Gear Fault Diagnosis and Classification Using Data Vibration: October 2017

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/321334684
Gear fault diagnosis and classiﬁcation using data vibration
Article · October 2017
CITATIONS READS
4 473
3 authors, including:
Yassine Elyassami
Mohammadia School of Engineers
3 PUBLICATIONS 14 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Gear and Bearing faults classification based vibration data extraction View project
All content following this page was uploaded by Yassine Elyassami on 13 March 2018.
The user has requested enhancement of the downloaded file.

VOL. 12, NO. 19, OCTOBER 2017 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2017 Asian Research Publishing Network (ARPN). All rights reserved.
www.arpnjournals.com
GEAR FAULT DIAGNOSIS AND CLASSIFICATION USING

DATA VIBRATION
Yassine Elyassami, Khalid Benjelloun and Mohammed Er-Rajouany
Electrical Department, LA2I, Mohamamdia Engineer’s School, Mohammed V University, Agdal, Rabat, Morocco
E-Mail: elyassami@yahoo.fr
ABSTRACT
Gear is considered as a critical component in machinery elements, their failure caused an unexpected disturbance
at industrial processes. Many researchers studied diagnose of gear faults by vibration data analysis. In this regard, we
create and develop a vibration database from an industrial plant, then we apply many methods to extract features and we
classify gear-faults based on different algorithms. We consider four gear fault classes: healthy gear, with pinion defect,
with wheel defect and with both pinion and wheel defects. We perform diagnosis using temporal and spectral analysis, than
we improve the fault classification results using appropriate feature extraction combined to nonlinear classifiers. The
excellent classification scores in the experimental phase proofs the effectiveness of the proposed methods.
Keywords: machine vibration, gear condition monitoring, faults diagnosis, classification, feature extraction.
1. INTRODUCTION section 4 exposed the obtained results and discussions.

Gears are rotating machine part needed for Finally, section 5 gives conclusions and remarks.
motion transmission in industrial applications such as
conveyors, wind power generators, elevators and pumps. 2. THE RESEARCH METHOD
There failures caused unscheduled downtime leading high
productivity losses and huge maintenance costs. Vibration 2. 1 Feature extraction
analysis can be considered as a useful way to diagnose and To diagnose gear faults, we process in three
classify typical faults [1], such as pinion and wheel faults, phases (see Figure-1): Data acquisition, a learning phase
cracked shaft, misalignment, unbalance, rubbing, motor by a feature extraction, and fault condition classification.
faults and bent shaft defects. To avoid the complexity due In the first phase, we construct a gear vibration database
to gear dynamical model, we use, for faults diagnosis, data using radial accelerometers, in the second one, we
driven models based on historical data [8, 9]. The gear diagnose faults gear based Fast Fourier Transformation
diagnosis can be done in the time or in the frequency (FFT) and we extract reliable features based many
domain [1, 16]. In the first one, Root Mean Square (RMS), methods and algorithms: Temporal, spectral (based
skewness, kurtosis and crest factor are the most popular acceleration spectrum (AS), Power Spectrum (PS),
statistical time domain parameter for faults diagnosis. In Density Power Spectrum (DPS)), wavelets and α-stable
the frequency domain, the enveloping technique has been distribution.
proven to be an efficient method for gear defects These features are then used to learn several
detection. In recent years, the neural networks, genetic behavioral models corresponding to different initial states
algorithms and the wavelet transform have emerged as and operating conditions of gears (Healthy and fault
excellent process for fault detection [3, 5, 17]. states). The last phase, we detect the component’s
A gear fault diagnosis can be decomposed into condition by exploiting the learned models, in this phase
three parts: data acquisition, feature extraction, and fault we use KNN and SVM classifier algorithms, for estimate
condition classification. The aim of this paper is to present gear state and we compute the identification scores,
a methodology by which different gear defects can be classification scores and the corresponding recognition
detected and classified. We present the conception and the accuracy. The learned models are exploited to identify the
exploitation of a gear vibration database taking in an gear defect type (in our study we have considered four
industrial plant and we present also different methods used gear fault classes: Normal, pinion fault, wheel fault and
to extract and classify gear-faults. For improving the pinion/wheel faults).
classification accuracy, we use K-Nearest Neighbors
(KNN) and multi-class Support Vector Machine (SVM)
classifiers as machine learning algorithms. Those methods
are combined with feature extraction parameters:
Temporal, spectral, alpha stable and wavelet. The
experimental measures and treatment are applied to gear
radial axis. The proposed methods are built and applied to
perform to automate the fault diagnosis and classification
procedure.
We present in Section 2, the adopted approach,
the experimental phase is detailed in section 3, while
5617
We use also the K-Nearest Neighbors algorithm

(KNN) [10, 15] a machine learning algorithm used in
pattern recognition as a non-parametric method. Its input
is the k closest training examples in the feature space and
its output consists of a class membership.
2.3 Experimental section

The experiment was carried on many gears with
the characteristics shown in the Figure-2. Those gears are
entrained by a three phase’s 18.5Kw and 1500 rpm motor,
they have parallel shafts, a bubbling lubrication and 1/14
as reduction ratio. For measurement, we consider full load
condition and we use as sensors the piezoelectric
accelerometers mounted with magnetic bases fixation. The
Figure-1. Gear fault classification approach. obtained radial vibration signals of four gears have a 25.6
KHz sampling rate and the acquisition duration for each
We use for temporal, spectral and wavelet Gear is 40Sec. The same experimental platform was
extractions [2, 6], five indicators (RMS, Peak Level, Crest performed for sleeve and ball bearing [13, 14].
Factor, Kurtosis and Entropy). In each case, we apply
them to acceleration signal, acceleration spectrum; power
spectrum, density power spectrum and the third level
wavelet decomposition (see Table-1).
Table-1. Feature extraction parameters.

Signals Parameters
Temporal Acceleration
Acceleration spectrum
(AS)
RMS
Spectral Power Spectrum (PS) Peak Level
Power Spectrum Density Crest Factor
(PSD) Kurtosis
The means of the eight Entropy Figure-2. Gears characteristics.
parquets using the third
Wavelet The industrial ICP accelerometers are connected
level wavelets
decomposition (dB02) to the vibration card NI 9234 as input NI CompactRIO-
Alpha stable 9022 controller module. The NI-controller is connected to
α Stable Acceleration a computer for advanced treatment: Features extraction
parameter’s
and fault classification. LABVIEW software (Real-time &
We use also an extraction based alpha stable FPGA) is used to obtain the acceleration signal data, for
distribution [7, 12], which includes Gaussian, Cauchy fault diagnosis, feature extraction and fault classification.
and Levy distributions as special cases. This distribution The obtained acceleration and spectrum are used to
was developed from the generalized central limit theorem. identify four different conditions of gear: (a) normal
The general characteristic function for this probability healthy gear, (b) gear with pinion defect, (c) gear with
density is characterized by a four-parameter family of wheel defect, (d) gear with both pinion and wheel defects.
continuous probability distributions form roughly
corresponding to measures of asymmetry and Table-2. RMS/Crest factor/Kurtosis Means values.
concentration. We use the four parameters of alpha stable Crest
distribution as features in fault diagnosis of gear. The RMS Kurtosis
factor
calculated features represent the degradation of gear.
Healthy gear 0.48 3.67 3.15
2.2 Faults classification Pinion defect 2.22 8.03 6.88
We use as nonlinear classifiers, the multi-class Wheel defect 3.13 7.97 13.31
Support Vector Machine (SVM) [4, 11] a computational
learning method which employs structural risk to Pinion and wheel defect 2.58 8.65 6.88
minimize an upper bound on the expected risk, it’s
considered as one of the most efficient classification For isolate gear vibration signature and reduce
algorithm. effect of the bearing signature, all signals are filtered by
using a 1KHz-10KHz pass-band filter. To diagnose those
5618
gears. We consider the evaluation of some temporal variation of the RMS, Crest factor and Kurtosis
parameters like Histogram/RMS/Crest factor/Kurtosis as parameters. High value and huge variation are correlated
shown in Table-2 and Figure-3. We see the difference with the defect level.
between the four gears in means values and in curves

Figure-3. Histogram/ RMS/ Crest factor /Kurtosis, for gear fault class: (a: Green) healthy, (b: Yellow)
Gear with pinion defect, (c: Blue) Gear with wheel defect, (d: Red) Gear with pinion/wheel defects.
To diagnose and to identify specified gearbox pinion 1.8Hz (∆T=0.56sec). With sidebands (1*, 2*,
faults, we use as exposed in Figures (from 4 to 7), FFT 3*∆f=1.8Hz) around tooth mesh frequency.
spectrum based on radial acceleration and correlated with Gear acceleration and spectrum, for the third gear
gear frequency specifications. In the Figure-4 we can see condition exposed in Figure-6, are distinct by a wheel
that gear acceleration and spectrum are normal for the first defect as appear in in the rotational frequency of wheel 9.1
gear condition and indicate a healthy gear, only electrical Hz (∆T=0.109sec). With sidebands (1*, 2*, 3*∆f=9.1Hz)
network frequency 50Hz, gear mesh frequency 118.3Hz around mesh frequency.
and their corresponding harmonics (150Hz, 250Hz, The forth gear condition is characterized by a
350Hz…) and (236.6Hz, 354.9Hz, 473.2Hz…) appear serious defect level, see Figure 7, with mixed faults:
with moderates pics and without sidebands around mesh Pinion defect in 1.8Hz and wheel defect in 9.1Hz. With
frequency. sidebands (1*, 2*, 3* ∆f=1.8Hz) and (1*, 2*, 3*∆f=9.1Hz)
The second gear condition, in Figure-5 indicate a around mesh frequency.
gear pinion defect, as appear in the rotational frequency of
5619
Figure-4. Acceleration signal and its corresponding spectrum for healthy gear.
5620
Figure-5. Acceleration signal and its corresponding spectrum for gear pinion defect.
5621
Figure-6. Acceleration signal and its corresponding spectrum for gear wheel defect.
5622
Figure-7. Acceleration signal and its corresponding spectrum for gear pinion/wheel defects.
3. RESULTS AND DISCUSSIONS training windows: 10% of readings were taken for each
To show the effectiveness of the proposed gear to calculate the statistical features used as input to
methods, we compare identification and classification classification algorithms.
scores, based feature extraction (temporal, spectral, alpha The recognition identification scores results are
stable and wavelet) using the same experimental signal detailed in Table-3. It was seen that those scores are
data and processing with KNN and SVM algorithms. mostly excellent: Higher than 99% and up to 99,917%
Radial accelerometers measurement is organized as 40 sets with better scores using SVM classifier. The identification
of data for each gear. We use just a weak amount of score indicates the similarity between the input and the
5623
assigned class. It is equal to (1 - d) × 100, where d is the REFERENCES

normalized distance between the input vector and the
assigned class. [1] A. Aherwar and M. S. Khalid. 2012. Vibration
analysis techniques for gearbox diagnostic: a review.
Table-3. Recognition identification score based International Journal of Advanced Engineering
vibration data.
Technology. 3(2).
KNN SVM
[2] A. Malhi and R. X. Gao. 2003. Feature selection for
Temporal 99,097 99,875
defect classification in machine condition monitoring.
Spectral (AS) 99,662 99,889 in Conference Record IEEE Instrumentation and
Spectral (PS) 95,567 99,917 Measurement Technology Conference. pp. 36-41.
Spectral (PSD) 98,967 99.806
[3] Cerrada M.; Sánchez R.V.; Cabrera D.; Zurita G.; Li
Wavelet 99,212 99,868 C. 2015. Multi-stage feature selection by using
αStable 94,272 99,813 genetic algorithms for fault diagnosis in gearboxes
based on vibration signal. Sensors. 15, 23903-23926.
Table-4. Recognition classification score based
vibration data. [4] Debnath R., Takahidel N., Takahashi H. 2004. A
decision based one-against-one method for multi-class
KNN SVM
support vector machine. Pattern Analysis &
Temporal 95,571 99,097 Applications. 7, 164-175.
Spectral (AS) 97,569 99,083
[5] D.P. Jena, S.N. Panigrahi, R. Kumar. 2013. Gear fault
Spectral (PS) 97,222 99,057
identification and localization using analytic wavelet
Spectral (DSP) 96,554 99,148 transform of vibration signal. Measurement. 46(3):
Wavelet 96,454 99,104 1115-1124.
αStable 99,305 99,135
[6] Farag K.Omar & A.M.Gaouda. 2012. Dynamic
wavelet-based tool for gearbox diagnosis. Mechanical
The classification confidence score indicates the
degree to which the assigned class represents the input Systems and Signal Processing. 26: 190-204.
better than the other classes represent the input. It is equal
to (1 - d1/d2) × 100, where d1 is the distance to the closest [7] Gang Yu and Ningning Shi. 2012. Gear Fault Signal
class, and d2 is the distance to the second closest class. As Modeling and Detection Based on Alpha Stable
exposed in Table-4, the SVM classifier offer a very high Distribution. 2012 International Symposium on
classification score between 99,057% and 99,148%, the Instrumentation & Measurement, Sensor Network and
KNN classifier combined with the alpha stable extraction Automation (IMSNA).
give the highest score 99,305% but it gives, when
combined to author extraction methods, acceptable scores [8] Gao Z.; Cecati C.; Ding S.X. 2015. A survey of fault
between 95, 571% and 97,569%. From the above results, it
diagnosis and fault-tolerant techniques-Part I: Fault
was inferred that we can improve classification by an
appropriate choice of feature extraction and classifier diagnosis with model-based and signal-based
algorithms. approaches. IEEE Trans. Ind. Electron. 62, 3757-
3767.
4. CONCLUSIONS
In this paper we study four different gear [9] Gao Z.; Cecati C.; Ding S.X. 2015. A survey of fault
behaviours, one healthy and three with different defect diagnosis and fault-tolerant techniques-Part II: Fault
signs were used. The classification is then obtained using diagnosis with knowledge-based and hybrid/active-
some extraction methods combined with classifiers based approaches. IEEE Trans. Ind. Electron. 62,
algorithms based radial vibration acceleration data. The
3768-3774.
percentage of correct classification was higher than 95, 5%
and up to 99, 3% for our proposed feature extraction and
[10] L.I. Yong-long, S.H.A.O. Ren-ping, X.U.E. Teng.
the chosen nonlinear classifiers.
2010. Gear fault diagnosis based on wavelet and
neural network, J. Aerosp. Power. 25(1): 234-240.
[11] Manikandan J., Venkataramani B. 2010. Study and

evaluation of a multi-class SVM classifier using
5624
diminishing learning technique. Neurocomputing. 73,

1676-1685.
[12] Shao Min, Chrysostomos L.N. 1993. Signal

processing with fractional lower order moments:
stable processes and their applications [J].
Proceedings of the IEEE. 81(7): 986-1010.
[13] Y. Elyassami, K. Benjelloun and M. El Aroussi. 2016.

Sleeve Bearing Fault Diagnosis and Classification.
12: 163-170.
[14] Y. Elyassami, K. Benjelloun and M. El Aroussi. 2016.

Bearing Fault Diagnosis and Classification Based on
KDA and Alpha-Stable Fusion. Contemporary
Engineering Sciences. 9(10): 453-465.
[15] Zhou Yuqing, Sun Bingtao, Li Fengping, and Song

Wenlei. NC Machine Tools Fault Diagnosis Based on
Kernel PCA and ݇-Nearest Neighbor Using Vibration
Signals. Shock and Vibration. 2015(Article ID
139217): 10.
[16] Z. Feng, M. Liang, and F. Chu. 2013. Recent

advances in time frequency analysis methods for
machinery fault diagnosis: a review with application
examples. Mechanical Systems and Signal
Processing. 38(1): 165-205.
[17] Zhi Qiang Chen, Chuan Li and René-Vinicio

Sanchez. Gearbox Fault Identification and
Classification with Convolutional Neural Networks.
Shock and Vibration. 2015 (Article ID 390134).
5625
View publication stats

Gear Fault Diagnosis and Classification Using Data Vibration: October 2017

Uploaded by

Copyright:

Available Formats

Gear Fault Diagnosis and Classification Using Data Vibration: October 2017

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Gear Fault Diagnosis and Classification Using Data Vibration: October 2017

Uploaded by

Copyright:

Available Formats

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

Gear fault diagnosis and classiﬁcation using data vibration

Article · October 2017

The user has requested enhancement of the downloaded file.

GEAR FAULT DIAGNOSIS AND CLASSIFICATION USING

1. INTRODUCTION section 4 exposed the obtained results and discussions.

We use also the K-Nearest Neighbors algorithm

2.3 Experimental section

Table-1. Feature extraction parameters.

assigned class. It is equal to (1 - d) × 100, where d is the REFERENCES

[11] Manikandan J., Venkataramani B. 2010. Study and

diminishing learning technique. Neurocomputing. 73,

[12] Shao Min, Chrysostomos L.N. 1993. Signal

[13] Y. Elyassami, K. Benjelloun and M. El Aroussi. 2016.

[14] Y. Elyassami, K. Benjelloun and M. El Aroussi. 2016.

[15] Zhou Yuqing, Sun Bingtao, Li Fengping, and Song

[16] Z. Feng, M. Liang, and F. Chu. 2013. Recent

[17] Zhi Qiang Chen, Chuan Li and René-Vinicio

View publication stats

You might also like