Available online at www.sciencedirect.

com

ScienceDirect
Procedia Engineering 97 (2014) 2092 – 2098

12th GLOBAL CONGRESS ON MANUFACTURING AND MANAGEMENT, GCMM 2014

Fault diagnosis of automobile gearbox based on machine learning

techniques
T Praveenkumara,*, M Saimuruganb, P Krishnakumarc, K I Ramachandrand
a
Research Scholar, Department of Mechanical Engineering, Amrita Vishwa Vidyapeetham, Coimbatore 641 112, India
b,c
Assistant Professor, Department of Mechanical Engineering, Amrita Vishwa Vidyapeetham, Coimbatore 641 112, India
d
Professor, Department of Mechanical Engineering, Amrita Vishwa Vidyapeetham, Coimbatore 641 112, India

Abstract

Gearbox is an essential device employed in industries to vary speed and load conditions according to the requirements. More
advancement in its design and operation leads to increase in industrial applications. The failure in any of the components of
gearbox can lead to production loss and increase maintenance cost. The component failure has to be detected earlier to avoid
unexpected breakdown. Vibration measurements are used to monitor the condition of the machine for predictive maintenance and
to predict the gearbox faults successfully. This paper addresses the use of vibration signal for automated fault diagnosis of
gearbox. In the experimental studies, good gears and face wear gears are used to collect vibration signals for good and faulty
conditions of the gearbox. Each gear is tested with two different speeds and loading conditions. The statistical features are
extracted from the acquired vibration signals. The extracted features are given as an input to the support vector machine (SVM)
for fault identification. The Performance of the fault identification system using vibration signals are discussed and compared.

©2014
© 2014Published
The Authors. Published
by Elsevier byisElsevier
Ltd. This Ltd. article under the CC BY-NC-ND license
an open access
(http://creativecommons.org/licenses/by-nc-nd/3.0/).
Selection and peer-review under responsibility of the Organizing Committee of GCMM 2014.
Selection and peer-review under responsibility of the Organizing Committee of GCMM 2014

Keywords:Fault diagnosis; Gearbox; Statistical features; Support Vector Machine (SVM).

* Corresponding author. Tel.:+91-9942948046.

E-mail address: praveent.mtp@gmail.com

1877-7058 © 2014 Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/3.0/).
Selection and peer-review under responsibility of the Organizing Committee of GCMM 2014
doi:10.1016/j.proeng.2014.12.452
 T. Praveenkumar et al. / Procedia Engineering 97 (2014) 2092 – 2098 2093

1. Introduction

Gears are very critical and a controlling component in power transmission systems, it is having wide area of
application in the field of rotating machinery. In recent years the operational accuracy of gears are increasing as
demand increases. Some of the minor errors like gear misalignment, backlash, and tooth cracks after running for
some period of time this defect will create some form of vibration which will affect the supporting components like
bearings, shafts to fail. So it is very important to monitor the condition of the gears to avoid run time breakdown of
the machine. The time-domain diagnostic algorithm is developed for monitoring the gear faults which has
improvement in the extraction capability [1].
Vibration is the behaviour of machine’s mechanical component as they react to internal or external forces.Vibration
signal analysis is mainly employed as it is an effective approach in detecting and diagnosis of faults in gears in a
power transmission system [2]. Continuous monitoring of gearbox vibration signals detects the deterioration due to
propagation of faults which tells the defect in gear, in order to schedule the proper shutdown to avoid unexpected
shutdown cost and time [3]. Excess load, manufacturing error and installation error affects the dynamic behaviour of
gears in transmissions [4]. The fault in gear creates periodic impulses which in turn transmitted to adjacent parts. So
the vibration signal should be detected early to avoid failure of whole system [5]. The future of condition based
maintenance research would be a design of intelligent device, which continuously monitor its own health using
online data acquisition, online signal processing and online diagnostic tools [6]. Compared to other monitoring
techniques vibration signals are better and it is a very popular technique to reduce cost and enable diagnosis [7].
SVM is a new generation learning method based on statistical learning theory. Due to its better performance and
accuracy, SVM is still best in feature classification for machine condition monitoring. SVM is mainly used for its
separation of classification with a large margin which reduces the generalization error. Comparison of ANN & SVM
for its feature selection from time-domain vibration signal [8]. After investigation of different vibration signal and
signal pre-processing technique it is concluded that SVM have been found to be better than ANN for the entire
feature set and the training time for SVM is also less than ANN. For wind speed application SVM outperforms over
MLP (NN) on actual and predicted data [9]. The popularity of using SVM is high, but when it relates to the
researchers application is rare for machine condition monitoring and SVM is moving towards the expertise
orientation side [10].
In this work, a tri-axial piezoelectric accelerometer is fixed on the Gearbox to acquire vibration signal for good gear
and a faulty gear. Statistical features were extracted from the acquired data and it is passed to the Support Vector
Machine for feature classification. The methodology flow chart is shown in figure. 1.

Accelerometer on gearbox

Data acquisition and signal conditioning

Statistical feature extraction from vibration data

Training of SVM using statistical features

Gearbox fault diagnosis using trained SVM

Fig.1. Methodology flow chart.

2094 T. Praveenkumar et al. / Procedia Engineering 97 (2014) 2092 – 2098

2. Experimental studies

The experimental studies include experimental set-up and experimental procedure.

2.1. Experimental set-up

The experimental setup comprises of 3 major equipment’s- Motor, Gearbox and Dynamometer. Motor power is used
as a input for gearbox drive and it has separate control panel for varying motor speed from 50 rpm to 1440rpm and it
also consists of a data logger in it to record and store O/P current, voltage and frequency values according to the
varying speed and operating load conditions.

Fig. 2.Experimental set-up along with control panel and torque controller.

Gearbox is made up of 4 set of gears with 4 different speed ranges. Gear shifter lever is employed in the gearbox for
changing the gears on demand. Eddy current dynamometer is coupled on the gearbox output shaft to apply load to
the gearbox, and its load can be varied using torque controller in terms of Kg-m.

Fig. 3.Accelerometer on gearbox.

 T. Praveenkumar et al. / Procedia Engineering 97 (2014) 2092 – 2098 2095

2.2. Experimental procedure

The experiment was conducted based on the Experimental conditions. 2 motor input speeds, 2 loads, 4 gear speeds
and 2 gear fault conditions with total of 32 test conditions has been carried out as mentioned in Experimental
conditions Table 1. Piezo electric accelerometer is used to acquire vibration data and it is temporarily fixed at the
top of gearbox using additive glues. The output from accelerometer is connected to SO analyser DAQ.

Fig. 4. Face wear introduced on 1st gear.

Table 1.Experimental conditions.

Gear type Engaged Gear Torque (N-m) Speed (rpm)
500
0
750
1st
500
5
750
500
0
750
2nd
500
5
750
Good gear
500
0
750
3rd
500
5
750
500
0
750
4th
500
5
750
500
0
750
1st
500
5
750
500
0
750
2nd
500
5
750
Face wear in 1st gear
500
0
750
3rd
500
5
750
500
0
750
4th
500
5
750
2096 T. Praveenkumar et al. / Procedia Engineering 97 (2014) 2092 – 2098

Initially the gearbox is fitted with all good gears and the test was started with motor input speed of 500rpm and
gearbox engaged with 1st gear with load of 0-Nm. There should be some consistency in the speed and load
operation, so the equipment is allowed to run for some time with that particular condition to make consistent
operation. After consistency the data acquisition will start and acquisition will carry for a time period of 300seconds
(5minutes) in the direction of Y-axis and the total of 8192 samples per second will be stored for further operations.
The same procedure will be followed for remaining 31 experimental conditions.

3. Statistical features

Feature extraction is a special form of dimensionality reduction of signal from original signal. Normally the input
signal need to be analysed is too large and it is very tough to process this huge data’s. So transforming the input
signal into a set of features is called feature extraction. Involvement of more number of variables is a major problem
in performing analysis of complex data. Feature extraction is used to combine variables to reduce these problems.
Time domain statistical features such as mean, median, mode, root mean square, standard deviation, variance,
skewness, kurtosis, maximum and minimum is used in this work.

4. Support vector machine (SVM)

SVM is a universal learning machine performs classification by N-dimensional hyper-plane and it produces a better
classification efficiency. SVM has been used extensively because of the use of well-established pattern recognition
approach. The principle of two class approach is presented in paper [11]. SVM creates hyper-plane between two sets
of data. The figure 5 clearly shows two class problems, where x denotes class ‘A+’ and o denotes class ‘A-’. SVM
try to place two boundary planes as far as possible and the plane orientation can be changed such a way to maximize
the margin between two planes and to reduce the quantisation error. The hyper-plane will be placed in between two
boundary planes. The feature present near the boundary planes are called support vectors. It is unable to separate
two classes properly in the input space of linear boundary. By transforming the feature into higher dimensional
space it is possible to create a plane that allows linear separation of higher dimensions [12].

Fig. 5.SVM classification method [M. Saimurugan et al., 13].

5. Results and discussions

As mentioned in Table 1the test has been carried for good and face wear gears with considering all other parameter
such as load, speed etc. The obtained time-domain features will be stored and that huge amount of data will be
extracted using Mat lab with the help of extracting variables such as Mean, Median, Mode, RMS, Kurtosis,
 T. Praveenkumar et al. / Procedia Engineering 97 (2014) 2092 – 2098 2097

Skewness, Maximum, Minimum, Variance, and Standard deviation. These extracted data contain values separately
for all experimental conditions.
Table 2.SVM confusion matrix for gear 1.

Category FG50010 FG50015 FG75010 FG75015 GG50010 GG50015 GG75010 GG75015

FG50010 98 0 1 0 0 1 0 0

FG50015 0 100 0 0 0 0 0 0

FG75010 2 0 89 0 0 0 9 0

FG75015 0 0 0 100 0 0 0 0

GG50010 0 0 0 0 100 0 0 0

GG50015 0 0 0 0 0 100 0 0

GG75010 0 0 14 0 0 0 86 0

GG75015 0 0 0 0 0 0 0 100

The extracted features for 1st gear with good and faulty gear are combined as a single file by representing the all
conditions in it. Then this file is analysed for SVM classification with training and testing of data files. The SVM
classification efficiency is also calculated with the help of confusion matrix as mentioned in Table 2. In the Table 2
FG denotes the Faulty Gear (Face wear) GG denotes the Good Gear. 0 N-m and 5 N-m are the two loading
parameters considered.
FG 500 1 0

Faulty Gear 500rpm 1 st gear 0 N-m load

The sample testing result of the classifier for Gear 1 is represented in the form of the confusion matrix in Table 2.
The following were interpreted from the confusion matrix.

The diagonal elements in the confusion matrix represent the number of correctly classified instances. The element in
the first row and first column represents the number of data points belongs to class ‘FG50010’ i.e. Faulty gear and
has been classified as Faulty Gear, but two instances misclassified to class FG75010 and GG50015 each.
The second element in the second row shows the number of data points belonging to class ‘FG50015’ but has been
correctly classified in class ‘FG50015’ i.e. with 100% efficiency. Similarly all the eight data sets may have some
misclassified and its classification efficiency is calculated. The same procedure is repeated for gear 2, gear 3, and
gear 4.

Table 3.Classification efficiency.

Gears Classification Efficiency (%)
Gear1 96.62
Gear2 92.375
Gear3 98.75
Gear4 100

The classification efficiency for all the 4 gears is mentioned in Table 3. Gear 1 consists of correctly classified
2098 T. Praveenkumar et al. / Procedia Engineering 97 (2014) 2092 – 2098

instances of 96.62%. Gear 2 classification efficiency is lesser compared to other gears. Gear4 provides a 100%
classification efficiency without misclassification. For a two class problem the classification efficiency of SVM is
quiet impressive and it can be used for automated fault diagnosis.

6. Conclusion

Fault diagnosis of automobile gearbox based on machine learning techniques is one of the important studies in the
field of condition monitoring of rotating machinery. Classification efficiency of good and faulty gear with two
different speeds and load conditions was carried out and the classification efficiency of four gears was quoted.
Support Vector Machine shows better classification ability in identification of a various faults in the gearbox and it
can be used for automated fault diagnosis.

Acknowledgement
The authors are grateful to the Defence Research and Development Organisation (DRDO) for the funding to carry
out the research work on fault diagnosis of dynamic mechanical systems.

References
[1]Liu Hong, Jaspreet Singh Dhupia, A time domain approach to diagnose gearbox fault based on measured vibration signals, J. SoundVib.
333(2014)2164–2180.
[2] Juliang Yin, Wenyi Wang, ZhihongManc, SuiyangKhoo, Statistical modeling of gear vibration signals and its application to detecting and
diagnosing gear faults, Inf. Sci. 259 (2014) 295–303.
[3] Omar D. Mohammed, MattiRantatalo, Jan-OlovAidanpaa, UdayKumar, Vibration signal analysis for gear fault diagnosis with various crack
progression scenarios, Mech. Syst. Sig. Process. 41 (2013) 176–195.
[4] F.Chaari, T.Fakhfakh, M.Haddar,Analyticalmodeling of spur gear tooth crack and influence on gear mesh stiffness, Eur. J. Mech. A. Solids.
28(2009) 461–468.
[5] Dae-HoonSeon, Jong-HoonJeon,Yang-Hann Kim, A novel sensing method of fault in moving machine, Mech. Syst. Sig. Process. 45 (2014)
154–169.
[6] Andrew K.S. Jardine, Daming Lin, DraganBanjevic, A review on machinery diagnostics and prognosticsimplementing condition-based
maintenance, Mech. Syst. Sig. Process. 20 (2006) 1483–1510.
[7] G.Diwakar, Dr. M. R.S.Satyanarayana, P. Ravi Kumar, Detection of Gear fault using vibration analysis, Int. J. Emerg. Technol. Adv. Eng. 2
(9) (2012)ISSN 2250-2459.
[8] B. Samanta, K.R. Al-Balushi, S. A. Al-Araimi, Artificial neural networks and support vector machines with genetic algorithm for bearing
fault detection, Eng. Appl. Artif. Intell. 16 (2006) 657–665.
[9]M.A. Mohandes, T. O. Halawani, S. Rehman, Ahmed A. Hussain, Support vector machines for wind speed prediction, Renewable Energy.29
(2004) 939–947.
[10]AchmadWidodo, Bo-Suk Yang, Tian Han, Combination of independent component analysis and support vector machines for intelligent
faults diagnosis of induction motors, Expert Syst. Appl. 32 (2007) 299–312.
[11] Victor L. Brailovsky, OfirBarzilay, Rabin Shahave, On global, local, mixed and neighborhood kernels for supportvector machines, Pattern
Recognit. Lett.20 (1999) 1183-1190.
[12] L. B. Jack, A. K. Nandi, Fault detection using support vector machines and artificial neural networks augmented by genetic algorithms,
Mech. Syst. Sig. Process. 16(2–3) (2002) 373–390.
[13] M. Saimurugan, K.I. Ramachandran, V. Sugumaran, N.R. Sakthivel, Multi component fault diagnosis of rotational mechanical system
basedon decision tree and support vector machine, Expert Syst. Appl.38 (2011) 3819–3826.