Vous êtes sur la page 1sur 13

Neurocomputing 175 (2016) 838–850

Contents lists available at ScienceDirect

Neurocomputing
journal homepage: www.elsevier.com/locate/neucom

Classification of unbalance and misalignment in induction motors


using orbital analysis and associative memories
José Juan Carbajal-Hernández n, Luis P. Sánchez-Fernández, Ignacio Hernández-Bautista,
José de J. Medel-Juárez, Luis A. Sánchez-Pérez
Center of Computer Research, National Polytechnic Institute, Av. Juan de Dios Bátiz s/n, Nueva. Industrial Vallejo, Gustavo A. Madero, México D.F. C.P. 07738,
México

art ic l e i nf o a b s t r a c t

Article history: Fault detection in induction motors is an important task in industry when production greatly depends of
Received 30 October 2014 the functioning of the machine. This paper presents a new computational model for detecting mis-
Received in revised form alignment and unbalance problems in electrical induction motors. Through orbital analysis and signal
24 June 2015
vibrations, unbalance and misalignment motor faults can be mapped into patterns, which are processed
Accepted 29 June 2015
Available online 6 November 2015
by a classifier: the Steinbuch Lernmatrix. This associative memory has been widely used as classifier in
the pattern recognition field. A modification of the Lernmatrix is proposed in order to process real valued
Keywords: data and improve the efficiency and performance of the classifier. Experimental patterns obtained from
Induction motors induction motors in real situations and with a certain level of unbalance or misalignment were processed
Fault detection
by the proposed model. Classification results obtained in an experimental phase indicate a good per-
Associative memories
formance of the associative memory, providing an alternative way for recognizing induction motor faults.
Orbital analysis
Vibrations & 2015 Elsevier B.V. All rights reserved.

1. Introduction In literature, several models for analyzing electro-mechanical faults


have been helpful in the decision making of expert criteria. Those
Nowadays, induction motors are used widely in industry, where works have been implemented using different mathematical theories
exposure to hard environments and conditions affects the optimal such as spectral analyses, wavelet analysis, time warping, time fre-
functioning, making them more vulnerable to electrical or mechanical quency analysis, Wigner–Ville distribution, stator current analysis, etc.
faults. Those faults should be detected early in order to avoid degra- [10,28,29,35,38,58,59]. Those models allow determining useful char-
dation or failures on critical electro-mechanical engine parts. Most acteristics for preventing and detecting possible failures in induction
electric motor failures interrupt processes, reduce production and may motors. Other models based in artificial intelligence have been pro-
damage other related machinery. In some factories, a very expensive posed in order to complement this area, providing a strong base for
scheduled maintenance is performed in order to prevent sudden developing more complemented systems in this field; e.g. artificial
motor failures. Normal operation of motors is extremely significant neural networks [14,16,34,51,7], Bayesian networks [33], support
due to it ensuring precision equipment and manufacture process vector machine [4], fuzzy logic [15,61], entropy and optimization
safety [35,37]. An appropriated supervision and monitoring for motor methods [53] among others. Moreover, analyses of incipient motor
faults can be made using electrical signals that include features that
faults can be helpful for detecting bad performance in its operation. In
are useful for identifying certain motor faults. Those features can be
this sense, several analyses have been studied by using some char-
studied and explored using the pattern recognition theory, which
acteristic behaviors produced by the induction motor; e.g. acoustic
has proved extremely useful for motor fault identification. Lately,
analysis [17], electrical problems [57], orbital electro-magnetic analysis
several works using pattern processing have been proposed. In those
[19] or motor vibrations [24].
models, incipient motor fault signals must be correctly preprocessed
and transformed into patterns in order to correctly identify different
n
Corresponding author. motor fault states ([21,43,48,62]). On the other hand, using asso-
E-mail addresses: jcarbajalh@cic.ipn.mx (J. Juan Carbajal-Hernández), ciative memories is considered, because they have demonstrated to
lsanchez@cic.ipn.mx (L.P. Sánchez-Fernández),
be very efficient algorithms for pattern classification, providing
jjmedelj@cic.ipn.mx (I. Hernández-Bautista),
ignaciohb@gmail.com (J.d.J. Medel-Juárez), good performance and efficiency. In literature, several types of
l.alejandro.2011@gmail.com (L.A. Sánchez-Pérez). associative memories have been developed, trying to increase their

http://dx.doi.org/10.1016/j.neucom.2015.06.094
0925-2312/& 2015 Elsevier B.V. All rights reserved.
J. Juan Carbajal-Hernández et al. / Neurocomputing 175 (2016) 838–850 839

classification capabilities or their implementation z 1. Rotor unbalance can be defined as the unequal mass distribution
[1,2,22,27,30,31,40,45,46,49,52,56]. The first associative memories on the motor rotation center and this is the most common
was created by karl [49], where his Lernmatrix is a crucial precedent problem in induction motors (Fig. 1a). In this case, the main
in the development of current associative memory models and is harmonic presents amplitudes higher than normal [3].
one of the first successful attempts to encode information in an 2. Shaft misalignment is the most common cause of machine
arrangement, known as crossbar grids. In general, a particular vibration after imbalance, which leads additional dynamic load
drawback in associative memories is that data information is based and accelerate machine deterioration (Fig. 1b). This type of
on processing binary patterns; it means, real values (ℜ) must be vibration is often from reactive forces in the couplings between
binarized, which implies information loss or large vector sizes in two rotating shafts. Generally, a misalignment presents the
order to save as much information as possible. In order to resolve fundamental harmonic at 2x rpm.
this gap, actual improvements consider the use of real data infor- 3. Extremely shaft misalignment occurs when slight shaft mis-
mation trying to avoid this problem using new or classical models alignments are not corrected and the evolution increases. In an
as those proposed by Esmi et al. [13], Salavati et al. [47] and extreme misalignment the range can vary from 4x to 8x rpm
[18,26,32].
Zheng [60]. The Lernmatrix has been a widely studied architecture
and its recognition capabilities can be increased if a new algorithm
for storing and recovering real valued patterns is adapted as a new
2.2. Signal acquisition
extension. In this way, associative memories have considerable
importance in pattern recognition applications, and their char-
Vibration is a common symptom derived from mechanical faults in
acteristics provide a good solution for incipient motor faults in
induction motors. Such vibrations can be measured using a piezo-
pattern classification. Therefore, this work proposes a different way
electric accelerometer sensor, which generates an electrical signal that
for induction motor fault pattern recognition. The importance of
using orbital analyses is that no spectral analyses are needed in is proportional to the acceleration vibration of a seismic mass [20]. As
the developing of this research, avoiding the developing of too each motor has a different rotation speed, standards such as ISO 10816
complex models as those proposed by the literature. In this sense, [25] and VDI 2056 [54] have established sampling frequency rates for
we have hypothesized that unbalance and misalignment motor motor measuring. According to them, this work used a sampling
faults can be modeled using orbits, where amplitude vibrations frequency of 50 kHz, being large enough to obtain a good quality
measured, surely will define a characteristic orbit shape; this signal and avoid aliasing over tested induction motors. In this work,
behavior is used for recognizing the motor fault using the Lern- induction motors with ½ to 1 HP and from 1000 to 2000 rpm were
matrix. Hence, the core of this work is based on two contribu- used for measuring vibration signals.
tions: (a) the feature extraction of vibration signals for building Signal vibrations produced by an accelerometer depend of its
orbital patterns and (b) the modification of a binary classifier orientation. In order to measure orbital patterns, two piezo-electric
(Lernmatrix) for real-value data processing. accelerometers were placed in the motor chassis near each axis (x, y).
The rest of the paper is organized as follows: Section 2 presents A rotor orbital vibration can be measured if the accelerometers are
some concepts about incipient faults in induction motors and the placed orthogonally (Fig. 2).
main characteristics of each one. In Section 3, orbital vibration Once the accelerometers were placed correctly, the vibration
pattern analysis is developed, remaking the behaviors of the signals were obtained. Those signals are contaminated with other
orbital signal when a motor fault is present. Also, pattern con- undesirable signals (Fig.3) and must be removed in order to
struction methodology is described in this section using the digi- extract the orbit to be used; the following section explains this
talized electrical signals. In Section 4, the Lernmatrix associative
important process.
memory is used for pattern classification, explaining the archi-
tecture of the Lernmatrix core and the proposed modification. In
Section 5, experimental patterns of motor faults obtained from
real situations are used for testing the proposed model. Finally,
Section 6 discusses the results obtained and gives some conclu-
sions and future research directions.

2. Induction motor faults measurement

2.1. Definitions
Fig. 1. Different Motor fault illustrations: (a) rotor unbalance and (b) shaft
Motor vibrations can cause motor wearing, fissuring by fatigue, misalignment.
loss of effectiveness, breakage of seals, noise, etc. Vibrations are
good indicators about mechanical motor conditions and they can
be a suitable way for diagnosing fault evolutions. According to this,
catastrophic motor failures can be predicted if behavior changes of
motor vibrations are studied and interpreted.
Failures in motors may occur frequently in three main compo-
nents: rotor, stator and bearings [3,36,41,6,7,50,55]. Faults in electric
machines can be classified in mechanical or electrical faults. This work
is mainly interested in studying three types of motor fault due to
being the most commonly presented faults: rotor unbalance, shaft
misalignment and extremely shaft misalignment. These motor faults
are described as follows: Fig. 2. Accelerometers placement at 90° over the chassis and between sensors.
840 J. Juan Carbajal-Hernández et al. / Neurocomputing 175 (2016) 838–850

Unfiltered Orbits Filtered Orbits


1.00 1.00
0.80 0.80
0.60 0.60
0.40 0.40
0.20 0.20
x2 0.00 x2 0.00
-0.20 -0.20
-0.40 -0.40
-0.60 -0.60
-0.80 -0.80
-1.00 -1.00
-1.00
-0.80
-0.60
-0.40
-0.20
0.00
0.20
0.40
0.60
0.80
1.00

-1.00
-0.80
-0.60
-0.40
-0.20
0.00
0.20
0.40
0.60
0.80
1.00
x1 x1
Fig. 3. Examples of filtered and unfiltered orbit signals using a Butterworth filter for removing spurious frequencies.

3. Orbital pattern building complete extracted motor signal, where the first perfect orbit is
captured by the model.
In literature, induction motor faults have been studied using orbital Vibration signals in displacement units are compounded by
analysis by Dongfeng et al. [11] and Ha et al. [19], where some basis several harmonics; each can be related with the normal operation
about how motor faults can be represented by orbits using vibrations or with a motor fault. For unbalance and misalignment faults, the
or electro-magnetic force are given. Orbital analysis is used for main harmonic is extracted for creating the orbits; other harmo-
building specific patterns obtained from different induction motor nics can distort the orbit shape (Fig. 3), changing notably the main
faults and using mechanical vibrations. In this paper, orbital analysis is characteristic of a fault shape. Additionally, the orbit shape can be
used for detecting a specific set of motor faults due to them generates affected by the effects of many other types of faults; however, only
particular features that can be transformed into patterns that will be unbalances and misalignments are analyzed in this work and
processed in the recognition process. Nevertheless, it is necessary to other faults still are a case study, so they are treated as undesirable
apply some preprocessing steps to the original signal in order to harmonics. Therefore, those undesirable harmonics must be
extract one orbit from a measured dataset without any representative removed in order to have a good quality orbit. In this case, a digital
characteristic. filter will be helpful for removing those frequencies and it can be
defined as a system that performs mathematical operations on a
3.1. Preprocessing sampled, discrete-time signal to reduce or enhance certain aspects
of that signal. A digital filter is characterized by its transfer func-
Orbits are constructed using two vibration signals, which are tion, or equivalent. Mathematical analysis of the transfer function
plotted in a (x, y) chart. In order to create an orbit, two different (h) can describe how it will respond to a specific input. Designing a
displacement signals x1(t) and x2(t) are measured orthogonally in filter consists in developing appropriate specifications to the pro-
the same motor chassis (Fig. 2) and then are plotted orthogonally blem and then producing a transfer function which meets these
as f(x, y) ¼ f(x1, x2). However, a signal preprocessing treatment is specifications. The following expression defines the filter equation
needed for avoiding undesirable effects such as noise or spurious used for cleaning the displacement signal:
frequencies. First, measured vibration signals are preprocessed
separately in order to preserve their particular features. Shape X
N 1
sðnÞ ¼ hðkÞxðn  kÞ; ð3Þ
orbits are built using displacement signals, that were measured k¼0
indirectly by acquiring two accelerometer vibration signals placed
orthogonally; nevertheless, as an accelerometer measures vibra- where sðnÞ is the filtered signal and hðkÞ is the transfer function.
tions in acceleration units a(t), the measured signals must be There are several ways for expressing a transfer function; however,
transformed to positioning units. This process can be performed describing it in frequency domain, hðkÞ-H ðωÞ shows the response
when the acceleration of one signal is transformed into velocity of the filter isolating a band of desirable harmonics. In this work, a
and then into displacement by integrating the acceleration signal Butterworth passband filter was implemented for removing those
in time domain according to the following definitions [20]: spurious harmonics according to the following magnitude
Z t response [39,42]:
vðt Þ ¼ aðt Þdt þ v0 ; ð1Þ  
 
0  
  1
Z H ðωÞ2 ¼  2N ð4Þ
t    cos ω
xðt Þ ¼ vðt Þdt þ d0 : ð2Þ   1 þ Ω0 sin
c
ω
0  
where ω ¼ 2fπf , f s is the sampling frequency, Ω0 ¼ tan ω20 and c
where a(t), v(t) and x(t) are the acceleration, velocity and dis- s

placement signals; vo and do are the initial velocity and displace- can be expressed as follows:
 
ment values respectively. In this work, the signal acquisition was sin ωpa þ ωpb
c¼ ; ð5Þ
measured in motors at rest that were turned on till they reached sin ωpa þ sin ωpb
their normal operation speed, establishing as initial conditions v0 h i
2π f 2π f
and d0 in zero. These values were considered for processing the where ωpa ¼ f pa , ωpb ¼ f pb and f pa ; f pb is the passband.
s s
J. Juan Carbajal-Hernández et al. / Neurocomputing 175 (2016) 838–850 841

 
This work uses a bandpass frequency of 10 Hz f 0 7 5 Hz ; this we can extract one from the signal (Fig. 4). In this sense, this
value was enough to provide a good quality in the orbit shape, behavior can be taken as an advantage of this model, due to motor
removing undesirable harmonics. The filter passband was adjusted speed differences among other engines do not represent a sig-
to the fundamental frequency of the measured signal ðf 0 Þ. nificant modeling change because orbits can be obtained when the
Fig. 3 shows an example of a vibration orbit measured directly speed is constant.
from an induction motor and the effect of the bandpass filter used
in the same signals. In this case, unfiltered orbits cannot capture 3.2. Orbit classification
the behavior of a motor fault because their shapes are distorted by
undesired harmonic effects. Once those harmonics were removed, Orbits shapes have correspondence with induction motor faults
the motor fault can be clearly observed (Fig. 3) (see Section 3.2 for (see in Section 3.1) and they have some specific characteristics that
more details about motor fault shapes). can be useful for pattern classification. Fig. 5, shows the most
Each filtered signal generates orbits continuously with the representative shapes for the described faults, where a perfect
same symmetry; nevertheless, just one orbit is enough to identify circular shape represents a good condition motor, an elliptical
a motor fault. In this case, it is important to obtain a closed and shape corresponds to an unbalanced rotor fault and a distorted
complete orbit in order to avoid a distorted pattern where the ellipse corresponds to a misalignment rotor fault.
main fault characteristic can be lost (Fig. 4). The procedure for The orbit shape suggests the motor fault type, but the way in
extracting an orbit is looking for the starting and ending points which the severity of the shaft unbalance and the machine mis-
with a low distance between them, defined by a tolerance. There is alignment changes, has been planted as a future work. However,
not a specific rule for establishing this tolerance; nevertheless, this the deformity generated in the orbit is enough to identify an
value must be enough for obtaining good shape orbits. For a unbalance or misalignment. In this sense, we use the shape orbit
practical purpose, this tolerance was computed determining the as a particular pattern to be classified into the described faults.
distance average of all points into the signal. In this case, the Observing the orbit shape, the motor fault can be easily
distance criterion was obtained using the Euclidian equation as recognized by a user. However, the proposal of a computational
follows [42]: model is to automatize this process using a classifier. In this sense,
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 2 associative memories were applied, because they have been
d ¼ ðx2  x1 Þ2 þ y2  y1 ð6Þ proved to be good classifiers, providing an effective way for fault
where d is the distance between points, and (x, y) are the orbit recognition. In other words, the orbital analysis provides a new
points coordinates respectively. Finally, extracted orbits should be way to represent a certain motor fault; nevertheless, the Lern-
normalized due to differences in their size. Therefore, all orbits are matrix addendum provides a good classifying model in order to
resized in a [  1, 1] range according with the following equation: reduce human dependence, having a reliable computational
system.
s1;2 ðnÞ
s01;2 ðnÞ ¼ ; 8 n ¼ 0; 1; 2; …; N  1 ð7Þ
maxfj s1 ðnÞj ; j s2 ðnÞj g
Since different kind of motors were used for measuring vibra- 4. Associative memories
tions, making relations between a specific motor with a particular
orbit size is a hard task. According to this, a normalization process In this section, an explanation of the classifier used for recog-
allows applying the proposed model to any kind of induction nizing orbital patterns is given. This work illustrates the associa-
motor, since an orbit shape only describes the type of fault that tive memories because they have proved good efficiency and the
will be recognized by a classifier. learning and recovering processes are implemented mostly by
If the velocity vibration is measured when the motor turns on, algebra equations, avoiding the complexity of minimization gra-
then the acceleration increases the velocity and this affects the dient processes. It is important to remark, that the learning pro-
orbit pattern generation. This behavior can be observed in the cess in an associative memory is made using pattern associations
orbit shape building, where a small orbit is drawn and its size with binary algebra operations; this means, there is not a learning
continuously grows due to the acceleration progress. This aspect convergence algorithm such as the artificial neural networks (this
represents a problem, because the orbit is open and cannot be is the main difference between these classifiers). This behavior
perfectly captured. When the motor reaches its normal velocity guarantees a certain learning level, which should be enough for a
operation, the generated orbits are closed and complete; therefore good recovering process.

Orbits Open orbit Closed orbit


0.15 0.30 1.00

x2 0.00 x2 0.00 x2
0.00

-0.15 -0.30 -1.00


-0.15

0.00

0.15

-0.32

0.00

0.32

-1.00

0.00

1.00

x1 x1 x1

Fig. 4. When the motor accelerates, the orbit size increases continuously generating a spiral. In order to obtain a closed orbit, velocity must be constant in order to preserve
the main motor fault characteristics.
842 J. Juan Carbajal-Hernández et al. / Neurocomputing 175 (2016) 838–850

Unbalanced Misalignment Extremely Misalignment


1.0 1.0 1.0
0.8 0.8 0.8
0.6 0.6 0.6
0.4 0.4 0.4
0.2 0.2 0.2
x2 0.0 x2 0.0 x2 0.0
-0.2 -0.2 -0.2
-0.4 -0.4 -0.4
-0.6 -0.6 -0.6
-0.8 -0.8 -0.8
-1.0 -1.0 -1.0
-1.0
-0.8
-0.6
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
1.0

-1.0
-0.8
-0.6
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
1.0

-1.0
-0.8
-0.6
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
1.0
x1 x1 x1

Good Condition
1.0
0.8
0.6
0.4
0.2
x2 0.0
-0.2
-0.4
-0.6
-0.8
-1.0
-1.0
-0.8
-0.6
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
1.0
x1

Fig. 5. Examples of orbit shapes that correspond to different motor faults.

As a classifier, the associative memories have a fundamental pur- There are several types of associative memories that have been
pose: recovering correctly patterns which in most cases have been widely used in the pattern recognition field, they can range from
altered with some kind of noise. The capacity of good learning, high simpler designs to some very complex to implement
storing rates and noise immunity are interesting properties in the [1,2,5,8,9,12,22,30,31,40,44,49,52]. The main motivation of using
pattern processing field. Therefore, an associative memory can be the Lernmatrix is that it is a well-known architecture that can be
defined as an input and output system that generates a relationship as customized in order to recognize non-binary vectors, providing a
follows: x-½M -y, which has as result a matrix M ([1]; Carbajal and good alternative in the pattern recognition field.
Sánchez, 2001). The input and output patterns are represented by
vectors and they are denoted by an x and y respectively. The goal of an
associative memory is to restore full patterns from input patterns that 4.1. The Lernmatrix
can be altered. Each association are represented by a (xk, yk) notation.
Thus, the associative memory can be expressed as a matrix M, which The Lernmatrix was proposed by Karl [49,50] and it is based on
has ijth elements denoted as mij [2,30,40,45,46]. The total number of coding data information using arrays, known as crossbars. According
associations conform the fundamental set (S), which can be expressed to the classical definition of the Lernmatrix, it is a binary classifier that
according to the next expression: works associating input vectors with output class vectors (hetero-
n  associative memory). This classifier, accepts binary input patterns such

s¼ xk ; yk k ¼ 1; 2; …p ð8Þ as xk A An and returns output binary patterns such as yk A Am, where
A¼ {0, 1}. Output patterns are orthonormal vectors that correspond to
where p denotes the fundamental set cardinality. Additionally, each a class that belongs to an input vector. According to this, there are m
input and output vector can be described according to the following different classes, defining yik with values according to ylk ¼1 for j¼m
equation: and yjk ¼0 for jam [1]. In the following sections, the training and
2 k3 2 k3 recalling phases are explained in order to introduce the reader to the
x1 y1
6 k7 6 k7 operation of the associative memories and the problems that are
6 x 7 6 y 7 actually detected in their functioning.
xk ¼ 6 27 n k 6 27 m
6 ⋮ 7 A A y ¼ 6 ⋮ 7 A A 8 k ¼ 1; 2; …; p ð9Þ
4 5 4 5
xkn ykm a) Learning phase
The Lernmatrix memory is built according to the p associations
where n and m are the number of components in A¼{0,1} for a binary between patterns (xk, yk) that have been included in the fun-
associative memory. damental set, where k¼ 1, 2, …, p. Each association creates a
Associative memories are classified in two types: (a) auto- matrix that is updated in each association. According to this,
associative, when the input vector is used as output vector too and the matrix generated by an association can be defined as
(b) hetero-associative, when input vectors are different to output follows:
vectors. This work used a hetero-associative memory since there
xk2 ⋯ xkj ⋯ xkn
are orbital patterns as inputs and classification vectors as outputs. xk1
J. Juan Carbajal-Hernández et al. / Neurocomputing 175 (2016) 838–850 843

2 3
yk1 m11 m12 ⋯ ⋯ Carbajal and Sánchez [5], where some additional algorithms have
6 m1j m1n 7
k 6 m21 m22 ⋯ ⋯ 7 been proposed in order to avoid this problem. Those works has been
y2 6 m2j m 7
6 2n 7 created for the original Lernmatrix, providing a basis for improving
⋮ 6⋮ ⋮ ⋱
mij

min 7
6 7 ¼ M; 8 k ¼ 1; 2; …; p ð10Þ our proposed modification, where real valued vectors are processed.
yki 6 m m ⋯
6 i1 i2
⋯ 7
7
6 ⋮ 7
⋮ 6 ⋮ ⋮ ⋱ ⋱⋮ 7 4.2. Proposed modification to the Lernmatrix model
4 5
ymk m m1 m m2 ⋯ m mj ⋯ mmn
Our proposed real valued Lernmatrix model is a modification to
the learning and recalling phase of the Steinbuch Lernmatrix, fol-
The M matrix is initialized with zeros and each mij component lowing the same operation rules. The new associative memory has
must be updated with each association as follows: two approaches; in the first, binary data information is processed
mij ¼ mij þ Δmij ð11Þ having better results than the original Lernmatrix (See Section 5); in
the second, real data information is used (real-valued Lernmatrix). The
real-valued Lernmatrix is helpful in avoiding the binarization process,
Essentially, calculating Δmij is the core of the Lernmatrix algo- reducing some preprocessing steps.
rithm, and the main efforts are focused on computing this Binary and real patterns can be stored in the proposed Lernmatrix
updating parameter as follows: memory introducing a modification in the learning phase, which
8
>
> þ ε if xkj ¼ 1 and yki ¼ 1 constitutes a novelty in this process. In addition, a second modification
<
is introduced in the recalling phase in order to recover the outputs
Δmij ¼  ε if xkj ¼ 0 and yki ¼ 1 8 k ¼ 1; 2; …; p ð12Þ
>
> patterns that correspond to input patterns. In the recalling phase, a
:
0 otherwise new algorithm was introduced and new concepts that jointly produce
where ε is a positive constant previously chosen. Commonly, a a better recall for fundamental input patterns. The proposed mod-
very small value for ε is enough for a good working of the ification introduces new steps in the learning and recalling phases of
algorithm. the Lernmatrix, which leads to improve patterns recall.
b) Recalling phase
The recalling phase consists of recovering an output class a) New learning phase
ω ω The Steinbuch Lernmatrix is useful when binary patterns are
vector (y ) that corresponds to an arbitrary input vector (x )
using the Lernmatrix memory as follows: classified. In order to avoid binarization processes and to
reduce noise effects in the recalling phase; it was proposed
X
n

i ¼ mij U xω
j ; i ¼ 1; 2; …; m: ð13Þ building the learning phase using linear algebra, in such a way
j¼1 that using the dot product between a vector and an unitary
vector (in the form of a pattern) can identify the class of the
pattern stored within a matrix containing n different vectors.
Finally, the output class vector is determined according to the
Therefore, the new learning phase consists in finding a new
following equation:
way to generate the matrix M, which stores the information of

1 if cωi ¼ maxfc g
ω the p associations of the fundamental set {(x1, y1),.., (xp, yp)}. In

i ¼ ; 8 i ¼ 1; 2; …; m and ω ¼ 1; 2; …; p:
0 otherwise this case, input vectors are real-valued patterns that can be
ð14Þ expressed as xk A An, A A ℜ and output vectors are binary
patterns that can be expressed as yk A Bm, B ¼{0, 1}. Ortho-
normal output vectors are also used for the Lernmatrix and
In the Steinbuch Lernmatrix, orthogonal output vectors are used their components yik are defined with values according to
for assigning an independent class to each input vector. This ylk ¼1 for j¼m and yjk ¼0 for jam. After Eq. (13) has been
guarantees a perfect recovery if a trained input pattern is pro- performed, the new process for determining the real valued
cessed by the associative memory as long as crossbar saturation is Lernmatrix can be enunciated using the following expression:
not present. In this sense, when input – output vectors are 8 k
associated in the training phase, their information is stored in the >
> x
< ‖xkj‖2 if yki ¼ 1
memory matrix. On the other hand, when an input trained pat- Δmij ¼ 8 i ¼ 1; 2; …; m: ð15Þ
>
>
tern is processed in the recovering phase, its corresponding :
0 otherwise
orthonormal vector (class) is obtained. According to this, untr-
ained input patterns with similitude to those used in the funda-
mental set (training process) are classified with a respective class b) New recalling phase
(orthonormal vector). When a cross bar saturation is present, an In this process, the binary Lernmatrix recalling phase was
orthonormal vector is duplicated or a non-orthonormal vector is modified in order to recover the real valued patterns stored in
used in the training process, then a misclassification is present in M. As in Eq. (13), the recalling phase consists of recovering an
ω
the recovering process [1,5]. output class vector (y ) that corresponds to an arbitrary input
ω
vector (x ) using the Lernmatrix memory as follows:
c) Crossbar saturation X
n

i ¼ mij ∙xω
j 8 i ¼ 1; 2; …; m: ð16Þ
Although the Lernmatrix is a good classifier, it has a drawback: j¼1

pattern misclassification occurs when crossbar saturation exists. A


crossbar saturation phenomenon occurs every time input pattern Then, the output vector is adjusted using the following expression:
characteristic set is a subset of another one. Additional pattern mis-
ð Þ
2
 Cω  1
classification occurs when yk has more than one bit as a class iden-

i
i ¼e ; 8 i ¼ 1; 2; …; m: ð17Þ
2
tifier (erroneously coded class vector) [5]. However, crossbar satura-
tion also occurs when the coded class vector is correct. More details
about crossbar saturation can be consulted in Aldape et al. [1] and Finally, the output class vector is determined according to the next
844 J. Juan Carbajal-Hernández et al. / Neurocomputing 175 (2016) 838–850

equation: (x2, y2), (x3, y3)} as follows:


8
ω
>
> 1 if Rω
i ¼ max R
0
5:3
1 0
3:1
1
>
< 0 1
B 7:2 C B 3:2 C

i ¼ ; 8 i ¼ 1; 2; …; m: ð18Þ B
B
C
C
1 B
B
C
C
>
> B C
>
: x ¼ B 1 C-y ¼ @ 0 A-x ¼ B
1 B C 1 2
B 0 C-
C
0 otherwise B C B C
@ 4 A 0 @ 5 A
1:2 2
0 1
The experimental section shows that these concepts allow obtain- 4:2
ing an improvement in the recall performance of the proposed 0 1 B  6:4 C 0 1
0 B C 0
B C B C
Lernmatrix.
y ¼ @ 1 A-x ¼ B
2 3
 1:3 C-y ¼ B
3
@ 0A
C
B C
c) Crossbar saturation B C
0 @ 2:1 A 1
Even though the proposed Lernmatrix has a good pattern recall 0
performance, it has the same misclassification problem as the
original Lernmatrix due to crossbar saturation. Crossbar saturation
occurs whenever the characteristic set of one or more learned 5. Learning phase
patterns is a subset of the characteristic set of another funda-
mental pattern [1,5]. Crossbar saturation can be avoided introdu- To create the matrix M using the p associations of the funda-
cing an additional vector which is built using the same informa- mental set, it is necessary to follow steps 1 and 2 of Section 4.2 as:
tion of the associative memory; each pattern introduced in the M Step 1 Step 2
memory gives information that could be useful when it is recalled. 0 1 0 1
00000 0:0538 0:0732 0:0101 0:0406 0:0122
B C
Therefore, this algorithm introduces some concepts at the end of M ¼ @ 00000 A B 0:0634  0:0655 0:0000 0:1023 0:0409 C
M ¼@ A
the recalling phase of the proposed Lernmatrix that allows incr- 00000
0:0649  0:0989  0:0201 0:0324 0:0000
easing the recall of patterns when the crossbar saturation effects
appear in the output patterns (see Section 4.3); thus, an additional
algorithm is executed following some steps, explained separately Once the learning phase and the associative memory M are
in order to facilitate the understanding of the proposal. First, a computed, the next step is the recalling phase.
weighted vector is built counting the number of zeros contained in
the Lernmatrix as follows:
8 6. Recalling phase
< ui þ 1 if mij ¼ 0
>
ui ¼ ; 8 j ¼ 1; 2; …; n;i ¼ 1; 2; …; m ð19Þ
>
:u The recalling phase for the proposed method is computed using
i otherwise
the matrix M and the previously defined rules. According to
ω
where ui is initialized with 1. Then, a static threshold is computed Eq. (16), the output vectors (Ci ) are computed as follows:
0 1
determining the maximum value of the weighted vector using the 0:0538 0:0732 0:0101  0:0406 0:0122
next expression: B C
C 1 ¼ Mx1 ¼ @ 0:0634  0:0655 0:0000  0:1023 0:0409 A

 0:0649  0:0989 0:0201 0:0324 0:0000
θ ¼ max ui :i ¼ 1; 2; …; n ð20Þ
0 1
5:3
The computed threshold is used to build a normalized vector B 7:2 C 0 1
B C 1
using the following equation: B C B C
B C
B 1 C ¼ @ 0:323 A;
θ B C
@ 4 A  0:518
si ¼ ; 8 i ¼ 1; 2; …; m ð21Þ
ui 1:2
Once the normalized vector is computed, it is multiplied by the 0 1
output vector as presented: 0:0538 0:0732 0:0101  0:0406 0:0122
2 B C
C ¼ Mx ¼ @ 0:0634  0:0655 0:0000  0:1023 0:0409 A
2
ω
di ¼ si ∙yω
i ; 8 i ¼ 1; 2; …; m ð22Þ 0:0649  0:0989 0:0201 0:0324 0:0000
0 1
Finally, the reclassified output vector is obtained according to 3:1
B  3:2 C 0 0:160 1
the next rule: B C
B C B C
8 ω
ω B C
B 0 C ¼ @ 1 A;
>
> 1 if di ¼ max d B C
>
< @ 5 A 0:355

i ¼ ; 8 i ¼ 1; 2; …; mandω ¼ 1; 2; …; p: 2
>
>
>
:
0 otherwise 0 1
0:0538 0:0732 0:0101  0:0406 0:0122
ð23Þ B C
C 3 ¼ Mx3 ¼ @ 0:0634  0:0655 0:0000  0:1023 0:0409 A
0:0649  0:0989 0:0201 0:0324 0:0000
0 1
4.3. Numerical example 4:2
B  6:4 C 0 0:340 1
B C
B C B
For a better understanding of the proposed Lernmatrix, this B  1:3 C ¼ @ 0:470 C
A:
B C
section exemplifies the operation of the new Lernmatrix. Suppose B C
@  2:1 A 1
a fundamental set with p¼ 3, n ¼ 5 and m ¼3. The fundamental
0
patterns and associations are expressed as ordered pairs {(x1, y1),
J. Juan Carbajal-Hernández et al. / Neurocomputing 175 (2016) 838–850 845

0 1
Recovering the class for each one of the 3 input vectors 4:2
B  6:4 C 0 0:051 1
according to the expression (17) and (18), obtains: B C
B C B C
0 1
0 1 B C
B  1:3 C ¼ @ 0:470 A;
1 1 B C
ð C 1  1Þ
2
@  2:1 A 1
 B C B C
R1 ¼ e 2 ¼ @ 0:795 A-@ 0 A-y1 ; 0
0:316 0

0 1 0 1 0 1
0:703 0 0:1185 0:1056 0:1112 0:0403 0:0338
ð Þ B C
2
C2  1
B C B C C 4 ¼ Mx4 ¼ @ 0:0634  0:0655 0:0000  0:1023 0:0409 A
R2 ¼ e  2 ¼@ 1 A @ 1 A-y2 ;
-
0:812 0 0:0649  0:0989 0:0201 0:0324 0:0000

Step 1 Step 2: updated


0 1 0 1
0:0538 0:0732 0:0101  0:0406 0:0122 0:1185 0:1056  0:1112 0:0403 0:0338
B C B C
M ¼ @ 0:0634 0:0655 0:0000  0:1023 0:0409 A M ¼ @ 0:0634  0:0655 0:0000  0:1023 0:0409 A
0:0649 0:0989  0:0201 0:0324 0:0000 0:0649  0:0989  0:0201 0:0324 0:0000

1
0
1 0 10 2:4
0:407 0 B 1:2 C 0 1:059 1
ð C 3  1Þ B
2
C B C B C
3 
¼ @ 0:869 A-@ 0 A-y3 : B C B
R ¼e 2
B  4:5 C ¼ @  0:201 C
A:
B C
1 1 B C
@ 3 A 0:225
The recalled output vectors correspond to their input vectors 0:8
respectively. Moreover, what is the Lernmatrix operation if there Recovering the class for each one of the 4 input vectors and
are more input vectors than classes? In this case, only three classes according to the expression (17) and (18), obtains:
were defined and four patterns should belong to one of them. In 0 1 0 1
order to illustrate this detail, in the next example we introduce 0:987 1
ðC 1  1Þ
2
B C B C
one additional input vector x4 that belongs to y1 class as: R ¼ e 2
1
¼ @ 0:795 A-@ 0 A-y1 ;
0 1 0:316 0
2:4
B 1:2 C 0 1 0 1
0 1
B C 1 0:544 0
B C ðC 2  1Þ
2

x4 ¼ B  4:5 C-y1 ¼ B @
C
0 A: 2
R ¼e  B
¼@
C B C
1 A-@ 1 A-y2 ;
B C 2
B C
@ 3 A 0 0:812 0
0:8 0 1 0 1
0:637 0
ðC 3  1Þ
2
Updating the previous real-valued Lernmatrix: 3  B C B C
ω R ¼e 2 ¼ @ 0:869 A-@ 0 A-y3 ;
Computing the output vectors (C ):
1 1
0 1
0:1185 0:1056  0:1112 0:0403 0:0338 0 1 0 1
B C 0:998 1
C 1 ¼ Mx1 ¼ @ 0:0634  0:0655 0:0000  0:1023 0:0409 A ðC 4  1Þ
2
B C B C

0:0649  0:0989  0:0201 0:0324 0:0000 R4 ¼ e 2 ¼ @ 0:486 A-@ 0 A-y1 :
0 1 0:740 0
5:3
B 7:2 C 0 1:157 1 Finally, the recalled output vectors correspond to their input
B C
B C B C vectors respectively. It can be observed that the recalling phase
B C
B 1 C ¼ @ 0:323 A;
B C was made perfectly to the four input patterns. Nevertheless, when
@ 4 A  0:518
the number of patterns increases a saturation phenomenon occurs,
1:2 which produces a misclassification problem. In order to avoid
saturation problems, in this work, all patterns used in the learning
0 1
0:1185 0:1056  0:1112 0:0403 0:0338 process were assigned to an orthonormal output vector, having a
B C
C 2 ¼ Mx2 ¼ @ 0:0634  0:0655 0:0000  0:1023 0:0409 A particular class for each input pattern. So as to identify a specific
0:0649  0:0989  0:0201 0:0324 0:0000 motor fault, a classified pattern should belong to a set of classes
0 1 associated to those patterns that represents a certain fault.
3:1
B 3:2 C 0  0:103 1
B C 6.1. Misclassification example
B C B C
B
B 0 C¼@
C 1 A;
B C
@ 5 A 0:355 For a better understanding of the crossbar saturation phe-
2 nomenon, this section exemplifies how the proposed modification
in the real valued Lernmatrix solves this problem. Suppose a
0 1 fundamental set with p ¼3, n ¼5 and m ¼3. Given the fundamental
0:1185 0:1056  0:1112 0:0403 0:0338
3 B C patterns {(xk, yk) | k¼ 1, 2, …, p} and the fundamental associations
C ¼ Mx ¼ @ 0:0634  0:0655 0:0000  0:1023 0:0409 A
3
expressed as ordered pairs {(x1, y1), (x2, y2), (x3, y3)}:
0:0649  0:0989  0:0201 0:0324 0:0000
846 J. Juan Carbajal-Hernández et al. / Neurocomputing 175 (2016) 838–850

0 1 0 1 problem, these outputs are processed using the proposed algorithm


5:3 5:3
B 7:2 C 0 1 B C 0 1 according to the expressions previously defined in Section 4.2.
B C 1 B 0 C 0 Using Eqs. (19) and (20), the weighted vector ui and the threshold of
B C B C
x1 ¼ B 1 C-y1 ¼ @ 0 A-x2 ¼ B 1 C-y2 ¼ B
B C
@
C
1 A-x3 ¼ the matrix M can be computed as follows:
B C B C
B C B C
@ 4 A 0 @ 0 A 0 0 1
1:2 1:2 1
0 1 B C
5:3 u ¼ @ 3 A;θ ¼ 4:
B 0 C 0 1
B C 0 4
B C
B 0 C-y3 ¼ B @0A
C Using Eq. (21), the normalized vector is obtained as follows:
B C
B C 0 1
@ 0 A 1
4
1:2 4 B C
si ¼ ¼ @ 1:333 A; 8 i ¼ 1; 2; …; n:
ui
1
7. Learning phase Finally, the corresponding classes can be computed using the
expressions (22) and (23), as follows:
For creating the matrix M using the p associations of the funda- 0 10 1 0 1 0 1
4 1 4 1
mental set, it is necessary to follow the steps 1 and 2 of Section 4.2: 1 B CB C B C B C
d ¼ sy ¼ @ 1:333 A@ 1 A ¼ @ 1:333 A-@ 0 A-y1; ;
1

Step 1 Step 2 1 1 1 0
0 1 0 1
00000 0:0538 0:0731 0:0101  0:0406 0:0122
B C B C 0
10 1 0 1 0 1
M ¼ @ 00000 A M ¼ @ 0:1736 0:0000 0:0327 0:0000 0:0393 A 4 0 0 0
2 B CB C B C B C
00000 0:1794 0:0000 0:0000 0:0000 0:0406 d ¼ sy ¼ @ 1:333 A@ 1 A ¼ @ 1:333 A-@ 1 A-y2; ;
2

1 1 1 0

Once the learning phase is computed and the associative


memory M is built, the next step is the recalling phase.
9. Experimental results

8. Recalling phase An experimental phase is made in order to validate the per-


formance of the proposed model, using the modified Lernmatrix
The recalling phase is computed using the same process that and some motors with a history of failure, generating different
Section 4.3 as follows: databases to be used in this section. A comparison using different
0 1 models designed for similar problems and that proposed in this
5:3
0 1B 0 1
0:0538 0:0731 0:0101  0:0406 0:0122 B 7:2 C
C 1 work has been computed in order to observe the results obtained
B 0:1736 0:0000 0:0327 0:0000 0:0393 CB C B C
by the real-valued Lernmatrix.
AB C
1 1
C ¼ Mx ¼ @ B 1 C ¼ @ 1 A;
B C As can be seen, an orbit shape corresponds to a certain motor
0:1794 0:0000 0:0000 0:0000 0:0406 @  4 A 1
1:2 condition, but each orbit can be extracted with a different length.
0 1
According to this, resulting orbits are not practical to be used in an
5:3 associative memory, due to homogeneous sized vectors must be
0 1B 0 1
0:0538 0:0731 0:0101  0:0406 0:0122 B 0 C C 0:310
B CB 1 C used. In order to have uniform patterns, all orbit signals were
C 2 ¼ Mx2 ¼ @ 0:1736 0:0000 0:0327 0:0000 0:0393 AB C¼B
@ 1
C
A;
B C resampled to 100 points of length by each axis (x, y), having a 200
B C
0:1794 0:0000 0:0000 0:0000 0:0406 @ 0 A 1
points pattern length (x þy). A database of 700 patterns was cre-
1:2
ated to be used in the learning phase of the associative memory.
0 1 Orbit patterns of this database were measured from different
5:3
0 1B 0 1
0:0538 0:0731 0:0101  0:0406 0:0122 B 0 C C 0:310 induction motors, which had different misalignment levels: 275
B CB C B
C ¼ @ 0:967 C
C 3 ¼ Mx3 ¼ @ 0:1736 0:0000 0:0327 0:0000 0:0393 ABB 0 C A: regular misalignment patterns, 106 extreme misalignments pat-
B C
0:1794 0:0000 0:0000 0:0000 0:0406 @ 0 A 1 terns, 314 unbalanced patterns and 5 patterns of good condition
1:2 motors. For testing the binary Lernmatrix (LM) and the binary
Recovering the respective output vector for each one of the Thresholded Lernmatrix (TLM), a binarization process for the
3 input vectors and according to the expressions (17) and (18), computed patterns was made using the IEEE-754 [23] standard,
obtains: where real values were transformed to 32 bit vectors (having 100
0 1 0 1 x-points þ100 y-points of 32 bits¼ 6400 bit vectors length). It is
1 1 important to remark that our proposed Lernmatrix processes real
ð C 1  1Þ
2
1  B C B C
R ¼e 2 ¼ @ 1 A -@ 1 A-y1 ; y2 or y3 ; valued data (RLM) and a non binarization process is needed.
1 1 As was mentioned in Section 4.1, The Lernmatrix relates a
binary vector (xk Þ with an orthonormal vector (yn Þ. The proposed
0 1 0 1
0:788 0 modification is based on the same definition (improving it to real-
ð C 2  1Þ
2
2  B C B C valued input vectors). In this case, one orthonormal vector is
R ¼e 2 ¼@ 1 A-@ 1 A-y2 or y3 ;
1 1 assigned to each pattern as a class, discerning the conventional
concept of an ANN or other associative memory designs where
0 1 0 1
0:782 0 only one class is assigned to a specific set of input patterns
ð C 3  1Þ
2
3  B C B C [2,52,9]. This detail can be considered as a drawback in the Lern-
R ¼e 2 ¼ @ 0:999 A-@ 0 A-y3
matrix due to big memory sizes being assigned. However, the
1 1
effect of a matrix saturation is minimized whenever inside ele-
In this case, the crossbar saturation misclassification problem ments are not altered when a new pattern updates the Lernmatrix
affects the output vectors for x1 and x2. In order to correct this in the learning process, having the whole pattern information
J. Juan Carbajal-Hernández et al. / Neurocomputing 175 (2016) 838–850 847

captured by the classifier (as long as crossbar saturation is avoi-

Efficiency
ded). The training database was compounded by 700 patterns

80.0%

99.2%
100%

100%
with their respective orthonormal output vectors. Nevertheless,
when a new never seen input pattern is processed by the Lern-
matrix, its class should belongs to the set of output vectors that

Recalled patterns
represents a specific fault (275 for misalignments, 106 for extreme,
314 for unbalance and 5 for good condition motor fault); e.g. if one
misalignment pattern is classified by the Lernmatrix, its output

ANN
class should corresponds to one of those 275 output vectors

273
105
314
4
defined for the misalignment fault.
A comparison using an artificial neural network (ANN) was

Efficiency
proposed as an alternative methodology. ANN's have proved to be

80.0%
98.4%
92.0%
85.8%
a good learning model with high rates of effectiveness, and they
could be a good tool for comparing the effectiveness of the real-
valued Lernmatrix. A Multilayer Perceptron (MLP) was used in this

Recalled patterns
work; the architecture that best fitted to the problem context was
of [200–150 – 2]. In this case, the inputs were fixed at 200 neu-
rons, where the first half is for x-axis and the other half for y-axis.

MAM
1000 epochs where used for training the ANN having a MSE of

309
253
91
1⨉10  4. The hidden layer was initialized between 2 and 200

4
neurons sequentially, where 150 neurons was the configuration

Efficiency
with the best results. Three sets were used to evaluate the per-

99.6%
98.1%
100%
100%
formance of the ANN (dividing the training database for different
purposes on the training process); one inside set for the training
algorithm (626 partners from the available data to train), one

Recalled patterns
validation set (35 partners from the available data) and one last
testing set to evaluate the final performance of the algorithm
(35 partners from the available data), with correlation coefficients
(r) of 0.965, 0.916 and 0.972 respectively. The training process used

TLM

104
274
314
by the ANN learning process was the Levenberg–Marquardt

5
algorithm. Finally, a Morphological Associative Memory (MAM)

Efficiency
proposed by Ritter et al. [44] has been planned to be used as a

99.6%
98.1%
different associative model for comparing the Lernmatrix results.

100%
100%
This associative memory has been designed as a particular artifi-
cial neural network (ANN), where the multiplication and addition
Recalled patterns
operations were changed to addition and maximum (or minimum)
respectively (called max or min memory type). This associative
memory accepts integer patterns as input and responds with an
orthonormal output vector, generating a different matrix for each
104
274
314
LM

association and by applying the maximum or minimum operator


5

all matrices are combined to build one final memory M. In this


Efficiency

case a max MAM was used in the experimental phase.


Using the procedure of Sections 4.1 and 4.2, the Lernmatrix
98.7%

99.0%
100%

100%

models were built using the proposed database. Once the Lern-
matrix has been computed, a preliminary test must be made using
the learning database for a recovering process in order to deter-
Recalled patterns

mine the classifier performance. In this case, high recovery rates


were reached in all the classifiers, where a 98.11–100% range of
recovering patterns can be observed in Table 1. On the other hand,
Motor fault recognition results using the learning database.

RLM

an experimental database was used for validating the performance


275
105
310
5

of the proposed system as part of a recovering process. This


database was built using different types of motor and with dif-
Number of patterns

ferent levels of misalignments. It is important to remark, that this


database was compounded by different motor measurements than
those used in the learning process. In this case, the database size
was of 1407 motor fault patterns as follows: 348 regular mis-
alignment patterns, 141 extreme misalignments patterns, 635
275
106
314
5

unbalanced patterns and 10 patterns of good condition motors. As


in Table 1, an evaluation using the original Lernmatrix is per-
Extreme misalignment

formed in order to compare the results obtained between the


proposal, the binary associative memories, the ANN and the MAM
Good condition

models.
Misalignment
Motor faults

Finally, the performance results of the motor fault recognition


Unbalance

using the experimental database, the associative memories, the


Table 1

ANN and the MAM are shown in Table 2. In this case, good con-
dition patterns were classified correctly by the proposed classifier,
848 J. Juan Carbajal-Hernández et al. / Neurocomputing 175 (2016) 838–850

because their orbits are represented by a perfect circle. However,

Efficiency
the rest of the orbital patterns have different shapes and they were

40.0%
94.9%
85.6%
65.9%
not always classified correctly. Unbalance patterns have the
highest correct recognition rate in the four models. Misalignment
and extreme misalignment patterns had a good recognition per-

Recalled patterns
formance using our proposed associative memory. Nevertheless,
those two databases presented high misclassified rates using the
binary Lernmatrix memories (LM and TLM). In this case, an
ANN

advantage of the real-valued Lernmatrix is that patterns with a

603
298
93
4 low noisy shape have low deviations according to the training set
and they can be perfectly recognized by the RLM. On the other
Efficiency

hand, a low deviation changes significantly binary patterns which


40.0%
69.2%
66.9%
40.4%
can be highly misclassified by the LM and the TLM. Extreme
misalignments had the highest misclassification rates in our model
due to they having significantly differences in the pattern shapes
Recalled patterns

(25.6% of error); however, unbalance and misalignment recogni-


tion processes have good classification rates (higher than 80%). The
ANN reports good classification rates in the unbalance and mis-
MAM

alignment classes. Furthermore, the good condition classes had


440
233
57
4

the higher misclassification rates for this classifier because the


whole database is unbalanced, and this generates a classification
Efficiency

trend to those classes with the higher number of patterns. This


90.0%
82.5%

25.5%

behavior is not present in the Lernmatrix due to there not being a


9.4%

convergence methodology in the learning process and the pattern


recovery is made using the previously computed associations in
Recalled patterns

the memory matrix (M). Finally, the MAM recognition process


presents rates below than the proposed RLM and ANN models.
This is caused due to saturation problems in the memory M,
whenever similar patterns that belong to different classes generate
TLM

524

36
33

this particular effect. An important remark is that recalling data-


9

bases were measured directly from real situations where signals


Efficiency

are exposed to noisy environments and human errors (such as bad


80.0%

28.3%
77.4%

sensor positioning, undesirable vibrations, etc.).


6.8%
Recalled patterns

10. Discussion and conclusions

This work proposed an alternative model for classifying


induction motor faults using two methodologies: orbital analysis
492
LM

and the Lernmatrix associative memory (as classifier). As a first


40
24
8

work, we introduced a computational model for classifying two


Efficiency

types of motor faults: unbalance and misalignment. The proposed


model uses orbital analysis for creating patterns according to an
88.1%
82.1%
74.4%
100%

orbital shape; however, the level of a motor fault was not con-
sidered in this work, were more detailed researches are needed in
order to establish shape-severity fault in a machine. In this sense,
Recalled patterns

as a future work this relation will be studied. As a first stage in this


Motor fault recognition results using the experimental database.

paper, orbital analysis provides a good tool for creating patterns


from certain types of faults in induction motors, since it creates a
RLM

perfect shape that can be directly related with a bad or good


560
286
105
10

condition in the machine. It is important to note, that other faults


can be present in the motor and they were removed in this work;
Number of patterns

nevertheless, more studies are needed. As a preliminary study, we


present a partial solution in motor fault classification where our
research suggest that new technics must be developed in order to
discover new orbit representations according to other motor
faults, including several aspects as new harmonics or different
348
635

141
10

ways for sensor positioning. Since conventional models have been


implemented using too complex methodologies that are difficult
Extreme misalignment

to reproduce, our approach gives a partial solution to the complete


set of motor faults with an easier way for detecting them, which
Good condition

can be easy to implemented in hardware restricted computational


Misalignment
Motor faults

systems (against the most powerful models). Therefore, our work


Unbalance

proposed orbital analysis is a feasible way to understand a motor


Table 2

fault without the use of complex systems. Orbital shapes are easy
to interpret and the can be combined with a huge set of
J. Juan Carbajal-Hernández et al. / Neurocomputing 175 (2016) 838–850 849

mathematical models; therefore, an associative memory is a good [7] M. Chow, Methodologies of using neural network and fuzzy logic technologies
technique for recognizing those shapes. for motor incipient fault detection, World Scientific, Singapore, 1997.
[8] L. Cleofas, V. García, R. oMartín, R. Valdovinos, J. Sánchez, O. Camacho, Hybrid
Furthermore, associative memories have demonstrated to be associative memories for imbalanced data classification: an experimental
good tools for the pattern processing field, where several types of study, Lect. Notes Comput. Sci. 7914 (2013) 325–334.
associative memories have been developed in order to correctly [9] B. Cruz, R. Barrón, H. Sossa, A new unsupervised learning for clustering using
geometric associative memories, Lect. Notes Comput. Sci. 5856 (2009)
classify patterns. Usually, those models has been built according to 239–246.
the pattern context, giving creative alternatives. Moreover, those [10] V. Climente, J. Antonino, M. Riera, R. Puche, L. Escobar, Application of the
models were implemented for binary image processing or having Wigner–Ville distribution for the detection of rotor asymmetries and eccen-
tricity through high-order harmonics, Electric Power Syst. Res. 91 (2012)
different types of architectures. In this sense, the proposed Lern- 28–36.
matrix was designed in order to create a classifier based on the [11] S. Dongfeng, O. Lianfsheng, B. Ming, Instantaneous purified orbit: a new tool
original Lernmatrix operation, where some modifications have for analysis of non-stationary vibration of rotor system, Int. J. Rotat. Mach.
7 (2) (2001) 105–115.
been made in order to increase its effectiveness or to extend its [12] E. Esmi, P. Sussner, M. Valle, F. Sakuray, L. Barros, Fuzzy associative memories
operation capabilities to other kinds of numerical patterns. In this based on subsethood and similarity measures with applications to speaker
sense, the use of the Lernmatrix provides an alternative way for identification. hybrid artificial intelligent systems, Lect. Notes Comput. Sci.
7209 (2012) 479–490.
recognizing real-valued patterns that can be processed in real [13] E. Esmi, P. Sussner, H. Bustince, J. Fernández, Theta-fuzzy associative memories
monitoring systems, having a good performance on classifying (Theta-FAMs) with stochastic weights, IEEE Trans. Fuzzy Syst. 23 (2015) 2.
motor fault patterns. Another advantage of our proposal against [14] F. Filippetti, G. Franceschini, C. Tassoni, Neural networks aided on-line diag-
nostics of induction motor rotor faults, IEEE Trans. Ind. Appl. 31 (4) (1995)
ANN´s, is that there are no convergence algorithms, which gives 892–899.
the certainty of a learning capacity, erasing the trial-error [15] Füssel, D. and Ballé, P., Combining neuro-fuzzy and machine learning for fault
approximation process that characterizes them. The success of diagnosis of a D.C. motor. in: Proceedings of the American Control Conference,
1997, pp. 37–41.
an associative memory depends in most cases on the problem [16] Gardel, P., Morinigo, D., Duque, O., Pérez, M. and Garcia, Neural network
itself. Due to this, a modification of the original binary Lernmatrix broken bar detection using time domain and current spectrum data, in: Pro-
for using with real-valued data was proposed. The results obtained ceedings of the 20th International Conference on Electrical Machines, No.
6350234: 2012, pp. 2492–2497.
in Table 1, show that the binary Lernmatrix had problems when [17] A. Gaylard, A. Meyer, C. Landy, Acoustic evaluation of faults in electrical
experimental patterns were processed. Nevertheless, our modified machines. Electrical Machines and Drives, 412, Conference Publication (1995),
Lernmatrix had a better performance with satisfactory results in p. 147–150. http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=497713.
[18] H. Grobler, G. Bright, R. Loubser, Vibration analysis of a rotor by considering
the classification process. One advantage to be considered in this misalignment and unbalance, 18th Int. Congr. Sound Vibr. 3 (2011) 2151–2158.
case, is that a binary pattern notably changes when some real [19] K. Ha, J. Hong, G. Kim, K. Chang, J. Lee, Orbital analysis of rotor due to elec-
values present slightly numerical deviations and it must be tromagnetic force for switched reluctance motor, IEEE Trans. Magn. 36 (4)
(2000) 1407–1411.
binarized. On the other hand, in the real-valued Lernmatrix those [20] S. Han, Retrieving the time history of displacement from measured accelera-
deviations do not represent a considerably pattern change. It is tion signal, J. Mech. Sci. Technol. 17 (2) (2003) 197–206.
concluded that the combination of the orbital analysis with the [21] S. Ho, K. Lau, Detection of faults in induction motors using artificial neural
networks. Electrical machines and drives, 412, IEEE Conference Publication
modification of the binary Lernmatrix into a real-valued Lernma- (1995), p. 176–180. http://ieeexplore.ieee.org/xpl/articleDetails.jsp?reload=true&
trix helps to implement an efficiency model for motor induction arnumber=497719.
fault identification. It is worthwhile mentioning that this paper [22] Hopfield, J.J., Neural networks and physical systems with emergent collective
computational abilities, in: Proceedings of the National Academy of Sciences,
provides a preliminary study of fault identification in induction vol. 79, 1982. pp. 2554–2558.
motors using orbital pattern analysis and future works are needed, [23] [IEEE] Institute of Electrical and Electronics Engineers, IEEE standard for
e.g. including additional patterns from different mechanical faults binary floating-point arithmetic, ANSI/IEEE standard 754, New York 22 (2)
(1985) 9–25.
in order to expand the capacities of the system, using a better [24] Iorgulescu, M., Beloiu, R. and Cazacu, D., Vibration monitoring for electrical
algorithm that further increases the performance obtained in this equipment faults detection using fast fourier transform. in: Proceedings of the
work. Thus, this model can be used as an alternative tool for 1st International Conference on Manufacturing Engineering, Quality and
Production Systems, 1, 2009, pp. 34–38.
preliminary motor fault classifications, when the good functioning [25] [ISO] International Organization for Standardization. 1995. ISO/10816,
of the machine is essential in critical time production industry. Mechanical vibration: evaluation of machine vibration by measurements on
non-rotating parts.
[26] L. Jun, J. Yu, L. Chih, T. Li, C. Hsin, Motor shaft misalignment detection using
multiscale entropy with wavelet denoising, Expert Syst. Appl. 37 (10) (2010)
Acknowledgments 7200–7204.
[27] Y. Katori, Y. Otsubo, M. Okada, K. Aihara, Associative memory network with dynamic
synapses, Advances in Cognitive Neurodynamics, Springer (2015), p. 479–483. http://
The authors of the present paper would like to thank to Center dx.doi.org/10.1007/978-94-017-9548-7_68.
of Computer Research of the National Polytechnic Institute for its [28] R. Kechida, A. Menacer, H. Talhaoui, Approach signal for rotor fault detection
support to develop this work and the project Grants SIP_20140597. in induction motors., J. Fail. Anal. Prevent. 13 (3) (2013) 346–352.
[29] K. Kim, A. Parlos, Induction motor fault diagnosis based on neuropredictors
and wavelet signal processing, IEEE/ASME Trans. Mechatron. 7 (2) (2002)
201–219.
References [30] T. Kohonen, Correlation matrix memories, IEEE Trans. Comput. C-2I (4) (1972)
353–359.
[31] B. Kosko, Bidirectional associative memories, IEEE Trans. Syst., Man Cybern.
[1] M. Aldape, I. Román, O. Camacho, Thresholded learning matrix for efficient 18 (1) (1998) 49–60.
pattern recalling. Lecture Notes in Computer Science, 5197, Springer Verlag, [32] Y. Lee, C. Lee, Modelling and analysis of misaligned rotor–ball bearing systems,
Berlin Heidelberg (2008), p. 445–452. J. Sound Vibr. 224 (1999) 17–32.
[2] Austin, J. 1987. ADAM: A distributed associative memory for scene analysis, in: [33] Z. Li, J. Zhu, X. Shen, C. Zhang, J. Guo, Fault diagnosis of motor bearing based on
Proceedings of First International Conference on Neural Networks, pp. 285–295. the Bayesian network, Procedia Eng., Elsevier 16 (2011) 18–26.
[3] K. Arun, A. Mohanty, Model based fault diagnosis of a rotor–bearing system for [34] B. Liang, S. Iwnicki, Y. Zhao, Application of power spectrum, cepstrum, higher
misalignment and unbalance under steady-state condition, J. Sound Vibr. order spectrum and neural network analyses for induction motor fault diag-
327 (3–5) (2009) 604–622. nosis, Mech. Syst. Signal Process. 39 (1–2) (2013) 342–360.
[4] T. Banerjee, S. Das, Multi-sensor data fusion using support vector machine for [35] Liu, D., Zhao, Y., Yang, B., Sun, J., A new motor fault detection method using
motor fault detection, Inform. Sci. Elsevier 217 (2012) 96–107. multiple window S-method time-frequency analysis. in: Proceedings of the
[5] J. Carbajal, L. Sánchez, New algorithm for efficient pattern recall using a static International Conference on Systems and Informatics, 2012, pp. 2563–2566.
threshold with the Steinbuch Lernmatrix, Connect. Sci. 23 (1) (2011) 33–44. [36] D. Matić, F. Kulić, SVM broken bar detection based on analysis of phase cur-
[6] F. Chen, S. Jhe, P. Min, T. Wen, Study of start-up vibration response for oil whirl, rent. 5th International Power Electronics and Motion, Control Conf. Expos.
oil whip and dry whip, Mech. Syst. Signal Process. 25 (8) (2011) 3102–3115. (2012), No. 6397456, pp. LS4c.21–LS4c.24.
850 J. Juan Carbajal-Hernández et al. / Neurocomputing 175 (2016) 838–850

[37] Masoud Haji, Hamid A. Toliyat, A technique for induction machines rotor José Juan Carbajal Hernández is a Research Professor of
broken bar detection, IEEE Trans. Energy Convers. 16 (4) (2001) 312–317. Intelligence Systems at Center for Computing Research
[38] Obaid, R., Habetler, T. and Tallam, R., Detecting load unbalance and shaft of National Polytechnic Institute, Mexico. Actually, his is
misalignment using stator current in inverter-driven induction motors, Elec- member of the National Researchers System of CON-
tric Machines and Drives Conference, IEMDC. IEEE International, 3, 2003, ACyT, México. His research interests are electronic
pp. 1454–1458. instrumentation, pattern recognition and environ-
[39] S. Orfanadis, Introduction to Signal Processing, Prentice Hall, 2009. mental modeling
[40] G. Palm, F. Schwenker, F.T. Sommer, A. Strey, Neural associative memories, in:
A. Krikelis, C.C. Weems (Eds.), Associative Processing and Processors, IEEE
Computer Society, Los Alamitos, 1997, pp. 307–326.
[41] E. Palomino, A. Sánchez, J. Cabrera, L. Sexto, Preliminary diagnosis of rotational
machinery. Experiences in the implementation of a predictive maintenance
program and certification of human resources in a Cuban cement industry.
Advances in vibration control and diagnostics, Polimetr. Int. Sci. (2006)
177–184.
[42] J. Proakis, D. Manolakis, Tratamiento digital de señales, 4a Ed., Pearson Edu- Luis Pastor Sánchez Fernández is a Research Professor
cation, España, 2007. of Automation and Control at Center for Computing
[43] Refaat, S., Abu, H., Saad, S., Aboul, E. and Iqbal, A., ANN-based for detection, Research of National Polytechnic Institute, Mexico.
diagnosis the bearing fault for three phase induction motors using current Actually, his is member of the National Researchers
signal, in: Proceedings of the IEEE International Conference on Industrial System of CONACyT, México. His research interests are
Technology, 6505681, 2013, pp. 253–258. virtual instrumentation, digital signal processing,
[44] G. Ritter, P. Sussner, J. Diaz, Morphological associative memories, IEEE Trans. automation and control theory.
Neural Netw. 9 (2) (1998) 281–293.
[45] Ruz, J., Suárez, D., García, R., Sánchez, E. and Suárez, M., Optimal training
algorithm application to design an associative memory for fault diagnosis at a
fossil electric power plant, in: Proceedings of the 8th IFAC Symposium on Fault
Detection, Supervision and Safety of Technical Processes. Mexico City, Mexico,
2012, pp. 756–762.
[46] J. Ruz, E. Sánchez, D. Suárez, Optimal training for associative memories:
application to fault diagnosis in fossil electric power plants. Hybrid Intelligent
Systems: Studies in Fuzziness and Soft Computing, 208, Springer Verlag
(2007), p. 329–356. http://dx.doi.org/10.1007/978-3-540-37421-3_21. Ignacio Hernández Bautista is graduated in electronic
[47] A. Salavati, K. Kumar, A. Shokrollahi, Nonbinary associative memory with engineering form ESIME-IPN in 2002. He obtained MSc
exponential pattern retrieval capacity and iterative learning, IEEE Trans. degree in computer engineering, digital systems spe-
Neural Netw. Learn. Syst. 25 (3) (2014) 557–570. cialty in 2009, and Sc. D degree in computer science in
[48] M. Seera, C. Lim, D. Ishak, H. Singh, Application of the fuzzy min-max neural 2014; both degrees are from the Center for Computing
network to fault detection and diagnosis of induction motors, Neural Comput. Research of National Polytechnic Institute, Mexico. His
Appl. 23 (2012) 191–200. research interests include image processing applica-
[49] k Steinbuch, Die Lernmatrix, Kybernetik 1 (1) (1961) 36–45. tions, image compression, neural networks and pattern
[50] k Steinbuch, H. Frank, Nichtdigitale Lernmatrizen als perzeptoren, Kybernetik recognition.
1 (3) (1961) 117–124.
[51] C. Sun, Z. Duan, Y. Yang, M. Wang, L. Hu, The motor fault diagnosis based on
neural network and the theory of D-S evidence. Advanced Materials Research,
683, Trans Tech Publications (2013), p. 881–884. http://dx.doi.org/10.4028/
www.scientific.net/AMR.683.881.
[52] A. Vázquez, J. Sossa, A bidirectional hetero-associative memory for true-color
patterns. Neural Process Letters, 28, Springer (2008), p. 131–153. http://dx.doi. José de Jesús Medel Juárez is a Professor developing on
org/10.1007/s11063-008-9086-9. intelligent estimation and identification and now in
[53] A. Verma, S. Sarangi, M. Kolekar, Stator winding fault prediction of induction compound and hybrid filters. He is an Aeronautical
motors using multiscale entropy and grey fuzzy optimization methods, Engineering (1994), Master Sciences (1996), and also a
Computers & Electrical Engineering, 40, Elsevier (2014), p. 2246–2258. http:// Doctor Sciences in Automatic Control (1998), Member
dx.doi.org/10.1016/j.compeleceng.2014.05.013. of National Council for Science and Technology (1999),
[54] [VDI] The German Association of Engineers. 1964. VDI/2056. Standards of member of the Mexican Academy of Sciences. He is a
evaluation for mechanical vibrations of machines. Germany. Full Professor and Researcher in the Computer Research
[55] J. Wang, R. Gao1, R. Yan, Broken-rotor-bar diagnosis for induction motors, Center (CIC, by the Spanish Acronym). His results
J. Phys. 305 (2011) 1–10. include: Digital filtering, Control, Real-time, and others.
[56] C. Xiu, C. Liu, Y. Cheng, Associative memory network and its hardware design. He is a laboratory member of the intelligent systems for
Neurocomputing, 158, Elsevier (2015), p. 204–209. http://dx.doi.org/10.1016/j. automation.
neucom.2015.01.047.
[57] L. Yahya, T. Saleem, B. Hasan, The application of neural network to electrical
motor's sound recognition system, J. Eng. Appl. Sci. 7 (2) (2012) 191–193.
[58] Yang, D., Induction motor bearing fault diagnosis using Hilbert-based bis-
pectral analysis, in: Proceedings of the International Symposium on Computer, Luis Alejandro Sánchez Pérez received his M.Sc. degree
Consumer and Control, IS3C. Taichung, Taiwan, 2012. with honors from the Computing Research Center of
[59] D. Zhen, T. Wang, F. Gu, A. Ball, Fault diagnosis of motor drives using stator the National Polytechnic Institute in Mexico, 2011. He is
current signal analysis based on dynamic time warping, Mech. Syst. Signal a Computer Systems Engineer graduated from the
Process. 34 (1–2) (2013) 191–202. Higher School of Computing in Mexico, 2008. He is
[60] P. Zheng, Threshold complex-valued neural associative memory, IEEE Trans. currently enrolled in Ph.D. program at the same center.
Neural Netw. Learn. Syst. 25 (2014) 9.
[61] F. Zidani, M. Benbouzid, D. Diallo, M. Naït, Induction motor stator faults
diagnosis by a current concordia pattern-based fuzzy decision system, IEEE
Trans. Energy Convers. 18 (4) (2003) 469–475.
[62] López, R., Sánchez, L., Progrebnyak, O. & Costa, A. 2008. Inter-turn Short Circuit
and Unbalanced Voltage Pattern Recognition for Three-Phase Induction Motors.
Lecture Notes in Computer Science, Springer Verlag, Vol. 5197: 470-478.