Available online at www.sciencedirect.

com

ScienceDirect
Procedia Engineering 109 (2015) 73 – 79

XXIII Italian Group of Fracture Meeting, IGFXXIII

Principal component analysis for characterization of fretting wear

experiments on spline couplings
Waqar Qureshia,∗, Francesca Curáa , Andrea Muraa
a Department of Mechanical and Aerospace Engineering, Politecnico di Torino, Corso Duca degli Abruzzi 24, 10129 Torino, Italy

Abstract
Spline couplings are often used in aerospace industry for the purpose of power and torque transfer. A spline coupling typically
consists of a hub and shaft with internal and external teeth respectively. As the teeth are engaged and the spline coupling is put into
operation, repeated relative surface movement of teeth begins and results in fretting wear. Principal component analysis (PCA)
is a statistical technique to identify the important parameters and their relationships with each other. This work deals with the
characterization of fretting wear by the help of principal component analysis. PCA has been applied to the experimental data
obtained from fretting wear tests in the lab and important variables with respect to fretting wear identified. This serves as a new
tool to interpret the experimental results and hypothesis generation in the field of fretting wear.
©©2015 The Authors.
2015Published 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/4.0/).
Peer-review under responsibility of the Gruppo Italiano Frattura (IGF).
Peer-review under responsibility of the Gruppo Italiano Frattura (IGF)
Keywords: principal component analysis; fretting wear; spline couplings

1. Introduction

Spline couplings typically consist of a hub and a shaft with internal and external teeth respectively. When a spline
coupling gets into operating mode, the teeth are engaged and a repeated relative movement between the surfaces of
the teeth starts which results into fretting wear. The fretting wear depends upon material properties, misalignment
between hub and shaft, rotational speeds, friction coefficients, lubricant properties, temperatures, etc. A review of
spline couplings can be found in [1]. Often a misalignment is induced in the spline couplings which severely affects
the life of components and hence becomes an important parameter in fretting wear. An analysis of rotating misaligned
spline couplings subjected to steady torque is given in [2]. Finite element methods are a popular tool to model and
predict fretting wear. Though computationally intensive, these methods provide quite significant insights. A finite
element based simulation to predict fretting wear for spline couplings is presented in [3].
After application of PCA the results are still not readily understood and cannot be easily interpreted owing to the
complexities associated with high dimensional data. A very useful work regarding interpretation of PCA is given in
[4]. Reference [5] explains an interesting usage of PCA as a tool to analyse signals and model reduction by applying
the mechanics of minimal realization using working approximations of controllable and unobservable subspaces. It

∗ Corresponding author. Tel.: +39 011 090 6930; fax: +39 011 090 6999
E-mail address: waqar.qureshi@polito.it

1877-7058 © 2015 Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer-review under responsibility of the Gruppo Italiano Frattura (IGF)
doi:10.1016/j.proeng.2015.06.209
74 Waqar Qureshi et al. / Procedia Engineering 109 (2015) 73 – 79

is worth noting that PCA is applicable only in case of linear relations. If the presence of non-linear relationships is
highly likely then Kernel-PCA [6] is recommended. Even though the work in [7] is directed towards statistical process
control in chemical processes it still gives a useful implementation of dynamic PCA to detect disturbances and their
isolation.
A combined usage of PCA and artificial neural networks to predict ball bearing fretting damage has been given
in [8]. In the current work a dedicated test bench has been established. Tests related to fretting wear have been
performed and values of several parameters have been acquired. The whole process has resulted in almost four
thousand readings for each of the thirteen variables under consideration. This accounts for high dimensional data
which is hard to interpret and it is not very straightforward to find the relationships between different variables. Hence
the use of PCA has been proposed to investigate the underlying characteristics of the data which gave some very
interesting results. The proposed approach can be quite helpful in designing future experiments at a lower cost and in
less time.

Nomenclature

T Average temperature of lubricant (°C)

F Reaction force at the driven component (N)
L Linear Variable Differential Transducer (LVDT) Reading (mm)
R Lubricating oil flow rate (lit/min)
N Revolutions per minute (RPM) of driving motor
θ Misalignment angle between shafts in minutes ( )
V RPM of actuating motor to maintain the misalignment
τ Applied Torque (Nm)
x Mean (Average) of column X
σx Standard Deviation of X
σ xx Variance of X
σ xy Covariance of X with Y or vice versa

2. Experimental set up

The test bench operates on the principle of power recirculation hence reducing the power requirements greatly from
1,000 kW to 6.3 kW. It is capable of introducing a misalignment between hub and shaft. The test bench is 2.5 m long,
1.2 meters wide and weighs 3,500 kg approximately. Design parameters of spline coupling under test and functional
specifications of the test bench have been given in Table 1. Test sample and experimental schematics are provided in
Figure 1 and Figure 2 respectively. Various sensors to measure the important parameters have been installed in the test
bench. The control and data acquisition of different variables have been achieved using the software tool developed
in LABVIEW [9].
Reference [9] gives more details of the test bench as, ”The test bench is composed of two coaxial shafts (inner
and outer shaft), both connected on the right side to a torque generator. The inner shaft is divided into two branches
connected by means of the spline coupling test article. At the left end, the inner and outer shafts are fixed to each

Table 1. Design parameters of spline coupling and functional specifications of test bench [9]
Spline coupling design parameters
Material Teeth profile No of teeth Modulus Pressure angle Crowing radius Face width
42CrMo4 Crowned 26 1.27 mm 30° 200 mm 12 mm
Test bench functional specifications
Max Torque Max RPM Inclination Increment Oil temp Oil flow rate Motor Power
5000 Nm 2000 10’ 0.5’ max 60 °C 1 l/min 6.3 kW
 Waqar Qureshi et al. / Procedia Engineering 109 (2015) 73 – 79 75

Motor
Test Article
Belt
CV Joint

Outer Shaft Torque

θ Inner Shaft Generator

Angular
Roller Bearings Velocity
Flexible Joint
Fig. 1. Spline Coupling [9] Fig. 2. Experimental Schematics [9]

other. When the torque generator applies a load to the inner shaft, it passes through the spline coupling test article and
comes back, by means of the outer shaft, to the torque generator. This apparatus creates a torque loop, letting the test
article under load. External and internal shafts (and the test article) can rotate thanks to an electric motor connected
to the external shaft by means of a belt”.

3. PCA Review

Principal component analysis is a statistical technique to deal with high dimensional data. It helps in reducing the
dimensions and redundancy of a data set. A very friendly tutorial of PCA is provided in [10]. Below is given a brief
summary of the basic concepts necessary for implementation of PCA from the same source.

3.1. Mean

Consider a single column matrix X with i number of elements. Its mean is given as;

n
X̃ = Xi /n (1)
i=1

3.2. Standard Deviation

For the same matrix X, its standard deviation is the amount of spread in the data and mathematically is given below
with X̃ as the mean of X found in Equation 1.

 n

σx = (Xi − X̃)2 /(n − 1) (2)
i=1

3.3. Variance

Using the Equation 2 to find standard deviation and taking its square gives variance. Mathematically

n
σ xx = (Xi − X̃)2 /(n − 1) (3)
i=1

3.4. Covariance

Covariance is a measure of how much two variables change together. It is simply a reimplementation of Equation 3.
Its positive value implies a direct correlation between both variables, whereas a negative value implies that if one
76 Waqar Qureshi et al. / Procedia Engineering 109 (2015) 73 – 79

variable increases, the other one decreases. It is worth mentioning that presence of non-zero covariance does not
always guarantee a relationship. It has to be further verified. If X and Y are two single column matrices with i number
of values their covariance will be given as below.

n
σ xy = (Xi − X̃)(Yi − Ỹ)/(n − 1) (4)
i=1

The same process can be repeated for all the dimensions in a data set. It means that if a five dimensional data were
provided, there will be total twenty five possible combinations of their variance (covariance) with each other. And the
resulting matrix is called as the covariance matrix.

3.5. Correlation

When mean centred data is divided by the respective standard deviations, its covariance is termed as correlation.
Correlation matrix can found in the same fashion as covariance matrix. Mathematically after necessary modifications
in Equation 4, it may be represented as;

n
σ xy = (Xi − X̃)(Yi − Ỹ)/σ x σy (n − 1) (5)
i=1

3.6. Steps of PCA

Now that the basic terminologies have been explained, the steps of PCA are given below.

• Each column in dataset to represent one variable

• Find mean of every column
• Centre the data in all columns around their respective means
• If scales are different, divide each column by its standard deviation
• If number of variables are m, find an m × m covariance or correlation matrix
• Find Eigen vectors and Eigen values of the covariance matrix
• Choose Eigen vectors with highest Eigen values to form the feature vector (FV)
• Multiply the original dataset with FV (each new dimension is a PC)
• Find correlation of each PC with original data set using Equation 5
• Draw loading plot by treating first and second PC as x and y axis

4. Results

Experiment was performed to measure the magnitudes of average temperature of lubricant (T ), reaction force
(F), LVDT reading (L), lubricant flow rate (R), RPM of driver motor (N), misalignment angle (θ), RPM of actuator
motor (V) and the applied torque (τ). The test was run for 10 million cycles at the speed of 1450 revolutions per
minute which took almost 5 days to complete. The sampling time was 100 seconds. Initial results provided a total of
thirteen variables with more than four thousand values each. Out of these thirteen, five were temperature readings of
lubricating oil at different points. Since the oil was in circulation so an average of temperature readings was taken to
reduce the data set to eight variables. The initial results after reduction to eight variables have been shown in Figure 3.

Table 2. Comparison of data before and after PCA

Observations Variables Readings Total readings Size per observation Size in Kilobytes
Original data 8 4135 33.08k 4 kB 132.32
Data after PCA 2 4135 8.27k 4 kB 33.08
Net reduction 6 24.81 75%
 Waqar Qureshi et al. / Procedia Engineering 109 (2015) 73 – 79 77

V 1

N
R
0
θ

L
T (0→68) R(0→0.3)
F(−80→954) N(0→1.5K) F
L(0→1.6) τ(0→1.1K)
θ(0→10) V(−103→34) T
-1
0 1k 2k 3k 4k T F L θ R N τ V
Fig. 3. Variables vs Time Fig. 4. Correlation Matrix

It can be observed that still the relationships between different variables are not very clear. Normally covariance of
every variable with each other should have been found which would result in a covariance matrix. However, the units
of measurements are not same and scales are drastically dissimilar hence correlation matrix has been found instead
after scaling the data and is shown in Figure 4. The results of a correlation matrix are always symmetric.
Then Eigen values and vectors of this matrix have been calculated. The results of Eigen value analysis are shown in
Figure 5. The figure shows Eigen values of all Eigen vectors in descending order. The Eigen vectors corresponding to
each Eigen value are the principal components (PCs) and only the components whose sum is more than sixty percent
have been chosen for further analysis. Next, feature vector comprising of the first two PCs is multiplied with the
original data set to give rise to a new data set with only two dimensions. This results in a great reduction in size
of data with the possibility of full recovery of original data. The reduction in size of data has been summarized in
Table 2.
Finally correlations of both variables in the new data set with original data are calculated, their results are used as
coordinates on a plot with first and second PCs (PC1 and PC2 ) as X and Y axis respectively. This gives rise to a so
called loading plot which is shown in the Figure 6. Loading plots give a lot of information about underlying dynamics
and relationships between different variables. Variables appearing together tend to behave similarly whereas those in

40
τ
1

R
30
PC2

L
N
20
0

θ
V
10
F
T
-1

0
1 2 3 4 5 6 7 8 -1 0 PC1 1
Fig. 5. Principal components with weights Fig. 6. Loading Plot
78 Waqar Qureshi et al. / Procedia Engineering 109 (2015) 73 – 79

Table 3. Operating conditions of independent variables

Variables Rotational speed (N) Misalignment (θ) Torque (τ) Number of cycles Sampling Time Readings Lubrication
Magnitude 1450 rpm 10’ 1000 Nm 10M 100 s 4135 yes

opposite quadrants exhibit contrasting behaviours. Farther a variable from the origin on a loading plot, statistically
more important the variable is likely to be.

5. Discussions

A proportional-integral-derivative (PID) control has been developed to maintain the operating conditions of various
parameters. The parameters being controlled by PID controller are the misalignment angle (θ), driving motor rpm (N)
and actuator motor rpm (V). LVDT (L) is simply being used as a sensor in PID control to maintain the misalignment
angle. The values of these four parameters have been controlled and maintained by the help of PID control. However,
torque is applied at a fixed value initially but it is not part of any feedback loop for its value to be maintained. The
changes in values of torque as a result of fretting wear phenomenon are hence natural and expected. So reaction force
(F) at the driven end of the shaft, average temperature (T ), oil flow rate (R) and torque (tau) are the parameters with
variable values and they can be expected to give some hint about how the fretting takes place during the experiment.

5.1. Corroboration with operating conditions

In order to understand, validate and correctly use the PCA it is necessary to corroborate the interpretation of results
obtained with the practical operating conditions. Also, it is of paramount importance to validate the experimental
procedure statistically which in itself is yet another advantage. To begin with, first the operating conditions / param-
eters are detailed in Table 3. These are the independent variables with one condition that they are supposed to be
constant throughout the experiment. So out of eight variables being monitored four have been defined to be operat-
ing conditions, whereas four variables are dependent on them and should change their values as the cycles progress.
Interaction between the hub and shaft of spline coupling is responsible for variations in the value of applied torque.
These conditions are reflected well in the loading plot as explained in the next section.

5.2. Loading plot interpretation

Loading plot shows a statistical picture of variables. The variables which appear closer to the origin should be
the constant / independent variables or operating conditions, whereas those farthest from it would be the dependent
(dynamic) parameters changing their values with each cycle. To summarize the conclusions about dependence and
independence of variables based upon a priori knowledge of experimental set up and the information from loading
plot, Table 4 gives the complete picture.

5.3. Relationships between variables

The loading plot is providing much more information about relationships between different variables which have
been summarized below;

• Statistically, most significant variables are the reaction force (F), torque (τ), lubricating oil flow rate (R) and
average temperature (T ) of the lubricating oil in descending order. Other variables are insignificant.

Table 4. A priori and loading plot based information on independent and dependent variables
A priori information Loading plot information
Independent N θ τ L Independent N θ V L
Dependent V T F R Dependent τ T F R
 Waqar Qureshi et al. / Procedia Engineering 109 (2015) 73 – 79 79

• Two pairs of variables behave similarly (increase or decrease together) which are:-
- Reaction force and average temperature of lubricating oil (less similar).
- Lubricating oil flow rate and torque (more similar).
• If any variables in the current set were to be used for the purpose of characterization of fretting wear under
given conditions, they have to be the reaction force and torque or lubricating oil flow rate.

6. Conclusions

In this paper the use of PCA as a tool to interpret experiments related to fretting wear in case of spline couplings
has been presented. Identification of most significant parameters, experimental validation and isolation of insignifi-
cant variables have been successfully carried out. The proposed approach lowers the memory requirements for data
processing and still provides a way to fully study the relationships between different variables without a significant
loss of information. It has been concluded for the current experimental data that torque, reaction force, lubricating
oil flow rate and its average temperature at different points play the most important role in the experiment. So future
experiments will be designed and conducted keeping in mind these interesting observations. Even though a specific
case has been considered but still the idea is general enough to be applied to other cases too. The proposed method
provides an unconventional statistical tool for validation, verification and explanation of experimental data. It also
gives information about the parameters which deserve more attention from a mathematical modelling point of view.
Usage of such techniques together with classical fretting wear modelling approaches is expected to provide more
insight into the phenomenon of fretting fatigue and wear. It may also act as a reliable tool to find inter-dependency of
various parameters even if their physical interactions in the experiment were not fully known. In the future work, the
same technique will be used in conjunction with the data of debris formation to find the most representative variables
of fretting wear.

References

[1] V. Cuffaro, Prediction Method for the Surface Damage in Splined Couplings, Ph.D. thesis, Politecnico di Torino, Italy (2013).
URL http://porto.polito.it/2506323/1/PhD_Thesis_Vincenzo_Cuffaro.pdf 1
[2] S. Medina, a. V. Olver, An analysis of misaligned spline couplings, Proceedings of the Institution of Mechanical Engineers, Part J, Journal of
engineering tribology 216 (5) (2002) 269–279. doi:10.1243/135065002760364813. 1
[3] J. Ding, I. R. McColl, S. B. Leen, The application of fretting wear modelling to a spline coupling, Wear 262 (9-10) (2007) 1205–1216.
doi:10.1016/j.wear.2006.11.017. 1
[4] C. R. Rao, The use and interpretation of principal component analysis in applied research, Sankhyā: The Indian Journal of Statistics, Series A
(1964) 329–358.
URL http://www.jstor.org/discover/10.2307/25049339?uid=3738296&uid=2&uid=4&sid=21106021547261 1
[5] B. Moore, Principal component analysis in linear systems: Controllability, observability, and model reduction, Automatic Control, IEEE
Transactions on 26 (1) (1981) 17–32. doi:10.1109/TAC.1981.1102568. 1
[6] B. Schlkopf, A. Smola, K.-R. Mller, Kernel principal component analysis, in: W. Gerstner, A. Germond, M. Hasler, J.-D. Nicoud (Eds.),
Artificial Neural Networks ICANN’97, Vol. 1327 of Lecture Notes in Computer Science, Springer Berlin Heidelberg, 1997, pp. 583–588.
doi:10.1007/BFb0020217. 1
[7] W. Ku, R. H. Storer, C. Georgakis, Disturbance detection and isolation by dynamic principal component analysis, Chemometrics and Intelligent
Laboratory Systems 30 (1) (1995) 179–196. doi:10.1016/0169-7439(95)00076-3. 1
[8] T. Kolodziejczyk, R. Toscano, S. Fouvry, G. Morales-Espejel, Artificial intelligence as efficient technique for ball bearing fretting wear damage
prediction, Wear 268 (1) (2010) 309–315. doi:10.1016/j.wear.2009.08.016. 1
[9] V. Cuffaro, F. Curà, A. Mura, Test Rig for Spline Couplings Working in Misaligned Conditions, Journal of Tribology 136 (1) (2013) 011104.
doi:10.1115/1.4025656. 2, 1, 1, 2
[10] L. I. Smith, A tutorial on Principal Components Analysis Introduction, Statistics 51 (2002) 52. doi:10.1080/03610928808829796. 3