process
66–67:679–698
20. Yaqub M, Loparo KA (2016) An automated approach for bearing damage detection. J Vib
Control 22:3253–3266