IEEE TRANSACTIONS ON EDUCATION, VOL. 46, NO.

1, FEBRUARY 2003

149

Basic Vibration Signal Processing for

Bearing Fault Detection
S. A. McInerny and Y. Dai

AbstractFaculty in the College of Engineering at the University of Alabama developed a multidisciplinary course in applied
spectral analysis that was first offered in 1996. The course is
aimed at juniors majoring in electrical, mechanical, industrial,
or aerospace engineering. No background in signal processing
or Fourier analysis is assumed; the requisite fundamentals are
covered early in the course and followed by a series of laboratories
in which the fundamental concepts are applied. In this paper, a
laboratory module on fault detection in rolling element bearings is
presented. This module is one of two laboratory modules focusing
on machine condition monitoring applications that were developed
for this course. Background on the basic operational characteristics of rolling element bearings is presented, and formulas given
for the calculation of the characteristic fault frequencies. The
shortcomings of conventional vibration spectral analysis for the
detection of bearing faults is examined in the context of a synthetic
vibration signal that students generate in MATLAB. This signal
shares several key features of vibration signatures measured on
bearing housings. Envelope analysis and the connection between
bearing fault signatures and amplitude modulation/demodulation
is explained. Finally, a graphically driven software utility (a set
of MATLAB m-files) is introduced. This software allows students
to explore envelope analysis using measured data or the synthetic
signal that they generated. The software utility and the material
presented in this paper constitute an instructional module on
bearing fault detection that can be used as a stand-alone tutorial
or incorporated into a course.
Index TermsBearings, envelope analysis, fault detection, kurtosis, spectral analysis.

I. INTRODUCTION

COURSE in applied spectral analysis team-taught by the

faculty from the Electrical Engineering and Aerospace
Engineering and Mechanics Departments of the University
of Alabama in Tuscaloosa, AL. was first offered in 1996 [1].
The course is geared for mechanical, electrical, and aerospace
engineering students in their junior year although more advanced students may also take the course. No signal processing
or Fourier analysis background is assumed. The only prerequisites are Classical Mechanics (Physics I) and Calculus II
(their second 4-h calculus course). The students have used
MATLAB in their calculus and freshman engineering courses,
but are generally not MATLAB proficient. Example MATLAB
programs are used in classroom illustrations; homework and
laboratory exercises build on these sample programs.
Manuscript received June 8, 2001; revised June 25, 2002.
S. A. McInerny is with the Aerospace Engineering and Mechanics Department, The University of Alabama, Tuscaloosa, AL 35487 USA (e-mail: [email protected]).
Y. Dai is with the Haldex Brake Products Corporation, Prattville, AL 36067
USA (e-mail: [email protected]).
Digital Object Identifier 10.1109/TE.2002.808234

The objectives for this course are straightforward. First and

foremost, students who complete the course should be able to
use and understand a two-channel spectral analyzer. This means
the students must have a fundamental understanding of time and
frequency domain analysis.[1] The course de-emphasizes digital signal processing (DSP) details, such as Z-transforms and
digital filter design, and focuses on a practical understanding of
the analysis and the spectral results. Students have the opportunity to pursue DSP, communications, acoustics, vibrations, electric machines, and electric power topics in greater depth in other
elective courses. The course is modular in structure, beginning
with a lecture module in which the fundamental concepts are
covered. A list of the topics and concepts covered in the Fundamentals Module is given in Table I.
The Fundamentals Module is followed by a series of laboratory modules in which these concepts are applied. These
modules can be taught in any order, providing considerable
scheduling flexibility for the involved faculty. The first few
times this course was offered, laboratory modules focused
on speech communication, machinery sound power measurements, machine condition monitoring (vibration monitoring),
and variable speed drives and motors. As time goes on, the
nature and content of the laboratory modules evolve to reflect
the needs of industry as well as the research interests of the
faculty. The last time that the course was offered, students
made use of experimental facilities developed to study variable speed fluid drive system efficiencies [2] in a module
focused on electrical power and power quality measurements
of variable speed drives.
This paper discusses a module on bearing diagnostics
developed as a second or alternate module on the application
of spectral analysis to machine condition monitoring. (The first
machine condition-monitoring laboratory developed for the
course focuses on diagnoses of imbalance and misalignment
in rotating machinery.) A software utility, actually a set of
graphically driven procedures that run under the MATLAB
environment, was developed to reinforce and illustrate the
signal processing techniques used in bearing diagnostics. In
this paper, background on rolling element bearings is presented
first. This presentation is followed by a discussion of basic
bearing fault diagnosis techniques before introducing the use of
the software utility to illustrate and reinforce these techniques.
II. BACKGROUND
A. Rolling Element Bearings
The main components of rolling bearings are the inner ring,
the outer ring, the rolling elements, and the cage (see Fig. 1).

0018-9359/03$17.00 2003 IEEE

150

IEEE TRANSACTIONS ON EDUCATION, VOL. 46, NO. 1, FEBRUARY 2003

TABLE I
TOPICS COVERED IN THE FUNDAMENTALS MODULE

(Some bearings have additional components, such as the guide

ring and seals used in spherical roller bearings.) Typically, the
inner ring of the bearing is mounted on a rotating shaft, and
the outer ring is mounted to a stationary housing. The rolling
elements may be balls or rollers. The balls in a ball bearing
transfer the load over a very small surface (ideally, point contact)
on the raceways. The load-carrying capacity of a ball bearing
is, therefore, lower than that of a roller bearing in which the
rollers transfer the load via line contact with the raceways. The
cage separates the rolling elements, preventing contact between
them during operation. It also helps to prevent poor lubrication
conditions and, in many cases, holds the bearing together during
handling.
Roller bearings (see Fig. 1) are generally used for applications
requiring high load carrying capability, but radial ball bearings
are the most common type of rolling element bearing. Radial
ball bearings are simple in design, suitable for high and even

Fig. 1.

Components of a rolling element bearing.

very high speeds, and robust in operation and require little maintenance. Angular-contact ball bearings have an angle between

McINERNY AND DAI: BASIC VIBRATION SIGNAL PROCESSING

Fig. 2.

151

Angular contact ball bearings.

Fig. 3.

Idealized vibration signature due to fault in outer bearing race.

the inner and outer rings, as shown in Fig. 2, which enables them
to support both radial and axial loads. This type of bearing is
used in the bearing fault simulator in the laboratory module, but
the basic ideas and equations developed for these bearings can
be readily adapted to roller bearings.
B. Bearing Failure Mechanisms
There are a number of mechanisms that can lead to bearing
failure, including mechanical damage, crack damage, wear
damage, lubricant deficiency, and corrosion. [3] Abusive
handling can induce nicks and dents, which are especially
harmful when located in areas tracked by the rolling elements.
When the smooth rolling contact surfaces are marred, higher
stress conditions imposed on the surface reduce bearing life
significantly. Permanent indentation created by rolling element
overload is called brinelling. A crack in a bearing component
may begin as a manufacturing-related defect or be induced by
operating stress via overload or cyclic loading.
Wear results in a gradual deterioration of the bearing
components, leading to a loss of dimensioning and associated
problems. When lubrication conditions become inadequate, the
increased friction results in metal-to-metal contact. Operating
forces can cause enlarged plastic deformation by tearing the
locally friction-welded regions from the metal matrix. Lighter
adhesive damage is often called scuffing or scoring; whereas,
more intensive damage is referred to as seizing or galling.
Abrasive wear occurs when hard particles become entrained
between the contact surfaces.
As discussed in connection with wear damage, above, insufficient lubrication accelerates the adhesive wearing progress.
Poor lubrication increases bearing component temperatures,
which speeds up the deterioration processes. Bearings that
operate in an environment of high humidity may be subjected
to surface oxidation and produce rust particles and pits. These
particles can produce rapid wear via abrasion. The pits function
as stress raisers, providing sites for crack initiation.

Fig. 4. Geometric parameters used in bearing fault frequency calculations.

natural frequency, a response that decays quickly because of

damping (illustrated in Fig. 3). The excitation and response
occur each time one of the balls rolls over the flaw, so that the
fundamental frequency of the response waveforms is the rate
at which the elements roll over the flaw. It is this fundamental
frequency that is of interest in the detection of bearing faults,
not the resonance frequency at which the bearing rings, and it
can be predicted from the bearing geometry and the speeds at
which the inner and outer races rotate (see Fig. 4). Different
frequencies are obtained for flaws in an outer race, inner race,
on one of the balls, or in the cage. [4]
For an angular contact ball bearing in which the inner race
rotates and the outer race is stationary, the four characteristic
frequencies are:
Train or Cage Frequency (FTF)
(1)
Ball Pass Frequency, Outer race (BPFO)
(2)
Ball Pass Frequency, Inner race (BPFI)

C. Characteristic Fault Frequencies

Consider an example where the outer race of a ball bearing is
flawed because of one of the failure mechanisms just discussed.
Each time one of the balls rolls over the flaw, a high-level
short duration (impulsive) force is incurred that causes the
bearing to vibrate. The bearing responds by ringing at its

and

(3)

Ball Spin Frequency (BSF)

(4)

152

IEEE TRANSACTIONS ON EDUCATION, VOL. 46, NO. 1, FEBRUARY 2003

where
is the ball diameter,
the pitch diameter,
the
number of balls, and the shaft rotation rate in hertz. These
formulas are theoretical, and discrepancies arise when bearings
carry significant thrust loads or there is slippage. Software and
software modules (sold as options with vibration condition
monitoring packages) that incorporate databases of parameters
for a wide variety of bearings are commercially available.
III. BASIC BEARING FAULT DETECTION TECHNIQUES
If vibration signatures measured on bearings were as uncomplicated as that in Fig. 3, detection of bearing flaws would be
an easy matter. In practice, however, vibrations measured on a
bearing are dominated by high-level imbalance and misalignment components and include random vibrations associated
with friction and other sources. Imbalance vibration occurs at
the shaft rate of rotation (referred to as the 1X); misalignment
shows up at the fundamental (1X) and its harmonics. The
spectral components associated with the ringing pulse sequence
are not integer harmonics of the fundamental, and would not
be mistaken for them, but are of relatively small amplitude.
In a conventional spectrum, these components are lost in the
spectral noise floor generated by random vibrations and leakage
from the high level harmonics.

Fig. 5. Construction of the synthetic signal. (a) Square wave approximation.

(b) Random noise. (c) Ringing pulse sequence. (d) Their sum.

A. Synthetic Signal
In the laboratory module, students construct a synthetic
signal that shares some of the characteristic features of the vibration signatures measured on bearings. This signal is synthesized
as
a) A five-term approximation to a 0.05 amplitude 6-Hz
square wave [see Fig. 5(a)]. A square wave contains only
odd harmonics, thus this approximation contains only the
fundamental (1X), 3X, 5X, 7X, and 9X harmonics. An
imbalance generates a high 1X level, and a misalignment
peaks, but the absence of
generates
even harmonics in the square wave does not affect the
instructive value of the synthetic waveform. The peak
amplitude of the waveform in Fig. 5(a) is 0.059; the rms
amplitude is 0.049.
b) Random noise with a Gaussian amplitude distribution
[see Fig. 5(b)]. Random noise is present in all measurements and arises from a variety of sources (background
vibrations, rubs, electrical noise). Random noise with
a Gaussian amplitude distribution is readily generated
using the randn command in MATLAB. Here, the noise
was scaled so that in Fig. 5(b) the maximum amplitude
is 0.5 and the rms amplitude is 0.11.
c) A ringing pulse sequence with a repetition rate of 40 Hz
(6.67X) and a 0.1 peak amplitude [see Fig. 5(c)]. A pulse
sequence generated by appending multiple records of a
single measured 4096 point pulse (the response of a table
to an impact) is used. This pulse sequence models the
resonant response of a highly damped bearing housing as
ball rolls over a race fault.
The individual signal components are shown in Fig. 5 along
with their sum the composite signal. A MATLAB m-file, syn-

Fig. 6. Spectra of the (a) Square wave approximation. (b) Random noise.
(c) Ringing pulse sequence. (d) Composite waveform in Fig. 5.

Fig. 7. Illustration of amplitude modulation of a high-frequency carrier signal

by a low-frequency modulating signal.

thetic.m, included in the downloadable program files can be

used to generate these components, the synthetic signal, and the
plot in Fig. 5. The rationale for the relative amplitudes of the
three signal components is as follows. Even though the peak
amplitude of the pulse sequence is larger than that of the square
wave, the associated spectral peaks are smaller. The random
noise is so large that it completely buries the harmonics of the
pulse sequence in a conventional spectrum, yet the fundamental
frequency associated with this waveform can be extracted from
the composite signal using envelope analysis.

McINERNY AND DAI: BASIC VIBRATION SIGNAL PROCESSING

153

Fig. 8. Application of the Hilbert transform to two time waveforms. a) Modulated sine wave. b) Pulse sequence from Fig. 1(c).

B. Application of Traditional Spectral Analysis

Spectra of the synthetic waveform and its components are
used to explain the limitations of simple spectral analysis when
applied to bearing fault detection (see Fig. 6). The 0.75-Hz resolution spectra in Fig. 6(a)-(c) illustrate two related concepts in
spectral analysis: narrow in the time domain and then spread out
in the frequency domain, and smooth functions (continuous with
continuous derivatives) have spectra that drop off more rapidly
in the frequency domain. The peak amplitude of the pulse sequence in Fig. 5(c) is 0.1; whereas, the peak of amplitude of the
square wave in Fig. 5(a) is only 0.05. However, the largest peak

in the spectrum of the pulse sequence is only 0.0015, while the

largest peak in the spectrum of the square wave is 0.0358. The
higher harmonics in the spectrum of the pulse sequence drop
off more slowly than those in the spectrum of the square wave
approximation. These peaks are lost in the spectrum of the composite signal [see Fig. 6(d)].
C. Amplitude Modulation and the Hilbert Transform
Envelope analysis is often explained in terms of amplitude
modulation and demodulation. Amplitude modulation is defined as the multiplication of a high-frequency carrier signal by

154

IEEE TRANSACTIONS ON EDUCATION, VOL. 46, NO. 1, FEBRUARY 2003

a low-frequency modulating signal, as illustrated in Fig. 7. This

nonlinear process gives rise to frequency components not in the
spectra of the original signals. The spectrum of the modulated
signal contains a peak at the carrier frequency with sidebands
spaced at the modulation frequency. [5]
The sequence of ringing pulses in Fig. 5(c) models the
periodic response of a bearing housing to bearings rolling
over a flaw. This waveform can be viewed as a carrier signal
at the resonant frequency of the bearing housing modulated by
the decay envelope. The frequency of interest in the detection
of bearing defects is the modulating frequency. Because the
resonant frequency of the housing is high and the number of
spectral lines typically limited, it is difficult if not impossible
to resolve sidebands of the bearing resonant frequency in a
conventional spectrum. Fortunately, bandpass and demodulation
techniques allow one to get a direct measure of the frequency
of the modulating signal (i.e., the bearing defect frequency).
The first step in amplitude demodulation is to rectify the
signal. This step can be achieved using the Hilbert transform.
While a discussion of Hilbert transform theory [6] is beyond
the scope of this paper, a conceptual understanding may be
obtained by considering the simple harmonic signal

Fig. 9.

Diagram of the data processing procedures used in envelope analysis.

(1)
This signal can be viewed as the real part of the complex signal
Fig. 10. Two signals with different kurtosis coefficients. The kurtosis
coefficient of the signal in (a) is larger than that of the signal in (b).

E. Time-Domain Statistics Kurtosis

(2)
The imaginary part of this analytic signal is the Hilbert transform of the real part. The magnitude of the analytic signal is ,
which is the envelope of the original signal.
In MATLAB, the command Hilbert(x), where x is a real
vector and produces a complex vector, whose real part is x
and whose imaginary part is the Hilbert transform of x. [7]
The magnitude of this complex vector (i.e., this digital time
sequence) is the envelope of x. Fig. 8(a) and (b) illustrate
the application of the Hilbert transform to a simple modulated
waveform and to the ringing pulse sequence in Fig. 5(c).

The kurtosis coefficient is a metric that is also used in the detection of bearing faults. It is calculated from the time domain
data. [10][13] The variance, , of a zero mean (no dc component) signal is the mean square value [i.e., the square of the
root-mean-square (rms) value]. AC coupled accelerometer signals are zero mean signals. A rectified waveform has a nonzero
average or mean value. In this case, the variance is given by
(3)
where is the number of data points in the data sequence, and
is the average value. The coefficient of kurtosis is a fourth-order
statistic normalized by the square of the variance

D. Envelope Analysis
The phrase envelope analysis typically refers to the
sequence of procedures indicated in Fig. 9. The purpose of
bandpass filtering is to reject the low-frequency high-amplitude
signals associated with imbalance and misalignment and to
eliminate random noise outside the passband. [8], [9] Analyzers and data loggers that offer envelope analysis generally
offer user selectable bandpass settings, such as 12.5 kHz,
2.55 kHz, 510 kHz, 1020 kHz, and 2040 kHz. After
bandpass filtering and rectification (here, using the Hilbert
transform), the final step in the envelope analysis process is
calculation of the spectrum of the rectified bandpassed signal.

(4)
Large values of indicate large excursions from the RMS value.
The physical meaning of kurtosis may be best explained by illustration. Fig. 10(a) shows a signal with kurtosis value that is
large, relative to that of the signal in Fig. 10(b). This statistical
moment is sensitive to the pulses induced by bearing defects and
has proven effective in the rolling element bearing diagnostics.
IV. INSTRUCTIONAL SOFTWARE UTILITY
In this section, the use of the software utility developed for
this module is described. The utility is invoked from within

McINERNY AND DAI: BASIC VIBRATION SIGNAL PROCESSING

Fig. 11.

Envelope analysis GUI.

Fig. 12.

Data import capabilities.

the MATLAB environment, but could be compiled into a

stand-alone executable program that does not require the user
to have the MATLAB program and toolboxes. The disadvantage of working with an executable program is that you are
unable to go in and modify or add to the programs capabilities.
The source code for the software discussed in this paper can be
downloaded from http://mcinerny.eng.ua.edu.
Students first apply the program to the synthetic signal that
they generated, and later to time domain data acquired on
a bearing fault simulator. The simulator utilizes a dc motor
directly coupled to a shaft with a rotor that is supported
between two rolling element bearings. A bearing fault was
created by drilling a small hole in the outer race. It is convenient
to have a motor speed controller, especially if one uses the
same simulator to demonstrate misalignment and imbalance.
However, a pulsewidth modulation- (PWM) type controller
is not advisable because it introduces harmonics of the drive
switching frequency. This simulator was designed and built by
students, but a variety of bearing fault simulators are available
on the market and can be found via a Worl Wide Web search on
Bearing Fault Simulator or Bearing Fault Demonstrator.

155

A. Starting the Software Utility

Fig. 11 is the graphical user interface (GUI) that comes up
when a user types bearinggui at the MATLAB command line.
B. Data Importation
Note, the load data window in the lower left corner of
Fig. 11. The user can import and further process time-domain
data stored in a variety of data file types. Fig. 12 indicates the
file formats that are currently supported, the types commonly
encountered. They include ASCII text files from an arbitrary
source (*.txt), or those obtained when a binary data file
recorded using the laboratory analyzer is converted to ASCII
using the manufacturers translation utility (*.adt). Whenever
possible, it is preferable to work with binary files because
they are much smaller. The long time record (LTR) data files
are time-domain data recorded using the long time record
capability of the laboratory analyzer. These files include header
information that specifies the sample rate and scale factor.
For data sets that consist of a single column of data points
without header information, the user is prompted for the data

156

IEEE TRANSACTIONS ON EDUCATION, VOL. 46, NO. 1, FEBRUARY 2003

sample rate and a scale factor (default value is 1) to be applied to

the data. Using this information and the number of data points,
a corresponding time vector is constructed. After loading, the
data is plotted automatically.
C. Bandpass Filtering
The user can choose one of the standard bandpass filters,
available in a drop-down list, or specify their own lower-cutoff
and upper-cutoff frequencies. After specifying the bandpass
filter frequencies, the user then clicks on the bandpass icon.
The bandpassed data set is then displayed. The bandpass
filtering process cannot be skipped before proceeding to the
Hilbert transform, but one can skip the envelope spectrum
process, entirely, and calculate the spectrum of the original
time sequence.
D. Hilbert Transform
When the user clicks on the Hilbert icon, the Hilbert transform of the bandpassed data set is calculated and the envelope
displayed.
E. Kurtosis
The kurtosis value displayed in the envelope analysis GUI is
that of the last active time domain data set (i.e., the original,
bandpassed, or enveloped bandpassed time sequence).
F. Spectrum Analysis
Spectra are calculated using an FFT routine that employs a
Hanning window and 50% overlap processing. The user can
choose the bandwidth and spectral resolution. (The program
down samples the enveloped data set by a factor of 2 before calculating the spectrum, since the bandwidth of enveloped data is
half that of the original data set. This process is transparent to the
user.) The envelope spectrum of the synthetic data in Fig. 5(d) is
displayed in Fig. 11. The 40-Hz bearing fault frequency and its
harmonics are distinctly visible in this spectrum. Space does not
permit us to show the synthetic data set as one proceeds through
the entire envelope analysis process. This process is very instructive, however, allowing the student to explore the effect of
different choices of filter settings for a given data set.

envelope analysis for bearing fault analysis were then examined.

Finally, a set of graphically driven procedures developed to
illustrate bearing fault detection techniques was presented.
Students who have taken the applied spectral analysis course
report that it was an excellent preparation for subsequent courses
(e.g., digital signal processing, vibrations, or acoustics). Several
students have gone on to pursue graduate work in related fields
or to work in the preventive maintenance/machine condition
monitoring field.
REFERENCES
[1] S. McInerny, H. Stern, and T. Haskew, Applications of dynamic data
analysis, IEEE Trans. Educ., vol. 42, pp. 276280, Nov 1999.
[2] S. A. McInerny, X. Gao, and T. Haskew, A PWM drive laboratory
exploiting research and curricular synergy, Int. J. Power and Energy
Syst., vol. 21, pp. 511, Feb. 2001.
[3] T. A. Harris, Rolling Bearing Analysis, 3rd ed. New York: Wiley, 1984.
[4] J. I. Taylor, The Vibration Analysis Handbook, 1st ed. Tampa, FL: Vibration Consultants, Inc, 1994.
[5] G. White, Amplitude Demodulation Principles, 1.2 ed. Bainbridge Island, WA: DLI Engineering Corporation, 1995.
[6] N. Thrane, The Hilbert Transform, Bruel & Kjaer Instruments, Marlborough, MA, 1984.
[7] MATLAB Signal Processing Toolbox Users Guide, vol. 4, The Math
Works, Inc., Natick, MA, 1996.
[8] R. M. Jones, Enveloping for bearing analysis, Sound and Vibration,
vol. 30, pp. 1015, Feb. 1996.
[9] Revolutions, Special Edition on Bearing Monitoring, SKF Condition
Monitoring, San Diego, CA, vol. 7.
[10] L. Ott, An Introduction to Statistical Methods and Data Analysis. North Scituate, MA: Duxbury Press, 1977.
[11] H. Prashad, Functional performance of roller bearings for acceptance
in routine applications, J. Inst. Eng. (India), pp. 105113, Aug. 1995.
[12] F. Honaravar and H. R. Martin, New statistical moments for diagnostics
of rolling element bearings, J. Manuf. Sci. Eng., vol. 119, pp. 425432,
Aug. 1997.
[13] R. B. W. Heng and M. J. M. Nor, Statistical analysis of sound and vibration signals for monitoring rolling element bearing condition, Appl.
Acoust., vol. 53, pp. 211226, Jan-Mar 1998.

V. CONCLUSION

S. A. McInerny received the B.S. degree in mechanical engineering from California State University, Long Beach (CSULB), in 1979, and the M.S. degree in
engineering and the Ph.D. degree in mechanical engineering from the University of California, Los Angeles, in 1984 and 1987, respectively.
She was a Member of the Technical Staff at The Aerospace Corporationfrom
1987 to 1993 and an Associate Professor of Mechanical Engineering at CSULB
from 1988 to 1993. She conducts applied research in fault detection and diagnostics in mechanical systems and has spent the last several summers working
in the propulsion diagnostics group at the Naval Air Warfare Center, Patuxent
River, MD, on bearing and gear diagnostics.

An instructional module on fault detection in rolling element

bearings has been described. This module was developed for
an existing course in applied spectral analysis, but could be
used as a stand-alone tutorial or incorporated into a course
on machine condition monitoring. After reviewing the basic
operation of rolling element bearings and the characteristics of
idealized bearing fault vibration signatures, the shortcomings of
conventional spectral analysis were illustrated with a synthetic
signal generated in MATLAB. The basis and effectiveness of

Y. Dai received the B.S. degree in mechanical engineering from Nanchang

University, China, in 1996 and the M.S. and Ph.D. degrees in mechanical
engineering from the University of Alabama, in 2002 and 2002, respectively.
Currently, he is working on brake noise reduction in the research and development center of Haldex Brake Products, Inc. His primary research interests are
in the areas of fault detection and diagnostics, vibration and noise control, and
vehicle dynamics.
Dr. Dai is an Associate Member of the Society of Automotive Engineers
(SAE).