sensors

Article
A Reliable Fault Diagnosis Method for a Gearbox
System with Varying Rotational Speeds
Cong Dai Nguyen, Alexander Prosvirin and Jong-Myon Kim *
School of Electrical, Electronics and Computer Engineering, University of Ulsan, Ulsan 44610, Korea;
daimtavn@gmail.com (C.D.N.); a.prosvirin@hotmail.com (A.P.)
* Correspondence: jmkim07@ulsan.ac.kr; Tel.: +82-52-259-2217




Received: 9 April 2020; Accepted: 29 May 2020; Published: 31 May 2020


Abstract: The vibration signals of gearbox gear fault signatures are informative components that can be
used for gearbox fault diagnosis and early fault detection. However, the vibration signals are normally
non-linear and non-stationary, and they contain background noise caused by data acquisition systems
and the interference of other machine elements. Especially in conditions with varying rotational
speeds, the informative components are blended with complex, unwanted components inside the
vibration signal. Thus, to use the informative components from a vibration signal for gearbox
fault diagnosis, the noise needs to be properly distilled from the informational signal as much as
possible before analysis. This paper proposes a novel gearbox fault diagnosis method based on an
adaptive noise reducer–based Gaussian reference signal (ANR-GRS) technique that can significantly
reduce noise and improve classification from a one-against-one, multiclass support vector machine
(OAOMCSVM) for the fault types of a gearbox. The ANR-GRS processes the shaft rotation speed
to access and remove noise components in the narrowbands between two consecutive sideband
frequencies along the frequency spectrum of a vibration signal, enabling the removal of enormous
noise components with minimal distortion to the informative signal. The optimal output signal from
the ANR-GRS is then extracted into many signal feature vectors to generate a qualified classification
dataset. Finally, the OAOMCSVM classifies the health states of an experimental gearbox using the
dataset of extracted features. The signal processing and classification paths are generated using the
experimental testbed. The results indicate that the proposed method is reliable for fault diagnosis in a
varying rotational speed gearbox system.

Keywords: adaptive noise reducer; gaussian reference signal; gearbox fault diagnosis; one against on
multiclass support vector machine; varying rotational speed

1. Introduction
Gearboxes are widely used in industrial applications, usually in harsh and continuous conditions,
making them susceptible to a variety of failures. Defects can cause the gearbox system to break down
and potentially damage complex mechatronic equipment or even cause a serious threat to safety,
property, or customer satisfaction. Therefore, it is essential to diagnose gearbox faults regularly to
ensure their early detection. The vibrations of gearbox systems have been studied since the 1980s,
and previous researchers have found that gearbox vibrations have a keynote meshing frequency [1,2]
with complex sidebands around it and its harmonics [3,4]. Therefore, the sideband frequencies and the
meshing frequency and its harmonics are the informative components for identifying gear faults. Signal
analysis is a backbone procedure for rotational-machine fault diagnosis research and applications.
It works by decomposing the related fault features that are the groundwork for identifying fault
patterns. The vibration characteristics of gearbox systems produce two major signals that can be
analyzed for fault detection: acoustic signals and vibration signals [5]. Vibration signals are the

Sensors 2020, 20, 3105; doi:10.3390/s20113105 www.mdpi.com/journal/sensors

Sensors 2020, 20, 3105 2 of 25

most popularly used ones for gearbox fault monitoring because acquiring vibration data is easy [6].
However, vibration signals contain many types of noise from sources such as measurement systems
(data acquisition systems), the environment, shafts, gears, and other related components and their
impingement [7,8]. All that noise, which exceeds the signal, fills the frequency spectrum of the vibration
signal, and eclipses it.
Many signal processing methods using advanced techniques have been presented by many
researchers: frequency analysis focusing on Fourier transform [9], Wigner distribution [10], rank-order
morphological filter [11], cyclostationary signals for mechanical applications [12], and the envelope
analysis [13], which is the most well-known for rotational-machine fault diagnosis applications such as
bearing-fault diagnosis. It detects the repeating shock amplitudes that appear as faulty teeth traverse
each rotation cycle. Using this method, the vibration signal is first processed by a bandpass filter to
achieve a high signal-to-noise ratio, and second, the Hilbert transform is used to achieve the envelope.
If the sideband frequencies in a gearbox vibration signal appear in the envelope, the presence of faulty
teeth in the gearbox can be deduced [14,15]. However, when the vibration signal is submerged in noise,
it is difficult to recognize the informative components for fault diagnosis in the envelope.
Time-frequency analyses were developed to process non-stationary signals using a frequency
transformation process divided based on windows across the time axis to capture informative events.
The basic time-frequency analysis method is a short-time Fourier transform (STHT) or a spectrogram,
such as a limited time window–width Fourier spectral analysis [16,17]. The challenges of the STHT
method, such as a failure of the assumption that the pieces of a non-stationary signal are stationary,
difficulty adapting the observation window size to the size of a real stationary piece of signal, and the
conflict between frequency resolution and time resolution (which is related to the Heisenberg uncertainty
principle) limit its usability. To resolve the disadvantages of the STHT method, the wavelet approach
was developed as an adjustable window frequency spectral analysis method. The basic wavelet
function can be modified to meet special needs, so the wavelet transform produces outputs with good
resolution in the low-frequency range and good time resolution in the high-frequency range [18–20].
In the region relevant for rotational-machine fault diagnosis, wavelet-based decomposition has been
widely used to apprehend the useful components of a vibration signal in a non-stationary condition (in
this context, non-stationary is the notion that the sideband frequency information of a vibration signal is
time-variant). Wavelet transform decomposes a vibration signal into many sub-bands that express the
time-frequency distribution through the dilation and transition of the mother wavelet. The sub-bands
that contain fault-related intrinsic features can then be used in the fault diagnosis process [21,22].
Nevertheless, the efficiency of the wavelet-based method correlates with the basic wavelet function,
so informative components that do not correlate as well with the applied wavelet could be missed or
lost in the transformed outcome. In addition, the white noise that is frequently parasitic in a vibration
signal and appears across the whole range of the frequency spectrum gives correlated oscillations with
a high potential to appear as excitation. In this paper, the effect of noise on the wavelet applied to the
wavelet transform method is compared with the proposed method for processing the signal path in the
experimental results.
The Hilbert-Huang transform (HHT) was introduced as a better methodology for analyzing
non-linear and non-stationary signals [23]. This technique is now often used for rotational-machine
fault diagnosis [24–27]. The HHT method uses a time adaptive operation known as empirical mode
decomposition (EMD) to decompose the signal into a group of complete and orthogonal components,
denoted as intrinsic mode functions (IMFs), that represent the intrinsic oscillation modes of the
fault-related components of a vibration signal. The HHT method was shown to outperform wavelet
transform in rotational-machine fault diagnosis in [28–30]. To capture the advantages of the HHT,
several fault detection tools combine EMD with other methods, such as envelope analysis and the
wavelet-based technique. EMD and the envelope analysis combine in series: the vibration signal is
first decomposed by EMD to determine the number of IMFs; the envelope analysis then processes the
IMFs to monitor for fault-related components. Compared with previous methodologies, this combined
Sensors 2020, 20, 3105 3 of 25

technique had better results [28,31,32]. Combining wavelet transform and EMD for time-frequency
analysis is another currently used combination method. It takes advantage of the strong points of
the two techniques and minimizes their limitations, particularly aliasing in the high-frequency band
(wavelet transform) and difficulties in isolating the signals within the second harmonic (EMD) [33,34].
However, the EMD technique is sensitive to noise, so noise-related IMFs, which are not useful,
can be generated by the EMD. As illustrated by Van M. et al. [35], EMD performs well in processing
low-noise vibration signals and poorly in processing high-noise signals. In other words, even EMD
combination techniques are unreliable in noisy environments. Therefore, to effectively apply enhanced
signal analysis techniques to non-stationary vibration signals, a proper pre-processing method to reduce
noise is required, such as narrowband demodulation [36] or discrete wavelet transform (DWT) [37,38].
Applying those de-noising methods effectively reduces the measurement noise, but the original
informative signal is also distorted by the attenuation of a narrow bandpass filter in narrowband
demodulation or the threshold in DWT-based de-noising. In other words, using one of those noise
reduction methods can degrade the performance of a fault diagnosis system. Therefore, we have
developed a new de-noising technique to reduce the noise from an original vibration signal by
optimizing a process for filtering the weights and parameters of the reference input signals (adaptive)
and considering the noise characteristics and rotational speed. We call our new technique the Adaptive
Noise Reducer–based Gaussian Reference Signal (ANR-GRS).
The adaptive noise-controlling technique reduces noise by means of destructive interference.
It consists of an adaptive filter and reference signals. The adaptive noise filter is a digital filter with an
adaptive algorithm that adjusts the filtering coefficients (or tap weights) so the filter can be flexibly
and optimally operated in unknown conditions with non-stationary signals to effectively remove
low-level noise [39]. The typical performance criterion for adjusting the filtering weights (convergence
condition) is based on the error signal, which is the difference between the output of the filter and
the input reference signal as determined using the recursive least-squares or least mean square (LMS)
algorithm. Between them, the LMS is more widely used because of its robustness and simplicity [40].
The ANR-GRS technique has three main function blocks: Gaussian reference signal (GRS) generation,
adaptive noise filtering using the LMS algorithm, and optimal output sub-band selection. The generated
GRS is a special signal consisting of a white-noise reference signal and a Gaussian reference signal
with adjustable parameters (mean value and standard deviation) to identify noise components that are
independent of the informative components in the frequency domains of the vibration signal from a
varying speed gearbox. The adaptive noise filter consists of an M-tap digital Finite impulse response
(FIR) filter and the LMS adaptive algorithm; it has two inputs: a reference input for the GRS signal
with specific parameters and the desired input for a vibration signal. The noise-reduced sub-band is
achieved as the output of the adaptive noise filter. The optimal output sub-band selection adjusts the
parameter of the GRS signals to receive the set of noise-filtered sub-bands output by the adaptive filters
and then selects the sub-band with the minimum mean square as the optimal sub-band, which is the
final output of the ANR-GRS module. That output becomes the input for the feature pool configuration
process used to extract the statistical features in the time and frequency domains of the vibration signal
as feature vectors [41] to be classified.
The heterogeneous feature pool improves the efficiency of gearbox fault expression for fault
diagnosis process; however, the high dimensionality of the feature vectors can be a challenge for
various machine learning techniques that can be used for decision making. In comparison with the
other artificial intelligence algorithms, the classification performance of support vector machines (SVM)
classifier is not much sensitive to the dimensionality of the feature vectors, in other words, this algorithm
is not affected by the problem called ‘curse of dimensionality’. Furthermore, SVM demonstrates
excellent generalization performance, so this technique is capable of achieving high accuracy while
classifying mechanical faults in rotation machinery [42]. Also, with an appropriate kernel function,
SVM can accurately separate the non-linear datasets by hyperplanes in high-dimensional feature space
using the non-linear mapping [43]. SVMs are widely used for fault diagnosis in many real-world
Sensors 2020, 20, 3105 4 of 25

applications [44]. They were originally designed for binary classification and then improved for
multiclass classification using the one against one, one against all, or hierarchical strategy. Among them,
the one against one strategy is the most reliable for our purposes [45–47]. Therefore, a one-against-one
multiclass SVM (OAOMCSVM) is used in this proposed methodology.
The new hybrid technique employs the ANR-GRS, which produces an optimal sub-band, and then
uses a machine-learning classification of fault types based on the OAOMCSVM on features extracted
from that optimal subband to identify faults in a gearbox system. The experimental results show that
the proposed method outperforms the aforementioned denoising methods, which verifies that the
“clean” input can be used to produce correct output from the signal processing and classification paths.
The rest of this paper is organized as follows: Section 2 provides the characteristics of a gearbox
vibration signal and the experimental test setup used in this study. The proposed methodology is
explained in detail, from theory to the construction of the ANR-GRS, feature pool configuration,
and OAOMCSVM classification, in Section 3. Section 4 demonstrates our experimental results in signal
processing and classification. Finally, Section 5 concludes the paper.

2. The Characteristics of a Gearbox Vibration Signal and Experimental Testbed Setup

2.1. The Characteristics of a Gearbox Vibration Signal

The defects of a gearbox can be classified into three major categories: manufacturing defects
(tooth profile error, eccentricity of the wheel, etc.), installation defects (parallelism), and operational
defects (tooth wear, case wear, tooth spalling, tooth cracks). This research considers operational faults.
A one-stage transmission gearbox, which consists of two rigid blocks with a pinion (on the drive
side) and a gear (on the non-drive side), is illustrated in Figure 1. A healthy gear in normal condition
working smoothly and periodically generates a linear and periodic vibration signal [3]. The vibration
signal, sh [mV/(m/s2 )], of a fault-free normal pair of gears meshing under a constant load speed can be
formulated as [48]:
N
X
sh (t) = Si cos(2πi fM t + ϕi ), (1)
i=0

where Si and ϕi are the amplitude and phase of the i-th meshing frequency harmonics; fM is the
meshing frequency (for the pinion: P is the number of teeth in the pinion wheel, fP is the pinion
rotation frequency, fM = P. fP ; or fM = G. fG , G is the number of teeth in the gear wheel, fG is the gear
rotation frequency), and N is the total number of fM harmonics in the frequency range of a vibration
signal. Figure 2a shows the spectrum of the output vibration signal of a fault-free gearbox; it is filled
with the frequency tones of the meshing frequency and its harmonics.
In a fault case, when the motion transferred from the drive shaft to the non-drive shaft by
the rotation between the pinion wheel and the gear wheel traverses a defective tooth (chipped,
worn, or missing), an abnormal movement occurs that changes the impulses in the vibration signals.
The vibration signal contains amplitude and phase modulations of the carrier frequency as the meshing
frequency; its frequency spectrum includes sidebands, frequency components on two sides of the
meshing frequency and its harmonics, as given in Equation (1). Thus, when the gear wheel has a
faulty tooth, the velocity of the gear angle changes impulsively within the rotating functionality and
generates a non-linear vibration signal in which issues such as speed variation, amplitude, and phase
modulation prevail [4]. The vibration signal is formulated [3] as given by Equation (2), and an example
of its spectrum is shown in Figure 2b:

N
X
s f (t) = Sk (1 + ak (t)) cos(2πk fM t + ϕk + pk (t)) (2)
k =0

PM   PM  
Here, ak (t) = j=0 A k j cos 2π j f G t + µk j and pk ( t ) = j=0 Pk j cos 2π j fG t + ξk j .
Sensors
Sensors 2019,
2019, 19,
19, xx FOR
FOR PEER
PEER REVIEW
REVIEW 55 of
of 25
25

𝐴𝐴 ,𝑃
,𝑃 are
are amplitudes
amplitudes andand µµ ,𝜉
,𝜉 areare phases
phases of
of the
the j-th
j-th sideband
sideband in
in the
the amplitude
amplitude and
and phase
phase
modulation
modulation signals,
signals, respectively,
respectively, around
around kk meshing
meshing harmonics.
harmonics.
Sensors 2020, 20, 3105 5 of 25

Sensors 2019, 19, x FOR PEER REVIEW 5 of 25

Ak j , Pk j are amplitudes and µk j , ξk j are phases of the j-th sideband in the amplitude and phase
𝐴 ,𝑃 respectively,
modulation signals, are amplitudes and µ ,𝜉 k
around aremeshing
phases of the j-th sideband in the amplitude and phase
harmonics.
modulation signals, respectively, around k meshing harmonics.

Figure
Figure
Figure 1.
1.The
1. The
FigureThe spur
spur gearbox
spur
spur
1. The gearbox
gearbox
gearbox model.
model.
model.
model.

(a) (b)
Figure 2. The frequency spectrum of the gearbox vibration signal: (a) a healthy gearbox and (b) a
faulty gearbox.

2.2.
(a)The Experimental Testbed Setup
(a) (b)
(b)
The experimental testbed is illustrated in Figure 3. The pinion wheel is fixed to a three-phase AC
Figure
Figure 2.
Figure 2. The
The
The frequency
frequency
frequency
induction spectrum
spectrum
spectrum
motor by of
of the
the gearbox
the
of(DS).
a drive shaft motionvibration
gearbox
gearbox
The vibration
vibration signal:
signal: (a)
signal: (a)
(torque) is transmitted(a)
fromaaahealthy
healthy
healthy
the gearbox
to the and
gearbox
AC motorgearbox and(b)
and (b) aaa
(b)
faulty
faulty load as adjustable blades, which are mounted on the end of the non-drive shaft by the engaged teeth
faultygearbox.
gearbox.
gearbox.
of a pinion wheel and a gear wheel (a gearbox with a gear reduction ratio of 1:1.52).

2.2.The
2.2.
2.2. TheExperimental
The ExperimentalTestbed
Experimental Testbed Setup
Testbed Setup
Setup
Theexperimental
The
The experimental testbed
experimental testbed is
testbed is illustrated
is illustrated in
illustrated in Figure
in Figure 3.
Figure 3. The
3. Thepinion
The pinionwheel
pinion wheelis
wheel isfixed
is fixedto
fixed toaaathree-phase
to three-phaseAC
three-phase AC
AC
induction
induction motor
induction motor by
motor by a drive
by aa drive shaft
drive shaft (DS).
shaft (DS). The
(DS). The motion
The motion (torque)
motion (torque)
(torque) isis transmitted
is transmitted
transmitted fromfrom
from thethe
the ACAC motor
AC motor to
motor to the
to the
the
load as
load
load as adjustable
as adjustable blades,
adjustable blades,
blades, which
whichare
which aremounted
are mountedon
mounted onthe
on theend
the endofof
end ofthe
thenon-drive
the non-driveshaft
non-drive shaft
shaft byby
bythe
the
theengaged
engaged
engagedteeth of
teeth
teeth
a
of pinion wheel
of aa pinion
pinion wheeland
wheel anda gear wheel
and aa gear
gear wheel (a
wheel (a gearbox
(a gearboxwith
gearbox with a gear reduction
with aa gear
gear reduction ratio
reduction ratio of 1:1.52).
ratio of
of 1:1.52).
1:1.52).

(a)

(a)
(a)

Figure 3. Cont.
Sensors 2020, 20, 3105 6 of 25
Sensors 2019, 19, x FOR PEER REVIEW 6 of 25

(b)

Figure 3. Experimental
Experimental testbed
testbed setup:
setup: (a)
(a) function
function block
block diagram;
diagram; (b)
(b) actual
actual experimental
experimental assembly.
assembly.

The number of teeth on the pinion wheel is 25 (P = = 25), the gear wheel has 38 (G = = 38), and the
length of each
each tooth
tooth isis99mm.
mm.Figure
Figure44depicts
depictsthe
theseeded
seededtooth
toothfailures
failures onon
thethe
gear wheel:
gear wheel:a perfect or
a perfect
healthy gear (H), tooth cut 10% (F1), tooth cut 30% (F2), and tooth cut 50% (F3).
or healthy gear (H), tooth cut 10% (F1), tooth cut 30% (F2), and tooth cut 50% (F3). To measure the To measure the speed
of the shaft
speed of therotation, a displacement
shaft rotation, transducer
a displacement is placed
transducer is to trackto
placed thetrack
holethe
in the DSinonce
hole peronce
the DS rotation.
per
The vibration
rotation. signals from
The vibration the gear
signals fromwheel in the
the gear normal
wheel in thecondition
normal and three levels
condition of tooth
and three cutofdefects
levels tooth
(shown
cut in Figure
defects (shown 4) in
were continuously
Figure acquired from
4) were continuously the vibration
acquired sensor
from the (accelerometer
vibration 622B01 made
sensor (accelerometer
by the IMI
622B01 madeSensor
by theCompany)
IMI Sensor mounted
Company)on themounted
end of theonDS,the72.5
endmm from
of the DS,the pinion
72.5 mm gear.
from The analog
the pinion
vibration
gear. The was digitized
analog vibrationusing PCI-2
was data acquisition
digitized using PCI-2(the data specifications of the
acquisition (the data acquisition
specifications system
of the data
are provided in Table 1). The sample datasets for each health condition (H, F1, F2,
acquisition system are provided in Table 1). The sample datasets for each health condition (H, F1, F2, F3) of the gearbox
under
F3) fourgearbox
of the shaft speeds
underare provided
four in Table
shaft speeds are 2.
provided in Table 2.

1. Specification of the sensors and data acquisition system.

Table 1.
Table

Device Device Specification Specification

Sensitivity (V/g):
SensitivitymV/(m/s
10.2 2)
(V/g): 10.2 mV/(m/s2 )
Vibration sensor
Operational frequency range: 0.42 to 10 kHz
(Accelerometer Operational frequency range: 0.42 to 10 kHz
Vibration sensor (Accelerometer 622B01) Resonant frequency: 30 kHz
622B01)
MeasurementResonant frequency:
range: ± 490 m/s2 30 kHz
Measurement ± 490 2
4- Channel DAQ 18-bit 40MHz AD conversion, a sampling frequencyrange:
of 65.536 kHzm/s
is used for each
PCI Board of18-bit
two channels
40MHz AD simultaneously
conversion, a sampling frequency of
4- Channel DAQ PCI Board
65.536 kHz is used for each of two channels simultaneously
Distance from the head of a transducer to a hole: 1.0 mm
Displacement Distance
Diameter from
of athe head
hole: of amm
12.80 transducer to a hole: 1.0 mm
transducer Sensitivity: 0 to −3dB
Diameter of a hole: 12.80 mm
Displacement transducer
Frequency response: 0–10 kHz
Sensitivity: 0 to −3dB
Each health state is sampled by sampling frequency 65536 Hz in continuous 1 second (1-sec
Frequency response: 0–10 kHz
sample) repeating by 300 times to receive 300 1-sec samples for each shaft speed. Hence, the number
of samples for each health state is 1200 vibration samples in four different shaft speeds, the total
Each
number of health
samplesstate is sampled
in this by sampling
experimental frequency
testbed is 65536
4800 of 1-sec Hz in continuous 1 s (1-s sample)
samples.
repeating by 300 times to receive 300 1-s samples for each shaft speed. Hence, the number of samples
for each health state is 1200 vibration samples in four different shaft speeds, the total number of
samples in this experimental testbed is 4800 of 1-s samples.
Sensors 2020, 20, 3105 7 of 25
Sensors 2019, 19, x FOR PEER REVIEW 7 of 25

(a) (b) (c) (d)

Thehealth
Figure4.4.The
Figure healthstates
statesofofthe
thegear
gearwheel
wheeland
andexamples
examplesof ofvibration
vibrationsignals
signalsatatthe
therotation
rotationspeed
speed
of300
of 300RPM:
RPM:(a)(a)no
noseeded
seededfault,
fault,healthy
healthygear,
gear,(b)
(b)tooth
toothcut
cut10%
10%(0.9
(0.9mm),
mm),(c)
(c)tooth
toothcut
cut30%
30%(2.7
(2.7mm),
mm),
(d)tooth
(d) toothcut
cut50%
50%(4.5
(4.5mm).
mm).

Table 2. A detailed description of the fault types and dataset.

Table 2. A detailed description of the fault types and dataset.
Gearbox Number of 1-s Data Samples Acquired for each Rotation SpeedSampling
Description Number of 1-sec data samples
Health State 300 RPM 600 RPM 900 RPM 1200 RPM Sampling
Frequency (Hz)
Gearbox health acquired for each rotation speed
Healthy No seeded fault inDescription
the frequency
state 300 300 300 600
300 900 300 1200 65536
(H) teeth of a gearbox (Hz)
RPM RPM RPM RPM
Fault type 1 Pinion tooth cut 10%
Healthy
(F1) No seeded
(0.9 mm) fault300
in the 300 300 300 65536
300 300 300 300 65536
Fault (H)
type 2 teeth
Pinion tooth cut 30%of a gearbox
300 300 300 300 65536
Fault
(F2)type 1 Pinion tooth cut 10%
(2.7 mm)
300 300 300 300 65536
Fault (F1)
type 3 (0.9 mm)300
Pinion tooth cut 50%
300 300 300 65536
(F3) (4.5 mm)
Fault type 2 Pinion tooth cut 30%
300 300 300 300 65536
(F2) (2.7 mm)
3. The Gearbox
Fault type 3 Fault Diagnosis
Pinion toothMethodology
cut 50%
300 300 300 300 65536
(F3) (4.5 mm)
A function block diagram of the method for gearbox fault diagnosis proposed in this study is
provided in Figure 5. We use four main processing blocks: the sensor and DAQ block, ANR-GRS,
3. The Gearbox Fault Diagnosis Methodology
feature pool configuration, and multiclass SVM-based classification. To acquire discrete samples of each
A function
captured signalblock
event diagram
containing of information
the method for gearbox
about fault diagnosis
the defective gearboxproposed in this study
in the acquisition is
dataset,
provided in Figure
the raw vibration 5. We
signal wasuse four main
sampled processing
at a high frequencyblocks: the sensor
of 65536 and DAQ
Hz to acquire rich block, ANR-GRS,
digitized vibration
feature pool configuration, and multiclass SVM-based classification. To acquire
sample data under different shaft rotation speeds (300, 600, 900 and 1200 RPM), and the adjustable discrete samples of
each
load captured signal event
was non-stationary. containing
Though information
the operation about range
frequency the defective gearbox
of a vibration in the
sensor acquisition
in this study is
dataset,
from 0.42theHzraw vibration
to 10 kHz (thissignal was sampled
is presented in Tableat1),a thus
highfault-related
frequency of 65536 Hz in
components to the
acquire rich
frequency
digitized vibration sample data under different shaft rotation speeds (300, 600,
domain of vibration signals mostly exist in the lowest segment 0–10 kHz of their frequency spectrums. 900 and 1200 RPM),
and the adjustable
Therefore, load was
the sampling non-stationary.
frequency of the rawThough
vibration thesignal
operation frequency
(all vibration rangeacquired
signals of a vibration
in this
sensor
paper) inwasthis study by
reduced is afrom
factor0.42 Hz tousing
of three 10 kHz (this is presented
a down-sampling in Table
technique. 1), thusimplementing
However, fault-related
components in the frequency
decimation involves aliasing, so domain of vibration
a low-pass Chebyshev signals
Typemostly
I Finiteexist in the
Impulse lowest segment
Response filter (the0–10
filter
kHz
withofthe
their frequency
order of 35 andspectrums.
a cut-off Therefore,
frequency the sampling
of 10 kHz) was frequency
used forofantialiasing
the raw vibration signal
[49]. The (all
output
vibration signals acquired in this paper) was reduced by a factor of three using
sub-band signal from the lowpass filter (lpf(n), n is denoted as discrete-time) which have frequency a down-sampling
technique.
spectrums However,
in the rangeimplementing
from 0–10 kHz, decimation
were theninvolves
optimized aliasing,
by theso a low-pass
ANR-GRS Chebyshev
to achieve Type I
the optimal
Finite Impulse
sub-band, Response
opt(n), from whichfiltertwenty-one
(the filter with the order
features of 35 andthrough
were extracted a cut-offafrequency
feature poolof configuration,
10 kHz) was
used
F(k),for
for antialiasing
classification[49]. TheOAOMCSVM.
by the output sub-band signal from the lowpass filter (lpf(n), n is denoted as
discrete-time) which have frequency spectrums in the range from 0–10 kHz, were then optimized by
the ANR-GRS to achieve the optimal sub-band, opt(n), from which twenty-one features were
extracted through a feature pool configuration, F(k), for classification by the OAOMCSVM.
Sensors 2020, 20, 3105 8 of 25
Sensors 2019, 19, x FOR PEER REVIEW 8 of 25

Figure 5. Function
Figure 5. Function block
block diagram
diagram of
of the
the proposed
proposed methodology.
methodology.

3.1. Adaptive Noise Reducer–based Gaussian Reference Signal

3.1. Adaptive Noise Reducer–based Gaussian Reference Signal
3.1.1. Adaptive Noise Filtering Technique
3.1.1. Adaptive Noise Filtering Technique
The Digital Filter
The digital filter
An adaptive filter combines the operation of a digital filter and an adaptive algorithm. The adaptive
An adaptive
algorithm optimizesfilter combines (or
the coefficient the weight)
operation of aofdigital
a digital
filterfilter and the
by using an adaptive algorithm.
feedback signal from The
the
adaptive algorithm optimizes the coefficient (or weight) of a digital filter by using the feedback
output (error signal) according to the signal condition or performance criteria [38]. Figure 6 illustrates signal
from
the the output
function of an(error signal)
adaptive according
filter to the
constructed signal
using condition
a FIR filter and oran
performance criteria [38].
adaptive algorithm. TheFigure
output6
illustrates the function of an adaptive filter constructed
of the FIR filter is calculated as given in Equation (3): using a FIR filter and an adaptive algorithm.
The output of the FIR filter is calculated as given in Equation (3):
XM
T
g(ng(n)
)== ∑ ccm ((n)r(n
n)r(n −−mm) )= (n)r(n)
= ccT(n)r(n) (3)
(3)
m=1

where,ccmm,, m
where, m= = 0,1,
0,1, …,
. . . ,M-1
M−1 (M(M
is is
the
thedigital filter
digital filterlength)
length)are
arethe
theadjustable
adjustableweights
weights(coefficients)
(coefficients)ofof
the filter,
the filter,which
whichdodonotnotdepend
dependononthethesample
sampletime.
time.TheTheweight
weightvector
vector(M(M×× 1)
1) is
is formed
formed as:
as:
c(n)≡ [c0, c1, …, cM-1]T, (4)
c(n)≡ [c0 , c1 , . . . , cM−1 ]T , (4)
and r(n-m), m = 0, 1, …, M-1 are samples of an input signal composed of the vector M × 1:
and r (n−m), m = 0, 1, . . . , M−1 are samples of an input signal composed of the vector M × 1:
r(n) ≡ [r(n), r(n-1), …, r(n-M+1)]T. (5)
T
r(n) ≡ [r(n),
T denotes the transpose operation of r(n−1), . . . , r(n−M+1)]
the matrix. .
Then, the error (5)
signal e(n) is the difference
between the FIR filter response, y(n), and desired signal, d(n), which can be calculated as:
T denotes the transpose operation of the matrix. Then, the error signal e(n) is the difference
e(n)and
between the FIR filter response, y(n), = d(n) - g(n)signal,
desired = d(n)d(n),
- cT(n)r(n).
which can be calculated as: (6)
A common criterion for tuning the convergence of the weight vector, c(n), is the minimization
T
of the mean-square error (MSE):e(n) = d(n) − g(n) = d(n) − c (n)r(n). (6)

A common criterion for tuning theJ ≡convergence

E{e2(n)} = E{[d(n)-cT(n)r(n)]2}
of the weight vector, c(n), is the minimization of
the mean-square error (MSE): (7)
J = cT(n)Rc(n)- 2PcT(n) + E{d2(n)},
where, R ≡ E{r(n)rT(n)} is the inputJ ≡ E{e2 (n)} = E{[d(n)−c
autocorrelation matrix,T (n)r(n)]
and P ≡2 }E{r(n)d(n)} is the cross-correlation
(7)
J = cthe
vector between the input signal and T (n)Rc(n) 2PcT (n)
desired−signal + E{d2 (n)},
vector.
Equation (7) indicates that the MSE is a quadratic function of the filter weights (c), and its
where, R ≡ E{r(n)rT (n)} is the input autocorrelation matrix, and P ≡ E{r(n)d(n)} is the cross-correlation
performance surface guarantees that it has a single global minimum MSE corresponding to the
vector between the input signal and the desired signal vector.
optimal vector co. The optimal vector co can be found by taking the first derivative of Equation (7)
Equation (7) indicates that the MSE is a quadratic function of the filter weights (c), and its
and setting it to zero, the result achieved by Wiener-Hopf equation (assuming that R has an inverse
performance surface guarantees that it has a single global minimum MSE corresponding to the optimal
matrix):
vector co . The optimal vector co can be found by taking the first derivative of Equation (7) and setting
co =R-1P, (assuming that R has an inverse matrix): (8)
it to zero, the result achieved by Wiener-Hopf equation
so that the minimum MSE is:
co =R−1 P, (8)
Jmin = E{d2(n)} – PTco. (9)
Sensors 2020, 20, 3105 9 of 25

so that the minimum MSE is:

Sensors 2019, 19, x FOR PEER REVIEW Jmin = E{d2 (n)} - PT co . (9)
9 of 25

Adaptive
AdaptiveAlgorithm
algorithm
The
Theadaptive
adaptivealgorithm
algorithmisisaa recursive
recursivefunction
function to
to automatically
automatically adjust
adjust the
the coefficient
coefficient vector,
vector, c(n),
c(n),
to minimize MSE (J
to minimize MSE (Jmin ) so that the weight vector converges to the optimum solution,
min) so that the weight vector converges to the optimum solution, c
coo, afteriteration
, after iteration
loops.
loops.Both
Boththe
theLMS
LMSand and recursive
recursiveleast-squares algorithms
least-squares algorithmscancan
be used to fetch
be used the optimal
to fetch solution
the optimal [39],
solution
but the LMS is the most broadly used. To calculate the updated weight vector in the
[39], but the LMS is the most broadly used. To calculate the updated weight vector in the recursive recursive loop,
the LMS
loop, thealgorithm is based
LMS algorithm is on the steepest-descent
based procedure
on the steepest-descent using a using
procedure negative gradientgradient
a negative of the instant
of the
square error, which was devised by Widrow and Stearns [50] as follows:
instant square error, which was devised by Widrow and Stearns [50] as follows:
c(n+1)==c(n)
c(n+1) c(n)++ µr(n)e(n),
µr(n)e(n), (10)
(10)
where µ is the step size (or convergence factor) that determines the stability and convergence rate of
where µ is the step size (or convergence factor) that determines the stability and convergence rate of
the LMS algorithm. The algorithm adapts the weight vector to the optimal Wiener-Hopf solution (co)
the LMS algorithm. The algorithm adapts the weight vector to the optimal Wiener-Hopf solution (co )
given in Eqaution (9) by an iterative process with the convergence factor. The step size is selected in
given in Eqaution (9) by an iterative process with the convergence factor. The step size is selected in
the range [40]:
the range [40]:
22
00<<µ µ< < ,, (11)
(11)
MS MS
u

whereSuS isisthe
where theaverage
averagepower
powerofofthe input
the signal
input r(n).
signal r(n).

Figure6.6. Function
Figure Function block
blockdiagram
diagramof
ofan
anadaptive
adaptivefilter.
filter.

Adaptive Noise Filtering Technique Applied to a Vibration Signal

Adaptive noise filtering technique applied to a vibration signal
To construct the adaptive noise filter, the noise reference signal and observed signal are applied as
To construct the adaptive noise filter, the noise reference signal and observed signal are applied
the input signal of an adaptive filter (the input response in Figure 6) and the desired signal (d(n) in
as the input signal of an adaptive filter (the input response in Figure 6) and the desired signal (d(n)
Figure 6), respectively. The observed signal is the vibration signal acquired from the accelerometer
in Figure 6), respectively. The observed signal is the vibration signal acquired from the accelerometer
sensor and digitized by the DAQ block reflecting gearbox behavior as expressed by the informative
sensor and digitized by the DAQ block reflecting gearbox behavior as expressed by the informative
signals (s(n)) and the noise (w(n)), as shown in Figure 7. As explained in Sections 1 and 2, the informative
signals (s(n)) and the noise (w(n)), as shown in Figure 7. As explained in sections 1 and 2, the
signals and noise are formed by different sources: the informative signal comes from the vibration
informative signals and noise are formed by different sources: the informative signal comes from the
of the gear and pinion teeth, whereas the noise comes from the measurement system, unrelated
vibration of the gear and pinion teeth, whereas the noise comes from the measurement system,
gearbox components, and mechanical resonances. Therefore, the informative signal and noise represent
unrelated gearbox components, and mechanical resonances. Therefore, the informative signal and
independent processes (E{s(n)w(n)} = 0). To implement an effective adaptive noise filtering system, the
noise represent independent processes (E{s(n)w(n)} = 0). To implement an effective adaptive noise
generated noise reference signal, r(n), should meet two conditions (A and B):
filtering system, the generated noise reference signal, r(n), should meet two conditions (A and B):
(A) The generated
A) The generatednoise reference,
noise r(n), and
reference, r(n),informative signal, s(n),
and informative are uncorrelated
signal, and independent
s(n), are uncorrelated and
(E{r(n)s(n)} = 0)
independent (E{r(n)s(n)} = 0)
(B) The characteristics
B) The characteristics of
of the
the generated
generated noise
noise reference,
reference, r(n),
r(n), and
and noise,
noise, w(n),
w(n), are
are homologous
homologous as
as
much as
much as possible.
possible.

When those conditions are met, the MSE of the adaptive noise filter can be calculated as follows:
E{e2(n)} = E{[s2(n) + (w(n)-ro(n))]2}, (12)
where ro(n) = cT(n)r(n), and the adaptive filter uses an FIR filter,
E{e2(n)} = E{s2(n)} + E{[w(n)- cT(n)r(n)]2}. (13)
Sensors 2020, 20, 3105 10 of 25

When those conditions are met, the MSE of the adaptive noise filter can be calculated as follows:

E{e2 (n)} = E{[s2 (n) + (w(n) − ro (n))]2 }, (12)

where ro (n) = cT (n)r(n), and the adaptive filter uses an FIR filter,

E{e2 (n)} = E{s2 (n)} + E{[w(n) − cT (n)r(n)]2 }. (13)

Sensors 2019, 19, x FOR PEER REVIEW 10 of 25
The informative signal is independent of both the noise (E{s(n)w(n)} = 0) and the generated noise
reference (E{r(n)s(n)} = 0).signal
The informative By implementing
is independentthe LMSthe
of both adaptive algorithm= to
noise (E{s(n)w(n)} adapt
0) and thethe filter coefficient
generated noise
vector,reference
c(n), to (E{r(n)s(n)}
the optimal = 0). By implementing
vector, thesquare
co , the mean LMS adaptive algorithm
of the output to adapt
signal thesignal)
(error filter coefficient
approaches
vector,
the single c(n), to the
minimum of optimal vector, co, surface.
the performance the meanFrom
squareEquation
of the output
(13), signal (error signal)
the minimum MSEapproaches
is taken to be:
the single minimum of the performance surface. From Equation (13), the minimum MSE is taken to
be:
n
2
o n
2
o 2
( )
minE e n ( ) ( ) T ( ) ( )]
= E s n + minE{[w n − c n r n } (14)
c(n) c(n)
min E{e (n)} = E{s (n)} + min E{ w(n) − c (n)r(n) } (14)
( ) ( )
Therefore, the output signal of the adaptive noise filtering system carries the complete informative
part of theTherefore,
gearbox the output signal
vibration signal throughout
of the adaptive noise filtering
the whole processsystem carries implementation.
of algorithm the complete
informative part of the gearbox vibration signal throughout the whole process of algorithm
In addition, the noise integrated into the vibration signal is reduced; in the ideal case, the noise
implementation. In addition, the noise
2 integrated into the vibration signal is reduced; in the ideal
is removed (min
case, the E{[w(n) − cT (n)r(E{ } =−0).
n)]w(n) Therefore,
c (n)r(n) forTherefore,
} = 0). adaptivefor
noise control,
noisewe implement
cnoise
(n) is removed (min
𝐜( )
adaptive control, we
the ANR-GRS.
implement the ANR-GRS.

Figure
Figure 7. Function
7. Function block
block diagramofofan
diagram anadaptive
adaptive noise
noisefiltering
filteringtechnique.
technique.

3.1.2. ANR-GRS
3.1.2. ANR-GRS
In thisInpaper,
this paper, the noise
the noise (w(n))
(w(n)) in in
thethe gearboxvibration
gearbox vibration signal
signal isisdivided
dividedinto two
into types:
two white
types: white
noise (u(n)) and band noise (b(n)). The white noise arises from the measurement system: the
noise (u(n)) and band noise (b(n)). The white noise arises from the measurement system: the amplifier,
amplifier, detector, DC power supply, thermal vibration of the semiconductor atoms, etc. In the
detector, DC power supply, thermal vibration of the semiconductor atoms, etc. In the frequency domain,
frequency domain, the power of the white noise is spread across the whole frequency spectrum of
the power of the white noise is spread across the whole frequency spectrum of the vibration signal
the vibration signal (theoretically, the power of white noise is spread from -∞ to ∞ in the frequency
(theoretically,
axis) [8]. the
Band power
noise,of
onwhite noise
the other is spread
hand, fromnoise
represents -∞ to ∞ in by
caused theunrelated
frequency axis) [8]. Band
components noise,
[7]. The
on thefrequency
other hand, harmonics of the band noise are distributed around the informative components of the of
represents noise caused by unrelated components [7]. The frequency harmonics
the band
gearnoise are frequency,
sideband distributed around
meshing the informative
frequency, and theircomponents of the gear
harmonics. Therefore, sideband frequency,
the informative signal
meshinginside the vibration
frequency, signalharmonics.
and their is separatelyTherefore,
independent
theofinformative
both types ofsignal
noise.inside
The ANR-GRS module
the vibration is is
signal
built independent
separately using the adaptive noise
of both filtering
types technique
of noise. and reference
The ANR-GRS noise-related
module generation
is built using signals, as
the adaptive noise
illustrated in Figure 8. To reduce the white noise, we apply a generated white noise
filtering technique and reference noise-related generation signals, as illustrated in Figure 8. To reduce signal with a
uniform, random distribution function (v(n)). The oscillation form of the generated white noise is
the white noise, we apply a generated white noise signal with a uniform, random distribution function
thus analogous to the white noise integrated into the vibration signal. Because its frequency spectrum
(v(n)). The oscillation form of the generated white noise is thus analogous to the white noise integrated
is within the observed frequency range, the maximum level of the power spectrum average (PSA) of
into the vibration signal. Because its frequency spectrum is within the observed frequency range,
the reference white noise is reduced to less than 10% (10% in this study) of the PSA of the vibration
signal to ensure that the informative signal can be eligible for conditions A or B. The GRS (g(n)) is
created to adapt to the band noise inside the vibration signal. To make the proposed methodology as
an invariant model, the GRS generation module uses the shaft rotation speed (RPM) information
from the displacement transducer and the vibration signal as the input parameters. Then, the mean
frequency (F ) and the standard deviation of the GRS are calculated based on the frequency of
Sensors 2020, 20, 3105 11 of 25

the maximum level of the power spectrum average (PSA) of the reference white noise is reduced to
less than 10% (10% in this study) of the PSA of the vibration signal to ensure that the informative
signal can be eligible for conditions A or B. The GRS (g(n)) is created to adapt to the band noise inside
the vibration signal. To make the proposed methodology as an invariant model, the GRS generation
module uses the shaft rotation speed (RPM) information from the displacement transducer and the
vibration signal as the input parameters. Then, the mean frequency (FCenter ) and the standard deviation
of the GRS are calculated based on the frequency of the defective wheel, which is a function of RPM
(the gear frequency in this paper). The GRS window is confined entirely within the frequency space
between two consecutive sideband frequencies (a sideband segment), pictorially described in Figure 9,
and computed as follows:
Nb
X (k−FCenter )2
WGRS (k) = e− 2∆ , (15)
k=1

where ∆ = σ2 is the variance, σ is the standard deviation of the GRS window, and FCenter is the mean
frequency of the GRS window. They function as the frequency of a faulty wheel (the gear frequency,
fG = P.RPM/G in this research).
FCenter = α. fFW . (16)

By linearization of the Gaussian function, the standard deviation (the characteristic of a Gaussian
distribution) can be approximately calculated as:

σ = 0.318.FCenter = 0.318.α.fFW . (17)

Nb is the number of frequency bins in a sideband segment and defined as follows:

2Ns
Nb = .f , (18)
Fs FW

where Ns is the number of samples of the vibration signal, Fs is the sampling frequency of the vibration
signal, and fFW is the frequency of the faulty wheel (the gear frequency, fG , in this paper).
To qualify condition B, the frequency components of a Gaussian window are separated from the
informative frequencies (sideband frequencies). A Gaussian window is placed completely inside the
space between two continual sideband frequencies in the visualization. Thus, the adaptation process
for a band-noise reduction is to preserve the original informative frequency component (significantly
reducing the noise components and causing negligible attenuation of the informative components).
From Equations (15)–(17) and Figure 9, the coefficient α is selected in the range from 0.25 to 0.75,
and the qualified Gaussian window signals are generated using the parameters in the following ranges:
The range of the mean value:

0.25. fFW ≤ FCenter ≤ 0.75.fFW . (19)

The range of the standard deviations of the Gaussian reference signal:

(
0.318.α. fFW when 0.25 ≤ α ≤ 0.5
σ= (20)
0.318.(1 − α). fFW when 0.5 < α ≤ 0.75

Therefore, the implementation of a stepping adjustment in the coefficient α drives a change in the
mean value and standard deviation (the position and shape) of the Gaussian window, which defines
the condition for fetching the optimal Gaussian window, as illustrated in Figure 9.

3.1.3. The Process for Calculating the Optimized Subband

First, the ANR-GRS algorithm germinates the initial parameters for the Gaussian signal generation
module: starting value of α = 0.25 in this paper, adaptive filter (M-tap, M = 40 in this study), coefficient
Sensors 2020, 20, 3105 12 of 25

vector c(n) = [0, 0, . . . , 0], and step size µ (µ = 0.01). The parameter α is scanned in the range [0.25 0.75]
in steps of 0.01 in company with the input rotation speed (RPM) to compute the FCenter (mean value)
and standard deviation (σ) using Equations (16) and (17). To generate the specific GRS needed for the
reference input of the adaptive filter r(n), the output of the adaptive filter is connected to the minus port
of the summation module, ro (n). The vibration signal, which contains both the informative component
and noise, is entered as the desired input and delayed for M sampling time steps to be compatible with
the delayed processing of the FIR digital filter. The LMS algorithm adjusts the coefficient vector to
receive the LMS of the error, which is the output of the summation module. The output error signal,
which has LMS (and to which the optimal coefficient vector is set), is pushed into the set of
Sensors 2019, 19, x FOR PEER REVIEW
proposed
12 of 25
optimized sub-bands.
Sensors 2019, 19, x FOR PEER REVIEW 12 of 25

Figure 8. A function block diagram of the ANR-GRS module and parameter adjustments.
FigureFigure 8. A function
8. A function block
block diagramofofthe
diagram theANR-GRS
ANR-GRS module
moduleand parameter
and adjustments.
parameter adjustments.

Figure
Figure
Figure 9. 9.
The9. The
The overall
overall
overall flowflow
flow chart
chart
chart ofof
of GRS
GRS
GRS signalgeneration
signal
signal generation
generation for
for thethe
the
for ANR-GRS
ANR-GRS module.
module.
ANR-GRS module.

Finally,
Finally, thethe algorithm
algorithm calculatesthe
calculates themean
meansquare
squarevalue
valueofofeach
eachsub-band
sub-bandin
inthat
thatset
set and
and then
then
selects the sub-band with the minimum value as the optimized sub-band and output
selects the sub-band with the minimum value as the optimized sub-band and output of the AND- of the AND-
GRSGRS module
module (Figure
(Figure 10).
10).
Sensors 2020, 20, 3105 13 of 25

Finally, the algorithm calculates the mean square value of each sub-band in that set and then
selects the sub-band with the minimum value as the optimized sub-band and output of the AND-GRS
module (Figure 10).
Sensors 2019, 19, x FOR PEER REVIEW 13 of 25

Figure
Figure 10. The algorithm
10. The algorithm flow
flow chart
chart of
of the
the ANR-GRS
ANR-GRS module.
module.
3.2. Feature Pool Configuration
3.2. Feature Pool Configuration
We found the ANR-GRS methodology to be highly effective in reducing most of the noise
We found the ANR-GRS methodology to be highly effective in reducing most of the noise
components from a 1-s raw vibration signal while leaving the information about gearbox faults intact.
components from a 1-sec raw vibration signal while leaving the information about gearbox faults
The optimized sub-band output from the ANR-GRS, i.e., the “clean” signal presenting the characteristics
intact. The optimized sub-band output from the ANR-GRS, i.e., the “clean” signal presenting the
of the gearbox component vibration with trivial noise effects, carries the intrinsic fault symptoms of
characteristics of the gearbox component vibration with trivial noise effects, carries the intrinsic fault
the cut tooth defects. We then use those optimal sub-bands, rather than the raw 1-s vibration signals,
symptoms of the cut tooth defects. We then use those optimal sub-bands, rather than the raw 1-sec
to extract features. According to Caesarendra et al. [51], the statistical parameters from the time and
vibration signals, to extract features. According to Caesarendra et al. [51], the statistical parameters
frequency domains of the signal are congruent and subservient for fault classification using machine
from the time and frequency domains of the signal are congruent and subservient for fault
learning. Table 3 displays twenty-one features, eighteen time-domain features (e.g., root means
classification using machine learning. Table 3 displays twenty-one features, eighteen time-domain
square, square mean root, kurtosis, skewness, margin, impulse, and peak-to-peak value) and three
features (e.g., root means square, square mean root, kurtosis, skewness, margin, impulse, and peak-
frequency-domain features (root mean square frequency, frequency center, and root variance frequency)
to-peak value) and three frequency-domain features (root mean square frequency, frequency center,
for each optimal sub-band. The feature pool dimensionality is NHS × N1-SEC × NF , where NHS is
and root variance frequency) for each optimal sub-band. The feature pool dimensionality is NHS × N1-
the number of gearbox health states (number of classes) that need to be classified (4 classes in this
SEC× NF, where NHS is the number of gearbox health states (number of classes) that need to be classified
study: healthy, pinion tooth cut 10%, pinion tooth cut 30%, and pinion tooth cut 50%), N1-SEC is the
(4 classes in this study: healthy, pinion tooth cut 10%, pinion tooth cut 30%, and pinion tooth cut
number of 1-s samples of each class (300 in this study), and NF defines the number of features (21 in
50%), N1-SEC is the number of 1-sec samples of each class (300 in this study), and NF defines the number
this study). Therefore, groups of 21-feature vectors were considered as the validating input dataset for
of features (21 in this study). Therefore, groups of 21-feature vectors were considered as the validating
our proposed intelligent fault-detection method based on a multiclass SVM.
input dataset for our proposed intelligent fault-detection method based on a multiclass SVM.
3.3. Gearbox Fault Classification Using a Multiclass SVM Classifier
3.3. Gearbox Fault Classification Using a Multiclass SVM Classifier
The principle operation of an SVM is based on the statistical learning theory of Vapnik [45] and
The principle
quadratic operation
programming ofwas
[46]. It an SVM is based
actually on the
designed statistical
to classify learning
binary theory
datasets of Vapnik
by finding the [45] and
optimal
quadratic programming [46]. It was actually designed to classify binary datasets by
plane, generally called the hyperplane, with the largest margin-gap separating it from both binary finding the
optimal plane,
classes. generally
Let {(xm called
, ym ), m=1, 2, .the
. . , hyperplane, with the
M} be the given largestdataset
training margin-gap
with M separating
samples, itwhere
from each
both
binary classes. Let {(x , y ), m=1, 2, ..., M} be the given training dataset with M samples, where each
sample data x ∈ ℝ , ℝ is a D-dimensional feature vector, and y (y ∈ {−1, +1}) are the class
labels. The SVM is used to find a set of linearly separable hyperplanes between two classes and
maintain the maximum distance (called the margin) from both of them.
Sensors 2020, 20, 3105 14 of 25

sample data xm ∈ RD , RD is a D-dimensional feature vector, and ym (ym ∈ {−1, +1}) are the class labels.
The SVM is used to find a set of linearly separable hyperplanes between two classes and maintain the
maximum distance (called the margin) from both of them.

Table 3. Definition of statistical features in the time and frequency domains.

Features Equations Features Equations Features Equations

s N
Prms 1 P
Peak Max(|s|) Shape factor 1 N Mean (s) sn
N n=1 |sn | N
n=1
Shape factor
s
Root mean N N s
1 P 2 Psrm
P
sn Entropy − pn .log2 (pn ) square mean 1 N
square (srms ) N N n=1 |sn |
n=1 n=1 root
N  N 
sn −s sn −s 3 max(s)
 
Kurtosis 1 P
Skewness 1 P
Margin factor
N σ N σ ssmr
n=1 n=1
2
Max(|s|) N √
Crest factor Square mean root (ssmr ) ( N1 Peak to peak max(s)−min(s)
P
srms |sn |)
n=1
Clearance 5th normalized N 
Max(|s|) sn −s 5 Kurtoris

1 P
Kurtosis factor .
factor ssmr moment N σ s4rms
n=1
6th normalized N  N
Max(|s|) sn −s 6

1
s2n
P P
Impulse factor 1PN
N σ Energy of signal
N n=1 |sn | moment n=1 n=1
Root mean
s s
Nf Nf Nf
Frequency 1 P ( )2 Root variance
1 1 (S( f ) − FC)2
P
S( f ) square
P
center (FC) Nf Nf S f frequency Nf
f frequency f f

Here is an input signal (i.e., optimized subband), N is the total number of samples, S(f) is the magnitude response of
q
2
the fast Fourier transform of the input signal s, N f is the total number of frequency bins, σ = N1 N
P
n=1 (sn − s) ,
s2
and pn = PN n 2 .
n=1 sn

The hyperplane, denoted as w, is determined as the maximized width of the margin and the
minimized structural risk, given by:

M
1 X
(w, b) = argmin wT w + C ξm , (21)
w,b 2 m=1

subject to: ym (wT ψ(xm ) + b) ≥ 1 − ξm , ∀m = 1, 2, . . . , M; − ξm ≤ 0, ∀m = 1, 2, . . . , M

Here, b is bias, C is the trade-off parameter, ξ = {ξ1 , ξ2 , . . . , ξN } is the set of slack variables,
and ψ(.) is a feature vector in the expanded feature space. Equation (21) can be solved by applying the
Lagrange duality solution [44] as shown below:

M M M
X 1 XX
argmaxw(α) = αm − αm αk ym yk ψT (xm )ψ(xk ), (22)
α 2
m=1 m=1 k=1

subject to: M m=1 αm ym = 0, 0 ≤ bm ≤ C, ∀m = 1, 2, . . . , M where, αm and αk are Lagrange multipliers,

P

xm and xk are two input training vectors, and K(xm , xk ) = ψT (xm )ψ(xk ) is a kernel function used to
map the input data space into a higher-dimensional feature space. Several kernel functions, such as
linear, polynomial, Gaussian, radial basis, and sigmoid functions, can be used in SVM classification
methods. Countless classification applications have more than two classes in their datasets and thus
require a solution beyond the binary SVM just described. Multiclass SVMs have been developed to
classify datasets of N different classes (N > 2), and they use one of three structures: one-against-one,
one-against-all, and hierarchical. Among those structures, OAOMCSVM requires more classifiers than
the others, but it also has the most reliable classification accuracy [45]. Therefore, we use OAOMCSVM,
illustrated in Figure 11, in the methodology proposed in this paper.
Sensors 2020, 20, 3105 15 of 25
Sensors 2019, 19, x FOR PEER REVIEW 15 of 25

Figure 11. The classification methodology of OAOMCSVM.

4. Experimental Results
4. Experimental Results
To verify the advantage of the ANR-GRS module in the proposed methodology, we implemented
To verify the advantage of the ANR-GRS module in the proposed methodology, we
experiments in two technological zones, signal processing and features dataset classification,
implemented experiments in two technological zones, signal processing and features dataset
and compared our results with those from conventional methodologies.
classification, and compared our results with those from conventional methodologies.
4.1. Signal Processing Experimental Results
4.1. Signal Processing Experimental Results
The 1-s vibration signals acquired using the experimental testbed described above contained the
The 1-sec vibration signals acquired using the experimental testbed described above contained
phase and amplitude modulation signals bearing information about the health states of a gearbox.
the phase and amplitude modulation signals bearing information about the health states of a gearbox.
To investigate the effectiveness of the proposed noise reduction technique, the experimental dataset
To investigate the effectiveness of the proposed noise reduction technique, the experimental dataset
was collected under various shaft rotation speeds that are equal to 300, 600, 900, and 1200 RPM,
was collected under various shaft rotation speeds that are equal to 300, 600, 900, and 1200 RPM,
respectively. The vibration signals output from the accelerometer are analog, so they were digitized
respectively. The vibration signals output from the accelerometer are analog, so they were digitized
with a high holding sample frequency of 65536 Hz to gather as much information (and noise) as possible
with a high holding sample frequency of 65536 Hz to gather as much information (and noise) as
on the wideband PCI-based data acquisition board (Table 1). Each 1-s digital vibration signal was
possible on the wideband PCI-based data acquisition board (Table 1). Each 1-sec digital vibration
down-sampled three times, incorporating a lowpass filter for antialiasing to output a digital vibration
signal was down-sampled three times, incorporating a lowpass filter for antialiasing to output a
signal realistically compatible with the working range frequency of the accelerometer (0–10 kHz,
digital vibration signal realistically compatible with the working range frequency of the
Table 1). Then, the vibration signal was input into the ANR-GRS module (Figure 5). The shaft rotation
accelerometer (0–10 kHz, Table 1). Then, the vibration signal was input into the ANR-GRS module
speed (RPM), measured by the displacement transducer, was observed by the ANR-GRS module
(Figure 5). The shaft rotation speed (RPM), measured by the displacement transducer, was observed
according to appropriate vibration signal data to generate the Gaussian reference signal. The optimal
by the ANR-GRS module according to appropriate vibration signal data to generate the Gaussian
subbands were the output of the ANR-GRS module.
reference signal. The optimal subbands were the output of the ANR-GRS module.
To demonstrate the superiority of the ANR-GRS technique, we compared its optimized subbands
To demonstrate the superiority of the ANR-GRS technique, we compared its optimized
with the outputs of other signal processing approaches for noise reduction: the Hilbert transform
subbands with the outputs of other signal processing approaches for noise reduction: the Hilbert
(HT), window bandpass filter (WBF), and wavelet transform with optimal subband-based maximum
transform (HT), window bandpass filter (WBF), and wavelet transform with optimal subband-based
kurtosis (WTK). We tested those approaches by replacing the ANR-GRS module with them. Figure 12
maximum kurtosis (WTK). We tested those approaches by replacing the ANR-GRS module with
illustrates the frequency spectra compared with the input vibration signal. Figure 12a shows the output
them. Figure 12 illustrates the frequency spectra compared with the input vibration signal. Figure
of a lowpass filter that received a 1-s vibration sample with 900 RPM (15 Hz) of fault type 2 (meshing
12a shows the output of a lowpass filter that received a 1-sec vibration sample with 900 RPM (15 Hz)
frequency, fM = P.RPM = 25.15 = 375 Hz and sideband gear frequency, fG = P.RPM/G = 9.87 Hz,
of fault type 2 (meshing frequency, f = P. RPM = 25.15 = 375 Hz and sideband gear frequency,
f = P. RPM/G = 9.87 Hz, shown as lpf(n) in Figure 5 and labeled as the OutLPF signal in Figure 12a).
The output signals from the noise-reduction modules are shown in Figure 12b (OutHT signal), Figure
Sensors 2020, 20, 3105 16 of 25

shown as lpf(n) in Figure 5 and labeled as the OutLPF signal in Figure 12a). The output signals from
the noise-reduction
Sensors 2019, 19, x FORmodules
PEER REVIEWare shown in Figure 12b (OutHT signal), Figure 12c (OutWBF 16 ofsignal),
25
Figure 12d (OutWTK signal), and Figure 13, the proposed ANR-GRS (OptANR signal). The three
12c (OutWBF
conventional methodssignal),(HT,
Figure 12d WTK)
WBF, (OutWTK signal),the
changed andoutLPF
Figure 13, the proposed
signal ANR-GRS
into different shapes(OptANR
and types
signal). The three conventional methods (HT, WBF, WTK) changed the outLPF signal into different
(the outLPF signal is an amplitude and phase modulation signal) regardless of the fault information
shapes and types (the outLPF signal is an amplitude and phase modulation signal) regardless of the
(meshing frequency and its harmonics and sideband gear frequencies). HT exalted the area of the
fault information (meshing frequency and its harmonics and sideband gear frequencies). HT exalted
low-frequency
the area of components,
the low-frequency whereas WTK fortified
components, whereastheWTK high-frequency components components
fortified the high-frequency in the frequency
spectrum
in the(Figure
frequency 12b,d). WBF(Figure
spectrum was better than the
12b, Figure HT
12d). WBFandwas
WTK methods
better than thebecause
HT and it
WTKfiltered the noise
methods
in some of the
because it meshing
filtered the frequency
noise in harmonics
some of theand sideband
meshing gear frequencies,
frequency harmonics and butsideband
it also reduced
gear or
removed significantly
frequencies, informative
but it also reduced orfrequency
removed components (Figure 12c).
significantly informative The outANR
frequency signal(Figure
components (Figure 13),
12c). The outANR signal (Figure 13), the output signal from the ANR-GRS module
the output signal from the ANR-GRS module proposed here, fulfilled the needs of signal processing: proposed here,
fulfilled
reducing thethe needs
noise of signal processing:
components reducingthe
and preserving the original
noise components
informativeand components.
preserving the original
It made the
informative
vibration signalcomponents. It made“cleaner”
from the gearbox the vibration signal from
(lowered the the gearbox
noise) and “cleaner” (lowered
approached the noise)
the characteristics
and approached the characteristics of the gearbox vibrations signal presented in Section 2.1. This
of the gearbox vibrations signal presented in Section 2.1. This comparison verifies that our accurate
comparison verifies that our accurate is a suitable technique for reducing the noise in gearbox
is a suitable technique for reducing the noise in gearbox vibration signals and returning an honest
vibration signals and returning an honest reflection of the health states of a gearbox along an
reflection of the
electronic health
signal states of a gearbox along an electronic signal path.
path.

(a)

(b)

(c)

Figure 12. Cont.

Sensors 2020, 20, 3105 17 of 25
Sensors 2019, 19, x FOR PEER REVIEW 17 of 25
Sensors 2019, 19, x FOR PEER REVIEW 17 of 25

(d)
(d)
Figure
Figure 12. The
12. The frequency
frequency spectrumanalysis
spectrum analysisfor
for the
the state-of-the-art
state-of-the-artmethodologies:
methodologies: (a) the
(a)input signal,
the input signal,
Figure
(b) (b)12.
the theThe
output frequency
output signal
signal spectrum
from
from thethe analysis
Hilbert
Hilbert for the state-of-the-art
transform
transform module,
module,(c) themethodologies:
(c)the output signal
output from
signal (a) thewindow
the
from input
the signal,
window
bandpass
(b) the
bandpass filterfilter
output module,
signal
module, andand
from (d)(d)
the thethe output
Hilbert
output signal from
transform
signal the wavelet
module,
from (c) the
the wavelet transform
output WTK
transformsignal
WTK module.
from the window
module.
bandpass filter module, and (d) the output signal from the wavelet transform WTK module.

Figure 13. Frequency spectrum analysis of the input and output signal of the ANR-GRS module.

4.2. Classification Results

The 13.
Figure
Figure classification
Frequencyperformance
13. Frequency spectrum of the of
spectrum analysis
analysis proposed
of the
the input
inputmethodology
and
and output is evaluated
output signal
signal of
of the in two experiments.
the ANR-GRS
ANR-GRS module.
module.
At the beginning, the feature sets of the dimensions 4 × 300 × 21 (300 samples from each of four health
4.2. Classification
states, as shown Results
in Table 2) for each speed (in this experiment, 300 RPM, 600 RPM, 900 RPM, and
4.2. Classification Results
1200 RPM) were selected. Then, in the first experiment, the dataset is created by merging the data of
The classification performance of the proposed methodology is evaluated in two experiments.
The classification
four available healthperformance
states under of the proposed
different methodology
rotating speeds resultingisinevaluated in twoset
the new feature experiments.
of the
At the beginning, the
dimensionality 4800
feature
× 21.
sets
This
of the
dataset
dimensions
was then randomly
4 ×divided
300 × 21 into
(300
a
samples
train and a
from
test sets
each
at
of four
ratio
At the beginning, the feature sets of the dimensions 4 × 300 × 21 (300 samples from each of four health
health8:2states, as shown in TableOAOMCSVM
2) for each speed (in and this general
experiment, 300 RPM, 600 RPM, 900faultRPM,
states, asforshown
trainingin and
Table testing
2) for each speed classifier
(in this experiment, evaluation
300 RPM, of 600theRPM,
proposed
900 RPM, and
and 1200 RPM)
diagnosis were selected. Then, in the first experiment, the dataset is created by merging the
methodology.
1200 RPM) were selected. Then, in the first experiment, the dataset is created by merging the data of
data of four available health
Furthermore, to prove states
the under different
robustness rotating
of the speeds
proposed resulting
ANR-GRS in the new
technique, thefeature
secondset of
four available health states under different rotating speeds resulting in the new feature set of the
the dimensionality 4800 × 21. This dataset was then randomly divided into a train and a testto
experiment is performed where the classifier is trained on data (i.e., feature sets) corresponding sets at
dimensionality 4800 × 21. This dataset was then randomly divided into a train and a test sets at ratio
ratio single
8:2 forrotating
training speed
and and tested
testing by data instances
OAOMCSVM collected
classifier andunder
generalother speeds. Specifically,
evaluation feature fault
of the proposed
8:2 for
set training and
formethodology.
a single testing
rotating OAOMCSVM
speed was used as classifier
a training and general
set (for evaluation
instance, feature setofcorresponding
the proposed to fault
diagnosis
diagnosis methodology.
300 RPM with the dimensionality of 4 × 300 × 21), and the feature sets corresponding to two other
Furthermore, to prove the robustness of the proposed ANR-GRS technique, the second experiment
Furthermore, to prove the robustness of the proposed ANR-GRS technique, the second
is performed where the classifier is trained on data (i.e., feature sets) corresponding to single rotating
experiment is performed where the classifier is trained on data (i.e., feature sets) corresponding to
speed and tested by data instances collected under other speeds. Specifically, feature set for a single
single rotating speed and tested by data instances collected under other speeds. Specifically, feature
rotating speed was used as a training set (for instance, feature set corresponding to 300 RPM with the
set for a single rotating speed was used as a training set (for instance, feature set corresponding to
dimensionality of 4 × 300 × 21), and the feature sets corresponding to two other speeds were used for
300 RPM with the dimensionality of 4 × 300 × 21), and the feature sets corresponding to two other
Sensors 2020, 20, 3105 18 of 25

testing (for instance, feature set corresponding to two other speeds 600 RPM and 900 RPM with the
dimensionality of 4 × 600 × 21).
Those processes were run four times. To construct the training model for classification, k-fold
cross-validation (k-cv) was used to estimate the accuracy of the generalized classification [52]. In k-cv,
the set of samples in the feature vector is split randomly into k mutual folds (k=10 in this study), denoted
as C1, C2, . . . , Ck. The classification OAOMCSVM operates on k-times of the accuracy estimation.
Some folds {Cj} (a random subset from k folds) are used as a training set, and the rest are used as a
validation set and alternative iteration k times. More specifically, for each speed, 300 feature vectors for
each health state in the training set were partitioned into ten folds (each fold containing 30 randomly
chosen feature vectors (30 × 21) for each health state); 9 of those folds were used for training, and the
1 remaining fold was used for validation. That process was repeated 10 times until all folds had been
used as the validation set. The final measure of performance in the training model is the average
value of the accuracies attained in each fold. These data are then used as the testing dataset (which
was not used at all in the training process) to verify the OAOMCSVM method and provide the final
classification result.
We also used the OAOMCSVM classification method to classify the feature pool configuration
datasets extracted from the comparison signal processing methodologies: the raw vibration signal
(lowpass filter output signal) extraction (methodology I), HT, (methodology II), WBF (methodology
III), WTK (methodology IV), and the IMFs and residuals from the EMD (methodology V).
The implementation of those methodologies for achieving the classification result for the four
health states complied strictly with the conditions used with input from the ANR-GRS module
just described. To estimate the classification result between methodologies (the proposed method
and others), all twenty-one features of the vibration signal were used as input feature vectors for the
OAOMCSVM module to ensure that the most informative features for and from each methodology
were used fairly for the classification. The classification results of the state-of-the-art methodologies
and the proposed ANR-GRS methodology obtained during two experiments are shown in Tables 4
and 5 and Figure 14 to visualize the results tabulated in Table 5. Those classification accuracies were
computed as follows: P
NTP
Caccuracy = L .100% (23)
Nsamples
where L is the number of categories (L = 4 as four health states), NTP is the number of true positives
(the number of fault samples in category i that are correctly classified as class i), and Nsamples is the
total number of samples used to estimate the performance of the proposed methodology.

Table 4. Classification results for state-of-the-art methodologies and the proposed ANR-GRS
methodology by a combination dataset of different speeds.

OAOMCSVM (4800 Samples) Accuracy (%)

Methodology Training Set Test Set Fault Fault Fault Overall
Healthy
(80%) (20%) Type 1 Type 2 Type 3 (%)
I 3840 960 59 73 69 75 69.0
II 3840 960 84 80 67 83 78.30
III 3840 960 92 89 76 83 84.6
IV 3840 960 85 87 58 74 73.10
V 3840 960 92 89 88 94 90.80
ANR-GRS 3840 960 100 99 99 100 99.70
Sensors 2020, 20, 3105 19 of 25

Table 5. Classification results for state-of-the-art methodologies and the proposed ANR-GRS
methodology by observation of separated speed dataset.

OAOMCSVM (10-Fold
Accuracy (%)
Cross Validation)
Methodology
Training Set Test Set Fault Fault Fault Overall
Healthy
(300 Samples) (600 Samples) Type 1 Type 2 Type 3 (%)
600 RPM,
300 RPM 53 78 69 52 63
900 RPM
900 RPM,
600 RPM 74 47 53 80 63.5
1200 RPM
I
600 RPM,
900 RPM 54 46 64 81 61.25
1200 RPM
300 RPM,
1200 RPM 53 53 68 77 62.75
600 RPM
Overall by health states 58.5 56 63.5 72.5 62.63
600 RPM,
300 RPM 51 99 63 85 74.5
900 RPM
900 RPM,
600 RPM 75 67 64 72 69.5
1200 RPM
II
600 RPM,
900 RPM 75 48 70 83 69
1200 RPM
300 RPM,
1200 RPM 74 62 74 84 73.5
600 RPM
Overall by health states 68.75 69 67.75 81 71.63
600 RPM,
300 RPM 75 58 69 93 73.75
900 RPM
900 RPM,
600 RPM 74 70 80 84 77
1200 RPM
III
600 RPM,
900 RPM 70 49 72 63 63.5
1200 RPM
300 RPM,
1200 RPM 83 53 72 66 68.5
600 RPM
Overall by health states 75.5 57.5 73.25 76.5 70.69
600 RPM,
300 RPM 64 74 87 63 72
900 RPM
900 RPM,
600 RPM 82 49 72 64 66.75
1200 RPM
IV
600 RPM,
900 RPM 63 47 69 76 63.75
1200 RPM
300 RPM,
1200 RPM 63 49 70 67 62.25
600 RPM
Overall by health states 68 54.75 74.5 67.5 66.19
Sensors 2020, 20, 3105 20 of 25

Table 5. Cont.

OAOMCSVM (10-Fold
Accuracy (%)
Cross Validation)
Methodology
Training Set Test Set Fault Fault Fault Overall
Healthy
(300 Samples) (600 Samples) Type 1 Type 2 Type 3 (%)
600 RPM,
300 RPM 77 94 72 89 83
900 RPM
900 RPM,
600 RPM 90 82 91 82 86.25
1200 RPM
V
600 RPM,
900 RPM 94 80 69 85 82
1200 RPM
300 RPM,
1200 RPM 98 65 69 83 78.75
600 RPM
Overall by health states 89.75 80.25 75.25 84.75 82.5
600 RPM,
300 RPM 100 95 98 100 98.25
900 RPM
ANR-GRS 900 RPM,
600 RPM 98 99 99 100 99
1200 RPM
600 RPM,
900 RPM 98 99 97 99 98.25
1200 RPM
300 RPM,
1200 RPM 99 98 95 99 97.75
600 RPM
Overall by health states 98.75 97.75 97.25 99.5 98.31

Table 4 illustrates that the proposed technique significantly outperforms its counterparts when
it is trained on the data instances corresponding to all available speeds and achieving the highest
accuracy of 99.7%.
Table 5 demonstrates that the proposed approach using ANR-GRS also yielded the highest average
classification accuracies (98.31%) in comparison with the other five state-of-the-art signal processing
methodologies when it is trained and validated on datasets corresponding to separate rotating speeds.
The methodology I extracted the feature vectors of all four speeds for classification by the
OAMCSVM directly from the raw vibration signal (OutLPF signal), in which non-linear and
non-stationary signals drown out the informative signal. Accordingly, those results are distributed
chaotically among the four classes, producing the lowest accuracy among the 6 methodologies (62.63%).
For methodologies II, III, and IV, the vibration signals change with the different characteristics of the
gearbox vibration signal (its amplitude and phase modulation signal), so their classification accuracy is
also low, around 70%. Methodology V (the EMD technique) is outstanding in comparison with the
first four approaches (82.5%) because it extracts IMFs, which contain fault-related information to better
discriminate between classes. However, IMFs can be mistakenly extracted from noise components,
which damaged the accuracy compared with the ANR-GRS technique by around 15%.
In addition, as a quantitative evaluation, we present the space distribution in a 3-dimensional
visualization (Figure 15) of samples belonging to four classes based on some features extracted from
the outLPF signal and the outANRsignal (signals before and after using the ANR-GRS technique,
respectively). The features of the outANR signal show better separation and clustering for different
health states of the gearbox fault diagnosis experimental scheme. Samples from the same class are
more closely clustered, whereas samples from different classes are discriminated and easy to classify.
On the contrary, before using the ANR-GRS, the features of different classes overlap, making it difficult
to distinguish the fault classes.
 (62.63%). For methodologies II, III, and IV, the vibration signals change with the different
characteristics of the gearbox vibration signal (its amplitude and phase modulation signal), so their
classification accuracy is also low, around 70%. Methodology V (the EMD technique) is outstanding
in comparison with the first four approaches (82.5%) because it extracts IMFs, which contain fault-
related information to better discriminate between classes. However, IMFs can be mistakenly
extracted
Sensors 2020, 20,from
3105 noise components, which damaged the accuracy compared with the ANR-GRS 21 of 25
technique by around 15%.

Figure 14.14.The
Figure Theaccuracy
accuracyof
ofeach
eachclass
class and
and the average accuracy
the average accuracyofofthe
thestate-of-the-art
state-of-the-art methodologies
methodologies
Sensorsand
2019,
the
and 19, xproposed
FOR PEER
proposed
the REVIEW methodology.
ANR-GRS
ANR-GRS methodology. 22 of 25

In addition, as a quantitative evaluation, we present the space distribution in a 3-dimensional

visualization (Figure 15) of samples belonging to four classes based on some features extracted from
the outLPF signal and the outANRsignal (signals before and after using the ANR-GRS technique,
respectively). The features of the outANR signal show better separation and clustering for different
health states of the gearbox fault diagnosis experimental scheme. Samples from the same class are
more closely clustered, whereas samples from different classes are discriminated and easy to classify.
On the contrary, before using the ANR-GRS, the features of different classes overlap, making it
difficult to distinguish the fault classes.
Moreover, confusion matrixes are shown in Figure 16 to demonstrate the reliability of the
varying-speed gearbox fault diagnosis methodology using the ANR-GRS module for effective noise
reduction. Using real-time tracking of the rotation speed (RPM) of a gearbox system, the ANR-GRS
generated speed-related function signals for real-time tracking of speed-dependent noise
components, and the optimized output signal was unaffected by speed during classification.
(a) (b)

Figure
Figure 15.
15.Three-dimensional
Three-dimensional visualization of features
visualization extracted
of features fromfrom
extracted (a) the
(a)input signalsignal
the input of the ANR-
of the
GRS module
ANR-GRS and (b)
module the(b)
and output signal signal
the output of the of
ANR-GRS module.
the ANR-GRS module.

Moreover, confusion matrixes are shown in Figure 16 to demonstrate the reliability of the
varying-speed gearbox fault diagnosis methodology using the ANR-GRS module for effective noise
reduction. Using real-time tracking of the rotation speed (RPM) of a gearbox system, the ANR-GRS
generated speed-related function signals for real-time tracking of speed-dependent noise components,
and the optimized output signal was unaffected by speed during classification.

(a) (b)
 (a) (b)

Figure 15. Three-dimensional visualization of features extracted from (a) the input signal of the ANR-
SensorsGRS 20, 3105and (b) the output signal of the ANR-GRS module.
2020,module 22 of 25

(a) (b)

(c) (d)

16. Confusion matrixes for four classification cases according to the speed dataset input for
Figure 16.
training: (a) 300 RPM, (b) 600 RPM, (c) 900 RPM, (d) 1200 RPM.

5. Conclusions
In this study we propose a reliable fault diagnosis methodology for gearbox systems under varying
speed conditions. It integrates adaptive noise control to significantly reduce noise with machine
learning classification to classify the fault states of the gearboxes. First, we created a set of Gaussian
reference signals that are a function of the rotation speed and consist of many noise components such
as white noise and band noise that are correlated to the parasitic noise in the vibration signals and
independent of the intrinsic informative components. Then, we applied those GRSs to an adaptive
noise control technique that produced an optimal sub-band as output for each 1-s vibration sample.
The most optimal sub-bands were then used in the feature pool configuration to extract feature vectors,
and an OAOMCSVM was used for classification. The experimental results indicated that the proposed
gearbox fault diagnosis methodology achieved the highest classification accuracies in both experiments
that are equal to 99.7% and 98.31% while significantly outperforming the counterpart state-of-the-art
methodologies used for the comparison. In future research, we will continue improving the robustness
of the proposed methodology and investigate it’s applicability to the real-time fault diagnosis scenarios.
Sensors 2020, 20, 3105 23 of 25

Author Contributions: All the authors contributed equally to the conception of the idea, the design of the
experiments, the analysis and interpretation of results, and the writing and improvement of the manuscript.
All authors have read and agreed to the published version of the manuscript.
Funding: This research was financially supported by the Ministry of Trade, Industry & Energy (MOTIE) of
the Republic of Korea and Korea Institute for Advancement of Technology (KIAT) through the Encouragement
Program for The Industries of Economic Cooperation Region. (P0006123)
Conflicts of Interest: The authors declare no conflict of interest.

References
1. MacFadden, P.D.; Smith, J.D. An Explanation for the Asymmetry of the Modulation Sidebands about the
Tooth Meshing Frequency in Epicyclic Gear Vibration. Proc. Inst. Mech. Eng. Part C J. Mech. Eng. Sci. 1985,
199, 65–70. [CrossRef]
2. McNames, J. Fourier series analysis of epicyclic gearbox vibration. J. Vib. Acoust. Trans. ASME 2002, 124,
150–160. [CrossRef]
3. Fakhfakh, T.; Chaari, F.; Haddar, M. Numerical and experimental analysis of a gear system with teeth defects.
Int. J. Adv. Manuf. Technol. 2005, 25, 542–550. [CrossRef]
4. Chaari, F.; Bartelmus, W.; Zimroz, R.; Fakhfakh, T.; Haddar, M. Gearbox vibration signal amplitude and
frequency modulation. Shock Vib. 2012, 19, 635–652. [CrossRef]
5. Mills, R. An Introduction to Machinery Analysis and Monitoring. Comput. Eng. 1991, 10, 314–315. [CrossRef]
6. Ghodake, S.B.; Mishra, P.A.K.; Deokar, P.A.V. A Review on Fault Diagnosis of Gear-Box by Using Vibration
Analysis Method. IPASJ Int. J. Mech. Eng. 2016, 4, 31–35.
7. Patil, C.R.; Kulkarni, P.P.; Sarode, N.N.; Shinde, K.U. Gearbox Noise & Vibration Prediction and Control.
Int. Res. J. Eng. Technol. 2017, 4, 873–877.
8. Boyat, A.K.; Joshi, B.K. A Review Paper: Noise Models in Digital Image Processing. Signal Image Process. Int.
J. 2015, 6, 63–75. [CrossRef]
9. Randall, R.B. Frequency analysis, 3rd ed.; Bruel & Kjaer: Nairobi, Denmark, 1987.
10. Matz, G.; Hlawatsch, F. Wigner distributions (nearly) everywhere: Time-frequency analysis of signals,
systems, random processes, signal spaces, and frames. Signal Process. 2003, 83, 1355–1378. [CrossRef]
11. Zhang, W.; Wang, H.; Teng, R.; Xu, S. Application of rank-order morphological filter in vibration signal
de-noising. In Proceedings of the 3rd International Congress on Image and Signal Processing, Yantai, China,
16–18 October 2010. [CrossRef]
12. Antoni, J. Cyclostationarity by examples. Mech. Syst. Signal Process. 2009, 23, 987–1036. [CrossRef]
13. Randall, R.B.; Antoni, J.; Chobsaard, S. The relationship between spectral correlation and envelope analysis
in the diagnostics of bearing faults and other cyclostationary machine signals. Mech. Syst. Signal Process.
2001, 15, 945–962. [CrossRef]
14. Touti, W.; Salah, M.; Ben Salem, S.; Bacha, K.; Chaari, A. Spur gearbox mixed fault detection using vibration
envelope and motor stator current signatures analysis. In Proceedings of the 17th International Conference
on Sciences and Techniques of Automatic Control and Computer Engineering (STA), Sousse, Tunisia, 19–21
December 2016. [CrossRef]
15. Jiang, R.; Liu, S.; Tang, Y.; Liu, Y. A novel method of fault diagnosis for rolling element bearings based on
the accumulated envelope spectrum of the wavelet packet. JVC/Journal Vib. Control 2015, 21, 1580–1593.
[CrossRef]
16. Xie, H.; Lin, J.; Lei, Y.; Liao, Y. Fast-varying AM-FM components extraction based on an adaptive STFT.
Digit. Signal Process. A Rev. J. 2012, 22, 664–670. [CrossRef]
17. Oppenheim, A.V.; Schafer, R.W. Discrete-Time Signal Processing; Prentice Hall: Englewood Cliffs, NJ, USA, 1989.
18. Vetterli, M.; Kovačević, J. Wavelets and Subband Coding; Prentice Hall PTR: Upper Saddle River, NJ, USA, 1995.
19. Olkkonen, J. Discrete Wavelet Transforms—Theory and Applications; Intech: Rijeka, Croatia, 2011;
ISBN 9789533071855.
20. Liu, J. Shannon wavelet spectrum analysis on truncated vibration signals for machine incipient fault detection.
Meas. Sci. Technol. 2012, 23, 055604. [CrossRef]
21. Aharamuthu, K.; Ayyasamy, E.P. Application of discrete wavelet transform and Zhao-Atlas-Marks transforms
in non stationary gear fault diagnosis. J. Mech. Sci. Technol. 2013, 27, 641–647. [CrossRef]
Sensors 2020, 20, 3105 24 of 25

22. Kang, M.; Kim, J.; Kim, J.M.; Tan, A.C.C.; Kim, E.Y.; Choi, B.K. Reliable fault diagnosis for low-speed bearings
using individually trained support vector machines with kernel discriminative feature analysis. IEEE Trans.
Power Electron. 2015, 30, 2786–2797. [CrossRef]
23. Huang, N.E.; Shen, Z.; Long, S.R.; Wu, M.C.; Snin, H.H.; Zheng, Q.; Yen, N.C.; Tung, C.C.; Liu, H.H. The
empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series
analysis. Proc. R. Soc. A Math. Phys. Eng. Sci. 1998, 454, 903–995. [CrossRef]
24. Loutridis, S.J. Damage detection in gear systems using empirical mode decomposition. Eng. Struct. 2004, 26,
1833–1841. [CrossRef]
25. Zhang, C.; Peng, Z.; Chen, S.; Li, Z.; Wang, J. A gearbox fault diagnosis method based on frequency-modulated
empirical mode decomposition and support vector machine. Proc. Inst. Mech. Eng. Part C J. Mech. Eng. Sci.
2018, 232, 369–380. [CrossRef]
26. Chen, K.; Zhou, X.C.; Fang, J.Q.; Zheng, P.F.; Wang, J. Fault Feature Extraction and Diagnosis of Gearbox
Based on EEMD and Deep Briefs Network. Int. J. Rotating Mach. 2017, 2017. [CrossRef]
27. Buzzoni, M.; Mucchi, E.; D’Elia, G.; Dalpiaz, G. Diagnosis of Localized Faults in Multistage Gearboxes:
A Vibrational Approach by Means of Automatic EMD-Based Algorithm. Shock Vib. 2017, 2017. [CrossRef]
28. Liu, B.; Riemenschneider, S.; Xu, Y. Gearbox fault diagnosis using empirical mode decomposition and Hilbert
spectrum. Mech. Syst. Signal Process. 2006, 20, 718–734. [CrossRef]
29. Goharrizi, A.Y.; Sepehri, N. Internal leakage detection in hydraulic actuators using empirical mode
decomposition and hilbert spectrum. IEEE Trans. Instrum. Meas. 2012, 61, 368–378. [CrossRef]
30. Han, G.D.; Wan, S.T.; Lv, Z.J.; Liu, R.H.; Wang, J.; Tang, G.J. The analysis of gearbox fault diagnosis research
based on the EMD and hilbert envelope demodulation. Adv. Mater. Res. 2014, 926–930, 1800–1805. [CrossRef]
31. Peng, Z.K.; Tse, P.W.; Chu, F.L. A comparison study of improved Hilbert-Huang transform and wavelet
transform: Application to fault diagnosis for rolling bearing. Mech. Syst. Signal Process. 2005, 19, 974–988.
[CrossRef]
32. Lin, J.; Dou, C.; Wang, Q. Comparisons of MFDFA, EMD and WT by neural network, Mahalanobis distance
and SVM in fault diagnosis of gearboxes. Sound Vib. 2018, 52, 11–15. [CrossRef]
33. Yang, Q.; An, D. EMD and wavelet transform based fault diagnosis for wind turbine gear box. Adv. Mech.
Eng. 2013, 2013. [CrossRef]
34. Zamanian, A.H.; Ohadi, A. Gear fault diagnosis based on Gaussian correlation of vibrations signals and
wavelet coefficients. Appl. Soft Comput. J. 2011, 11, 4807–4819. [CrossRef]
35. Van, M.; Kang, H.J.; Hin, K.S. Rolling element bearing fault diagnosis based on non-local means de-noising
and empirical mode decomposition. IET Sci. Meas. Technol. 2014, 8, 571–578. [CrossRef]
36. Guo, Y.; Liu, Q.N.; Wu, X.; Na, J. Gear fault diagnosis based on narrowband demodulation with frequency
shift and spectrum edit. Int. J. Eng. Technol. Innov. 2016, 6, 243–254.
37. Raj, A.S.; Murali, N. Morlet wavelet UDWT denoising and EMD based bearing fault diagnosis. Electronics
2013, 17, 1–8. [CrossRef]
38. Wang, X.; Zi, Y.; He, Z. Multiwavelet denoising with improved neighboring coefficients for application on
rolling bearing fault diagnosis. Mech. Syst. Signal Process. 2011, 25, 285–304. [CrossRef]
39. Haykin, S.S. Adaptive Filter Theory, 4th ed.; Prentice Hall: Upper Saddle River, NJ, USA, 2002; ISBN 0130901261.
40. Lee, K.A.; Gan, W.S.; Kuo, S.M. Subband Adaptive Filtering: Theory and Implementation; John Wiley and Sons:
Chichester, UK, 2009; ISBN 9780470516942.
41. Rauber, T.W.; De Assis Boldt, F.; Varejão, F.M. Heterogeneous feature models and feature selection applied to
bearing fault diagnosis. IEEE Trans. Ind. Electron. 2015, 62, 637–646. [CrossRef]
42. Liu, R.; Yang, B.; Zio, E.; Chen, X. Artificial intelligence for fault diagnosis of rotating machinery: A review.
Mech. Syst. Signal Process. 2018, 108, 33–47. [CrossRef]
43. Liu, J.; Zio, E. Feature vector regression with efficient hyperparameters tuning and geometric interpretation.
Neurocomputing 2016, 218, 411–422. [CrossRef]
44. Widodo, A.; Kim, E.Y.; Son, J.D.; Yang, B.S.; Tan, A.C.C.; Gu, D.S.; Choi, B.K.; Mathew, J. Fault diagnosis of
low speed bearing based on relevance vector machine and support vector machine. Expert Syst. Appl. 2009,
36, 7252–7261. [CrossRef]
45. Vapnik, V.N. The Nature of Statistical Learning Theory; Springer: New York, NY, USA, 1995.
46. Cristianini, N.; Shawe-Taylor, J. An Introduction to Support Vector Machines and Other Kernel-Based Learning
Methods; Cambridge University Press: Cambridge, UK, 1999; ISBN 0521780195.
Sensors 2020, 20, 3105 25 of 25

47. Hsu, C.W.; Lin, C.J. A comparison of methods for multiclass support vector machines. IEEE Trans. Neural
Networks 2002, 13, 415–425. [CrossRef]
48. Fan, X.; Zuo, M.J. Gearbox fault detection using Hilbert and wavelet packet transform. Mech. Syst. Signal
Process. 2006, 20, 966–982. [CrossRef]
49. Figliola, R.S.; Beasley, D.E. Sampling, digital devices, and data acquisition. In Theory and Design for Mechanical
Measurements, 5th ed.; John Wiley & Sons: Hoboken, NJ, USA, 2011; pp. 260–303.
50. Widrow, B.; Stearns, S.D. Adaptive Signal Processing; Prentice Hall: Upper Saddle River, NJ, USA, 1985.
51. Caesarendra, W.; Tjahjowidodo, T. A review of feature extraction methods in vibration-based condition
monitoring and its application for degradation trend estimation of low-speed slew bearing. Machines 2017,
5, 21. [CrossRef]
52. Rodríguez, J.D.; Pérez, A.; Lozano, J.A. Sensitivity Analysis of k-Fold Cross Validation in Prediction Error
Estimation. IEEE Trans. Pattern Anal. Mach. Intell. 2010, 32, 569–575. [CrossRef]

© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).