214 M. Cocconcelli et al.

Fig. 2. Ball bearing structure and possible damage locations: (a) rail, (b) external
bearing, (c) outer ring, (d) inner ring, (e) rolling element and (f) cage.

n
finner = fr (1 + λcosβ) (2)
2
fr 1 − (λcosβ)2
fball = (3)
2 λ
fr
fcage = (1 + λcosβ) (4)
2
d
where n is the number of rolling elements, λ = D , d is the ball diameter, D is
the pitch diameter, β is the ball contact angle and fr is the rolling frequency of
the outer ring. The characteristic fault frequencies in Eqs. 1–4 correspond to the
faults in Fig. 2(c)–(f) respectively. In the case of linear motors, two other types of
fault are present. The damage on the fixed rail (Fig. 2(a)) causes a single impact
as the bearing hits the fault on the rail surface. In case of multiple bearings
supporting the carts, there is a train of impacts whose characteristic frequency
is:
v
frail = (5)
l
Where l is the distance between contact points of two subsequent bearings
with the rail and v is the linear velocity of the bearing. Finally, the damage along
the external surface of the bearing (Fig. 2(b)), whose characteristic frequency is:
v
fsurf ace = (6)
πDout
Where v is the linear velocity and Dout is the external surface diameter.
Actually, the fault frequency can be quite different because these formulas are
computed with a kinematic model that does not take the slipping of the rolling
elements into account. Another key factor is the periodic modulation of the
vibration signal due to the load distribution. As a matter of fact the load on
the bearing can change for mechanical problems (e.g. unbalance, misalignment,
 Bearing Fault Model 215

etc.) or for the process needs (e.g. variation of the payload on the end-effector,
mechanical cams, etc.).

3 Model Implementation

This section explains the method used for the implementation of the model. In
this paper, the influence of lubrication is not taken into account, i.e. the bearing
is supposed to be correctly lubricated. The vibration signal of a localized fault
is modeled as follows [22]:
∞

x(t) = h(t − iT − τi )q(iT ) + n(t) (7)
i=−∞

Where h(t) is the impulse response to an impact on the damage point, q(t)
represents the modulation due to the load distribution, T is the interval between
two consecutive impacts, τ represents the uncertainties due to the random slip-
ping of the rolling elements and n(t) is the system noise. The model developed
can be used both with a constant velocity profile of the mover and with a vari-
able motion profile. The model takes into consideration the localized faults for
a rolling bearing, the inner race of which is fixed. For example, in the case of a
fault on the outer ring, it is possible to write Eq. 1 as the projection of the fault
on the external surface of the bearing as follows:
2π 1
Δpe = Dout (8)
n 1 − dcosβ
D
Where Δpe is the distance between two damages projected on the external
surface of the bearing. This formulation allows to consider the fault frequency
as a position variation of the bearing. In this way there is a direct correlation
between the damage of an element of the bearing and the position along the
motion profile. The same method is used for each aforementioned bearing fault
since each fault is defined as a series of equispaced impulses. The implementation
of the model can be divided into 5 steps:

1. Generate a vector S filled with zeros that represents the expected vibration
signal recorded by a sensor placed on the moving cart. The length of the
vector is equal to L. Each cell of the vector corresponds to a period of time
equal to t = f1s where fs is the sampling frequency of the sensor in Hz.
2. Resample the motion profile M (t) and the load vector q(t) at the same sam-
pling frequency fs of the sensor.
3. Place 1 in the cells of the S vector with index values defined dividing M (t)
by Δpe (Eq. 9). Δpe is a rational number and its value may be between
two samples of the motion profile. In this case an error is introduced and it
directly depends on the sampling frequency fs (the greater is fs , the lower is
the error), but it is possible to correlate each Δpe,i value with an error that
is equal to
216 M. Cocconcelli et al.

E(ti ) = |M (ti ) − Δpe,i | (9)

4. Weight the vector S with the vector of the load q(t) computed on the contact
point between the bearing and the rail.
5. Filter the generated vector with the FFT-based model with overlap-add
method, having a filter coefficient equal to the acceleration of the impulse
response of a SDOF system. The generated response of the SDOF system to
a unit impulse in the time domain is:

F/m −ζωn t
xSDOF (t) = e sin(ωd t) (10)
ωd
where F is the amplitude of the external excitation, m is the system mass,
ζ isthe damping coefficient, ωn is the natural frequency in [rad/s] and ωd =
ωn 1 − ζ 2 .

4 Numerical Validation
In this section the proposed algorithm is used for the simulation of the expected
vibration signal for a faulted bearing. The geometry of the bearing and the
rail are taken from a real application and the results will be used for future
development of condition monitoring techniques.

Fig. 3. Beckhoff XTS independent cart system with 12 carts.

The case of study takes into account the Beckhoff XTS system (Fig. 3) with
GXF Hepco Guidance and the cart shown in Fig. 1. The cart is connected to the
ellipsoid rail through three bearings. Table 1 lists the data of the bearing taken
into account and the fault frequency computed with the Eqs. 2–6 in the constant
speed case. The model simulates the expected vibration signal of the right upper
bearing for a single cart that is moving on the rail.
The rail consists of an upper and a lower linear paths, each of which are
1 m in length, and two circular paths that are 0.500 m long. The profile of the
radial load q(t), which acts on the bearing, has been computed by means of the
knowledge of the motion profile M (t). The forces taken into account are:
 Bearing Fault Model 217

Table 1. Vibration signal model data for localized outer ring fault in a rolling bearing

Model parameters Constant velocity

Dout (mm) 21.2
d (mm) 17.7
D (mm) 5
Number of spheres 7
Velocity (m/s) 1
Rotational frequency fr (Hz) 15.022
Sample frequency fs (Hz) 8E3
Outer-ring fault frequency (Hz) 37.889
SDOF spring stiffness k (N/m) 2E13
SDOF damping coefficient ζ 5%
SDOF natural frequency fn (Hz) 6E3

– gravity force
– friction along the track
– fictitious forces
– preload of the bearing
These forces are shown in Figs. 4 and 5. In the simulation the noise is equal
to zero only to make the output data more readable. All the aforementioned
possible damages have been taken into account and numerically validated, but,
for sake of brevity, only the results of the outer ring damage with the modulation
of the radial load are shown. The tests examine two situations:
1. Constant velocity motion profile
2. Variable velocity motion profile

4.1 Constant Velocity

In this case the cart has a constant velocity of 1 m/s. Results are shown in Fig. 4.
Figure 4(a) shows the linear displacement of the cart imposed in the simulation.
The Fig. 4(b) is the FFT of the vibration signal in the constant load zone,
it shows the resonance of the system clearly. Figure 4(c) shows the load profile
computed for the specific motion task. The resulting vibration signal (Fig. 4(d))
exhibits variation of the load profile due to the change of orientation of the
mover along the track and the two curves of the rail. The zoom of the spectrum
in Fig. 4(e) shows the damage frequency and its harmonics, spaced at the correct
fault frequency (fouter in this example).

4.2 Variable Velocity

In this case the cart follows the motion profile shown in Fig. 5(a), where the
cart decelerates and accelerates on the left curve in order to synchronize with a
218 M. Cocconcelli et al.

Fig. 4. Simulated vibration signal in the case of the inner ring fault at a constant veloc-
ity: (a) motion profile, (b) FFT of the vibration signal, (c) load profile, (d) simulated
vibration signal, (e) FFT zoomed.

Fig. 5. Simulated vibration signal in the case of the inner ring fault at a variable veloc-
ity: (a) motion profile, (b) FFT of the vibration signal, (c) load profile, (d) simulated
vibration signal, (e) FFT zoomed.

possible product placed along the lower linear path. The load profile is recom-
puted because of the velocity variation. In this case it is not possible to directly
compare the results of the formulas 1–6 with the results of the model because of
the variable fr at each time instant. In the zoomed spectrum in Fig. 5(e), there
are several peaks but they are not so clear as in the constant velocity profile.
This fact makes the diagnostics of the vibration signal more demanding.
The results show that the fault vibration signal is modulated not only by the
variable load as before but also by the variable motion profile.
 Bearing Fault Model 219

5 Conclusions
This paper shows a flexible model for the simulation of bearing fault vibration
signal in the case of independent cart conveyor system. Because experimental
data are not yet available, the model has been numerically tested referring to a
known case. In particular, constant speed application was simulated, looking for
expected bearing fault frequencies in the spectrum of the vibration data. The
geometry parameters used are those of the Beckhoff XTS system, considering
one cart along the rail and a fault on the upper right ring of the bearing. The
model allows to:

– Choose any track shape of the rail

– Use variable velocity and constant velocity profiles
– Use variable load profiles
– Simulate the most common damage of a rolling bearing along a linear rail
– Consider the resonance of the system
– Consider random contributions

The simulated data at a constant velocity show the fault frequency with a
modulation because of the variable load and the system resonance.

Acknowledgments. The authors are grateful for the National University Research
Fund (FAR 2016) of the University of Modena and Reggio Emilia - Departmental and
Interdisciplinary Projects (DR. 73/2017, Prot. n. 37510-27/02/2017) and the support
from Tetra Pak Packaging Solutions.

References
1. Rockwell Automation: iTRAK The Intelligent Track System Increase machine
flexibility and throughput to enhance overall productivity. http://literature.
rockwellautomation.com/idc/groups/literature/documents/br/motion-br007 -en-
p.pdf
2. Beckhoff Automation: XTS. The eXtended Transport System. https://download.
beckhoff.com/download/Document/Catalog/XTS Beckhoff e.pdf
3. B&R: ACOPOStrak Ultimate Production Effectiveness. https://www.br-
automation.com/smc/5adafdb3a7f954f17c8bb25652a8c971a38e4d94.pdf
4. Molano JCC, Rossi S, Cocconcelli M, Rubini R (2017) Dynamic model of an inde-
pendent carts system. In: Advances in Italian mechanism science. Mechanisms and
machine science, vol 47, pp 379–387
5. Bellini A, Filippetti F, Tassoni C, Capolino GA (2008) Advances in diagnostic
techniques for induction machines. IEEE Trans Ind Electron 55(12):4109–4126
6. Poyhonen S, Jover P, Hyotyniemi H (2004) Signal processing of vibrations for
condition monitoring of an induction motor. In: Proceedings of 1st international
symposium control, communications signal processing, pp 499–502
7. Nandi S, Toliyat H, Li X (2005) Condition monitoring and fault diagnosis of elec-
trical motors-a review. IEEE Trans Energy Convers 20(4):719–729
8. Randall RB (2011) Vibration-based condition monitoring: industrial, aerospace
and automotive application. Wiley, Hoboken
220 M. Cocconcelli et al.

9. Curcurú G, Cocconcelli M, Immovilli F, Rubini R (2001) On the detection of dis-

tributed roughness on ball bearings via stator current energy: experimental results.
Diagnostyka 51(3):17–21
10. Immovilli F, Cocconcelli M, Bellini A, Rubini R (2009) Detection of generalized-
roughness bearing fault by spectral-kurtosis energy of vibration or current signals.
IEEE Trans Ind Electron 56(11):4710–4717
11. Hayes M (1996) Statistical Digital Signal Processing and Modeling. Wiley, Hoboken
12. Trajin B, Chabert M, Regnier J, Faucher J (2009) Hilbert versus Concordia trans-
form for three-phase machine stator current time- frequency monitoring. Mech Syst
Signal Process 23(8):2648–2657
13. Salami M, Gani A, Pervez T (2001) Machine condition monitoring and fault diag-
nosis using spectral analysis techniques. In: Proceedings of the 1st international
conference on mechatronics, pp 690–700
14. Wang W, McFadden PD (1993) Early detection of gear failure by vibration anal-
ysis I. Calculation of the time-frequency distribution. Mech. Syst. Signal Process.
7(3):193–203
15. Wang W, McFadden PD (1996) Application of wavelets to gearbox vibration signals
for fault detection. J Sound Vib 192(5):927–939
16. Cerrada M, Snchez R-V, Li C, Pacheco F, Cabrera D, Valente de Oliveira J, Vsquez
RE (2018) A review on data-driven fault severity assessment in rolling bearings.
Mech Syst Signal Process 99:169–196
17. McFadden PD, Smith JD (1984) Vibration monitoring of rolling element bearings
by the high frequency resonance technique a review. Tribol Int 117:3–10
18. McFadden PD, Smith JD (1984) Model for the vibration produced by a single
point defect. J Sound Vib 96:69–82
19. McFadden PD, Smith JD (1984) The vibration produced by multiple point defects
in a rolling element bearing. J Sound Vib 98:263–273
20. Su YT, Lin SJ (1992) On initial detection of a tapered roller bearing frequency
domain analysis. J Sound Vib 155:75–84
21. Ho D, Randall RB (2000) Optimization of bearing diagnostic techniques using
simulated and actual bearing fault signals. Mech Syst Signal Process 14:763–788
22. D’Elia G, Cocconcelli M, Mucchi E (2018) An algorithm for the simulation of
faulted bearings in non-stationary conditions. Meccanica 53(45):1147–1166
 Fault Detection Methodology for a Fan
Matrix Based on SVM Classification
of Acoustic Images

Lara del Val1(&), Alberto Izquierdo2, Juan J. Villacorta2,

Marta Herráez1, and Luis Suárez3
1
Mechanical Engineering, School of Industrial Engineering,
University of Valladolid, Valladolid, Spain
{lvalpue,herraez}@eii.uva.es
2
Signal Theory and Communication Systems Department,
School of Telecommunication Engineering, University of Valladolid,
Valladolid, Spain
{alberto.izquierdo,juavil}@tel.uva.es
3
Civil Engineering Department, Superior Technical School,
University of Burgos, Burgos, Spain
luis.a.suarez.vivar@gmail.com

Abstract. A methodology to detect if a fan matrix is working properly has been

designed and is presented in this paper. This methodology is based on a Support
Vector Machine (SVM) classifier that uses geometrical parameters of the
acoustic images of the fan matrix. These acoustic images have been obtained
using a 16  16 planar array of MEMS microphones working at different fre-
quencies. A fan matrix that is not working properly implies that some of its fans
have failed, that is, it does not work. The designed fault detection methodology
supposes that these fans fail one by one. If one of the fans is not working, this
fact can be detected rapidly with the purposed methodology, and the fan can be
repaired or replaced by a new one. Although it is really unusual that more than
one fan fails at the same time, this paper also studies how this methodology
works if the number of faulty fans increases, in order to know if the method-
ology is robust enough in the presence of unexpected situations.

Keywords: Fault detection
Fan matrix
Acoustic images
SVM

1 Introduction

The detection of a fault consists of the determination of the existence either of a failure in
structural components or of an abnormal behavior of a system [1]. Condition monitoring
is the process of monitoring a parameter of condition in machinery (vibration, temper-
ature etc.), in order to identify a significant change which is indicative of a developing
fault. It is a major component of predictive maintenance. The use of condition monitoring
allows maintenance to be scheduled, or other actions to be taken to prevent consequential
damages and avoid its consequences, as a major failure. Condition monitoring techniques
are normally used on rotating equipment, auxiliary systems and other machinery
(compressors, pumps, electric motors, internal combustion engines, presses) [2].

© Springer Nature Switzerland AG 2019

A. Fernandez Del Rincon et al. (Eds.): CMMNO 2018, ACM 15, pp. 221–228, 2019.
https://doi.org/10.1007/978-3-030-11220-2_23
222 L. del Val et al.

Engineers have used vibration based analysis for decades to evaluate the condition of
complex mechanical systems such as rotating machinery [3, 4]. In these cases, the diffi-
culty arises when trying to relate faults with observable quantities, and many times these
relations are subject to human interpretation [5]. To avoid this subjective interpretation, an
automatic methodology should be desirable. The classic approach for monitoring is based
on making periodically vibration measurements of the equipment, and then comparing
them to known healthy/damaged data to assess the health status of the machine [6].
Sometimes, vibrational measurements need a sensor mounted on the machine, as
accelerometers, and this presence can imply disturbances on the machine response and
performance. As it known that vibrational responses are related to acoustic emissions,
one possible solution to this problem is the analysis of the related acoustic responses
instead of the vibrational ones. Acoustic-based diagnosis with non-contact measure-
ment is a good option, as sound field contains abundant information related to fault
pattern [7]. There are many examples of the use of microphone arrays in acoustic
imaging systems to measure this acoustic field [6–9].
An array is an arranged set of identical sensors, fed in a specific manner. The
beampattern of the array can be controlled by modifying the geometry of the array
(linear, planar…), the sensor spacing and the beampattern, the amplitude and phase
excitation of each sensor [10]. By using beamforming techniques [11], the array
beampattern can be electronically steered to different spatial positions, allowing spatial
filtering, i.e. the discrimination of acoustic sources on the basis of their position.
The authors of this paper have experience in the design and development of
acoustic arrays. This work is based on the use of a planar array of 8  8 MEMS
microphones [12] to acquire and process acoustic images of a fan matrix [13], in order
to detect if it is working properly. A fan matrix, fan array or fan wall is a system formed
by several fans located on a surface, working together in order to improve the per-
formance of one alone large fan with lower power consumption. Any type of appli-
cation that requires specific temperature conditions is a candidate for a fan matrix.
The failure detection methodology shown in this paper is based on a Support
Vector Machine (SVM) classifier that uses geometrical parameters of the acoustic
images of the fan matrix, as they offer useful information [14]. The acoustic images of
the fan matrix have been obtained using a 16  16 planar array of MEMS microphones
working at different frequencies. The designed fault detection methodology supposes
that the fans of the matrix fail one by one. If one of the fans is not working, this fact can
be detected rapidly with the purposed methodology, and the fan can be repaired or
replaced by a new one. It has been observed that the effect of the corresponding faulty
fan is shown on the geometry of the acoustic image of the fan matrix [15].

2 Hardware Setup
2.1 Processing and Acquisition System
This section shows the acquisition and processing system used in this work [12], based
on a 2D array of MEMS microphones. The acoustic images acquisition system used in
this paper is based on 4 Uniform Planar Arrays (UPA) of 8  8 2.125 cm-uniformly-
spaced MEMS microphones, forming a bigger UPA of 16  16 sensors. This 8  8
array module and the 16  16 array are shown in Fig. 1.
 Fault Detection Methodology for a Fan Matrix Based on SVM Classification 223

Fig. 1. (a) Array module with myRIO and MEMS array board. (b) 16  16 array

The system implements the acquisition of the acoustic signals, using the MEMS
array, and then the acquired signals are processed in order to generate the acoustic
images, using wideband beamforming; as it is shown in Fig. 2. The programming
language used is NI LabVIEW 2015, along with its Real Time, FPGA, and GPU
modules, which allows developing applications on different hardware platforms like
those used in the system: FPGA, Embedded Processor (EP), PC, and GPU.

Fig. 2. Software algorithms diagram

2.2 Test Fan Matrix

The system shown in previous sections has multiple applications: localization and
characterization of noise or vibration sources, spatial filtering and elimination of
acoustic interferences, etc. This study case is focused on obtaining acoustic images of a
3  3 fan matrix, specifically built to these tests with 9 coherent axial PC fans, which
move the air in the direction of the fan axis. Each of the fans used to build the fan
matrix is a Foxconn D90SM-12 3-Pin with 7 blades. One of these fans is shown in
Fig. 3(a) and (b) shows the fan matrix implemented for the tests. As it can be observed