Journal Pre-proof

A novel feature extraction method for bearing fault classification with one
dimensional ternary patterns

Melih Kuncan, Kaplan Kaplan, M. Recep Minaz, Yılmaz Kaya, H.

Metin Ertunç

PII: S0019-0578(19)30486-0
DOI: https://doi.org/10.1016/j.isatra.2019.11.006
Reference: ISATRA 3393

To appear in: ISA Transactions

Received date : 26 March 2019

Revised date : 10 October 2019
Accepted date : 4 November 2019

Please cite this article as: M. Kuncan, K. Kaplan, M.R. Minaz et al., A novel feature extraction
method for bearing fault classification with one dimensional ternary patterns. ISA Transactions
(2019), doi: https://doi.org/10.1016/j.isatra.2019.11.006.

This is a PDF file of an article that has undergone enhancements after acceptance, such as the
addition of a cover page and metadata, and formatting for readability, but it is not yet the definitive
version of record. This version will undergo additional copyediting, typesetting and review before it
is published in its final form, but we are providing this version to give early visibility of the article.
Please note that, during the production process, errors may be discovered which could affect the
content, and all legal disclaimers that apply to the journal pertain.

© 2019 Published by Elsevier Ltd on behalf of ISA.

*Title page showing Author Details

Journal Pre-proof

Melih KUNCAN, PhD

Siirt University, Engineering Faculty C Block, Electrical and Electronics Engineering
Department, Kezer Campus, Batman Road 10.km.
Merkez/SİİRT, 56100

Kaplan KAPLAN, PhD Candidate

Kocaeli University, Engineering Faculty A Block, Department of Mechatronics Engineering,
Umuttepe Center Campus, Old Istanbul Road 10. km.
İzmit / Kocaeli, 41380

of
M. Recep MİNAZ
Siirt University, Engineering Faculty C Block, Electrical and Electronics Engineering
Department, Kezer Campus, Batman Road 10.km.

pro
Merkez/SİİRT, 56100

Yılmaz KAYA, PhD

Siirt University, Engineering Faculty C Block, Computer Engineering Department, Kezer
Campus, Batman Road 10.km.
Merkez/SİİRT, 56100
re-
H. Metin ERTUNÇ, PhD
Kocaeli University, Engineering Faculty A Block, Department of Mechatronics Engineering,
Umuttepe Center Campus, Old Istanbul Road 10. km.
İzmit / Kocaeli, 41380
lP

Tel: + 90 (484) 223-1224 / ext. 3033

Corresponding author
E-mail: [email protected]
a
urn
Jo
*Highlights (for review)

Journal Pre-proof

> In this study, a novel approach based on 1D-LTP is proposed to extraction of quantitative
features from bearing vibration signals

> One advantage is that this method uses all data points for feature extraction

> It is fast and can be use in real-time application

> High accuracies achieved for bearing fault classification

of
> Original data (Experimental setup of authors)

pro
re-
a lP
urn
Jo
*Blinded Manuscript - without Author Details
Click here to view linked References

Journal Pre-proof

A Novel Feature Extraction Method for Bearing Fault Classification with One
Dimensional Ternary Patterns

Abstract

Bearing is one of the most critical parts used in rotary machines. Bearing faults break down

of
the mechanism where it is located. Moreover, the faults may cause to malfunction by
spreading to the entire system. Thus this may result in catastrophic failure eventually. Precise
and decisive feature extraction from the raw vibration signal maintains to be one of the

pro
current topics explored for fault diagnosis in bearings. In this study, vibration signals are
obtained from bearings which are formed with artificial faults of specific dimensions from a
bearing test setup. Instead of employing traditional feature extraction methods found in the
literature, a novel feature extraction method for bearing faults called one-dimensional ternary
re-
pattern (1D-TP) is applied. The proposed approach is a statistical method that uses patterns
obtained from comparisons between neighbors of each value on vibration signals. The study
aims to identify the size (mm) of the fault by determining the bearing part (inner ring, outer
ring, ball) from which the faults in the bearings are caused. Several classification techniques
lP

were performed by using ternary patterns with RF (Random Forest), k-NN (k-nearest
neighbor), SVM (Support Vector Machine), BayesNet, ANN (Artificial Neural Networks)
models. As a result of analyzing the signals obtained from the experimental setup with the
proposed model, 91.25% for dataset_1 (different speed), 100% for dataset_2 (fault type -
a

inner ring, outer ring, ball) and 100% for dataset_3 (fault size (mm)) success rates are
urn

determined.

Key Words: Feature Extraction, 1D-TP, Fault classification, diagnosis, artificial intelligence,
bearing fault in servo-motor
Jo

1. INTRODUCTION

Machine motion systems are usually provided with bearing-based rotational mechanisms.
Bearings are high-precision components that allow machines to move at extremely high
speeds and transport large loads efficiently and effectively. In addition to offering high
precision, reliability, and durability, the bearings provide the ability to rotate at high speeds
 Journal Pre-proof

with minimum speed, vibration, and friction force. During the process, a bearing is exposed to
friction forces by forming an interface between the stationary and moving part in the machine.
Also, the bearings are affected by overheating or overloading contamination and
manufacturing faults. This kind of faults that can occur in the roller bearing, which is an
indispensable component in automation systems can cause failure to the operation of the
system part and the malfunction in the machine. Predicting bearing faults in advance and
changing the bearing without causing significant damage in the machine is essential in terms

of
of machine health. The presence and level of fault in rotary systems can be determined by
vibration signals. For this reason, the machines must be observed in such a way that the

pro
standard operating vibrations remain the same. Vibration caused by the system or malfunction
by appropriate methods can be detected by various fault monitoring methods. Bearings can be
used in all rotating applications such as cars, airplanes, construction equipment, power tools,
electronics. The bearings are employed as the most important fundamental parts of the mobile
machines. Therefore, the detection fault type and size of bearing has great importance for
re-
efficient machine life [1-9].

The diagnosis or determination of faults in bearings is still one of the state-of-the-art topics.
Researchers usually use data from the on-line database (benchmark) in order to model the
lP

bearing signal. Tahir et al. classified local faults in ball bearings in the time domain by using
statistical features (RMS, mean, variance, skewness, kurtosis, crest factor, impulse factor,
shape factor, median, and range). To increase the ability of classifiers to distinguish, the
central tendency-based feature suggested as a pre-processing technique. In this way, they
a

reported that they reduced the effects of fluctuations and increased the accuracy of classifiers.
urn

They used Support Vector Machine(SVM), Bayesian network (BN), decision table, and
decision tree to classify inner and outer bearing defects. They declared that they classified the
features with a very high success rate with this pre-processing method [10]. Wang et al.
proposed a new hybrid approach to the random forest classifier for fault diagnosis of bearings.
They extracted fault features by applying Wavelet Packet Decomposition. They defined the
Jo

best central wavelet cluster for signal preprocessing with signal-to-noise ratio and mean
square error values. They stated that they classified as normal, inner, outer, and ball faulty
bearings with high accuracy of 88.23% using the method proposed in the comparative test
results [11]. Ding et al. proposed a new fusion feature based on locality preserving projections
(LPP) in order to investigate the active role of the excellent mold in error classification.
Individually, the proposed fusion feature was performed in two main steps. In the first step,
 Journal Pre-proof

they created a two-class model for both classes, primarily using this signal class and good
signals. Then, they created a fusion map for mathematically combining the maps of LPP or
for the improvement of all two-class models. The authors suggested that the local and global
principal component analysis method had a better performance than the traditional LPP,
whereas the test recognition rate was between 94.2 - 98.8% [12]. Yan et al. proposed a new
fault classification algorithm based on a multi-domain optimized SVM consisting of three
stages (i.e., multi-domain feature extraction, feature selection, and multi-field feature

of
extraction). In the first phase, they applied three approaches (i.e., statistical analysis, FFT and
VMD) separately to extract the fault property information from the multi-domain aspect (e.g.,

pro
time domain, frequency domain, and time-frequency domain). They employed Laplace
scoring algorithm for the feature selection. Finally, they have used the support vector machine
classification model based on particle swarm optimization (PSO-SVM) to implement the
identification of the multiple fault status of the bearings. According to experimental results,
multi-domain features provide higher diagnostic accuracy than single domain features [13].
re-
Ertunc et al. have developed a multi-step decision algorithm based on ANN and ANFIS
models. Both time and frequency domain feature obtained from vibration and current signals
were used to train ANN and ANFIS models. With the multi-stage classification model, they
lP

classify normal, an inner ring or outer ring fault as a heavy, medium, and new fault conditions
with high accuracy. They stated that the ANFIS method performed classification with higher
accuracy than the ANN method [14]. Kavathekar et al. extracted statistical features from the
data obtained from Case Western Reserve University. They then subdivided the features with
a

Principal Component Analysis (PCA) analysis. They have calculated new features with axis
urn

transformations. Thus, they declared that they increased fault classification accuracy.
According to their observation, Rotation Forest Technique performed better than other ANN,
ANFIS, and Random Forest (RF) Techniques [15]. Zhang et al. proposed Improved Symbolic
Aggregate Approach (ISAX) method by developing the Symbolic Aggregate Approach
(SAX) technique that transforms a real-time series into a symbol sequence. They stated that
Jo

with the new method, the characteristics of the mean value within the time frame have
changed with the multiple properties obtained from time, frequency, and time-frequency
domains to obtain comprehensive fault information. After transforming a vibration signal into
a set of symbolic sequences according to the multiple field characteristics by ISAX
transformation, they used Shannon entropy values as the eigenvalues of the symbol sequence.
They used the Kernel Principal Component Analysis (Kernel PCA) to reduce the dimensions
 Journal Pre-proof

of these eigenvalues and to extract key features for the classification tasks. The SVM and
ISAX method achieved better accuracy than other conventional methods [16]. Attoui et al.
have proposed a time-frequency procedure which includes both energy of classic wavelet
decomposition packet and also a new feature extraction technique. Initially, the most impulse
frequency bands were selected under different band conditions using a combination of Fast-
Fourier-Transform FFT and Short Time Fourier Transform (STFT) algorithms. In order to
maximize the detection of the frequency content, the Wavelet Packet Decomposition (WPD)

of
technique was used to separate the features into different frequency subbands. They have used
the STFT to know exactly which frequency components are obtained in these frequency

pro
subbands. After obtaining the feature vector, the three feature size selection techniques,
Linear Discriminant Analysis (LDA), a feedback wrapper method, and Locality Sensitive
Discriminant Analysis (LSDA) were used. Finally, the ANFIS algorithm used for the
immediate identification and classification of bearing faults. To evaluate the performance of
the proposed method, different test data were adjusted to the trained ANFIS model using
re-
different healthy and faulty bearing conditions under various load levels, fault severity, and
rotational speed. They reported that they achieved 99.83% success with LSDA and ANFIS
models [17]. Singh et al. have aimed to determine incorrect bearing detection, classification,
lP

and location in a three-phase asynchronous motor by using the Stockwell transformation and
the SVM. They applied the Stockwell transformation to the stator current signals to extract
some features in both the time and frequency domain. Many unrelated and high-ranking
features were selected according to the Fisher score order. These features were used in order
a

to classify ball, inner and outer ring defects using the SVM. After detecting the fault, they
have used Stockwell transform to determine whether the faulty bearing is on the fan side or
urn

the load side. They stated that they applied this algorithm with 91.66% success on the
experimental data of faulty bearings collected from the industry [18]. Jiang et al. suggested
semi-supervised kernel Marginal Fisher analysis (SSKMFA) that is a feature extraction
method that can explore the intrinsic manifold structure of this model data set and can
simultaneously take into account the inter-class compactness and inter-class discrimination.
Jo

SSKMFA method is more decisive feature since this method obtains their low-dimensional
features directly from the raw high-dimensional vibration signals by using both the labeled
and unlabeled samples of the internal manifold structure. These features are classified by K-
nearest neighbor (k-NN), which is one of the simplest methods of classification to recognize
different fault categories and the severity of the faults by making them optimally low
dimensional. They indicated that the proposed approach improves the fault recognition
 Journal Pre-proof

performance of this approach, and they perform classification with an accuracy rate of 98.5%
[19]. Kaplan et al. used the vibration data obtained from the experiment set that they created
to determine the size of the fault in the inner ring. Then, they used ANN, ANFIS, SVM
method in order to classified fault dimensions by extracting statistical features. According to
the authors' previous studies, they achieved a 100% accuracy by performing a more efficient
classification with the SVM model [20-24].

of
As seen in literature, some of the features extraction methods implemented in previous
research include Discrete Wavelet Transform, FFT, STFT, WPD, and other methods. Every
method has its own advantages and limitations. Bearing vibration signals are nonlinear, time-

pro
varying, and unstable. It is well-know that FFT is a global linear method. However, a
frequency spectrum does not reflect the frequency changes over the time domain; thus, FFT,
STFT has some limitations when applied to non-stationary signal analysis. The features
extracted by such methods exhibit poor compatibility. The future work must primarily focus
re-
on feature extraction from a bearing vibration signal using more statistical data. In this study,
a new feature extraction method is proposed for vibration signals. One-dimensional local
ternary patterns (1D-TP) method was used to extract features from raw vibration signals. The
TP method is generally used in image processing for texture analysis. In the study, the 1D-TP
lP

method has been developed to describe the sharp waves in the vibration signals adequately.
The most important advantages of the proposed method are; the first is that it uses all the
values on the vibration signal, and the second one is that it is simple, fast, and real-time
applicable. According to the results of the study, it was observed that the 1D-TP approach
a

provides efficient features for classification of vibration signals.

urn

2. EXPERIMENTAL TEST SETUP AND GENERATION OF DATA

The data used in the study was obtained from the bearing-shaft assembly connected to an AC
servo motor, as in Figure 1. The test setup consists of an AC servo motor, LVD drive, two-
axis vibration sensor, radial, and axial pressure valves, NIDAQ 6211 data acquisition card
Jo

(DAQ), and signal conditioners. This type of servo motors is generally two-phase squirrel
cage asynchronous motors. Two-phase asynchronous motors are made of large-power but
mostly used in automatic control systems as small servo motors. As there is no brush and
collector, they are less likely to malfunction and easy to maintain. The purpose of this engine
is to achieve the desired high-speed levels by minimizing external vibrations. The LVD (Low
Voltage Directive) is used as a servo motor driver. PCB 352C65 SN77267 sensor is used to
 Journal Pre-proof

sense the vibrations. These sensors operate according to gravity and with the resulting
vibrations; minimal signals are obtained using piezoelectric material. The data collected via
the sensor was passed through the amplifier and sent to the NIDAQ card. The signal from the
vibration sensor is converted to an electrical signal using a signal amplifier. The condenser
with one channel output can rise to 200 kHz signal. It can also output 24 Volt sensors from 2
mA to 20 mA. DAQ cards, which simplify testing and measurement, are the most suitable
systems for data collection. For the bearing data collection system, the NIDAQ - 6211 card is

of
used. There are also signal conditioners associated with the sensor to regulate and amplify the
vibration signals. Figure 2 shows a schematic representation of the test setup. [25].

pro
re-
a lP

Figure 1. The experimental setup

urn
Jo
 Journal Pre-proof

of
pro
Figure 2. Vibration data collection scheme [14]
re-
The main reason for bearing faults is mechanical friction. Usually, these errors take a long
time to form. In order to make an experimental observation in the laboratory environment,
artificial fault in certain diameters have been formed. Artificial faults must be carefully
lP

created in order to prevent the effects of an extra vibration on the collected data except for the
bearing faults. For this purpose, holes were drilled with 0.15 mm, 0.5 mm and 0.9 mm
diameter on the inner ring, outer ring and the ball with micron precision. Figure 3 shows the
laser-generated artificial faults, and Figure 4 shows a microscopic view of a sample fault.
a
urn
Jo

Figure 3. A laser-generated sample – artificial fault in the outer ring

 Journal Pre-proof

of
pro
Figure 4. A microscopic view of a laser-generated sample

Vibration data were collected at a sampling frequency of 24 kHz, during 20 second and stored
as a raw data packet. At the first stage, the data obtained from an error-free bearing was
recorded. The vibration sensors are connected to the bearing experimental setup to the motor
re-
shaft in the X and Y-axis directions. For this study, only data was collected from the
accelerometer sensor on the Y-axis. One bar axial and radial load pneumatic pressure is
applied to the bearing holding mechanism while data is collecting. The data obtained from the
lP

artificial faults (mm) in the inner ring (IR), outer ring (OR) and balls of the bearings (BS)
were recorded. The tests were repeated for five different speeds determined at the data
collection stage. Thus, different speed, different fault size (mm) and different fault types
(inner, outer ring, and ball fault) were obtained from the bearings. The following table shows
a

the different bearing fault details tested in the test setup for operation.
urn

Table 1. Dataset1 - data set created for different speeds

Dataset1 Artificial Fault Type (Speed)

Experiment IR (mm) OR (mm) BS (mm) Speed (rpm)
Number:
1 0.15 0.15 0 1500
Jo

2 0.15 0.15 0 1740

3 0.15 0.15 0 1800
4 0.15 0.15 0 1860
5 0.15 0.15 0 2100

Table 2. Dataset2 – data set generated for bearing fault types

Dataset2 Artificial Fault Type (variety)

 Journal Pre-proof

Experiment IR (mm) OR (mm) BS (mm) Speed (rpm)

Number
1 0.9 0.0 0.0 1500
2 0.0 0.9 0.0 1500
3 0.0 0.0 0.9 1500

Table 3. Dataset3 - data set for bearing error sizes

Dataset3 Artificial Fault Type (size)

of
Experiment IR (mm) OR (mm) BS (mm) Speed (rpm)
Number
1 0.00 0.00 0.00 1500

pro
2 0.15 0.15 0.15 1500
3 0.5 0.5 0.5 1500
4 0.9 0.9 0.9 1500

In this study, three different data sets were used. In the first data set, the signals are obtained
re-
at different rpm. Before the feature extraction step with the proposed method, the raw
vibration signal was divided into data packages containing 10000 samples. For visualization
of the signals, the sample vibration signals at 1500, 1740, 1800, 1860, and 2100 rpm are given
in Figure 5.
a lP
urn
Jo
 Journal Pre-proof

of
pro
re-
Figure 5. Sample signals obtained at different rpm (A) 1500 rpm, (B) 1740 rpm, (C) 1800
rpm, (D) 1860 rpm, (E) 2100 rpm.

In the second data set, there are signals of different fault types at 1500 rpm. In other words,
lP

this dataset contains the measured signals on the inner, outer ring, and the ball. The sample
signals of these fault conditions are given in Figure 6.
a
urn
Jo

Figure 6. Signals of different fault types (A) Inner ring fault, (B) Outer Ring Fault (C) Ball
Fault
 Journal Pre-proof

In the last data set, there are sample signals at different fault sizes taken at 1500 rpm. Faults
consist of vibration signals with inner ring, outer ring, ball in the size of 0.00mm, 0.15 mm,
0.5mm, 0.9mm. Samples of these faulty signals are given in Figure 7.

of
pro
re-
Figure 7. Sample signals of different fault sizes (A) 0.15mm Fault, (B) 0.9mm Fault, (C)
0.5mm Fault (D) Fault-free signals
lP

3. METHOD

3.1. Feature Extraction using One-Dimensional Ternary Patterns

a

One-dimensional ternary patterns (1D-TP) method was developed from the local ternary
patterns, which is widely used in image processing techniques. The developed method is used
urn

for feature extraction from one-dimensional signals. The value of each member on a vector of
signal is compared to neighbors and this comparison information is considered as new values
on the signal [18]. Binary codes are generated as a result of comparisons between neighbors
for each value of the signal. The decimal responses of these codes are considered as a new
Jo

signal expressing the signal. In the 1D-TP method, P represents the neighboring number of
each value on the signal. The number of P/2 neighbors is taken previously and after each
value of the signal. In Figure 8, the steps of the 1D-TP method are shown on an example. As
seen in the figure, four neighbor values were taken before and after the Pc (center) value [26].
 Journal Pre-proof

of
pro
re-
lP

Figure 8. Feature extraction steps: (A) A section of sample vibration signals, (B) signal
values, (C) comparison of the Pc with the Pi, (D) Separation of positive and negative values,
a

(E) conversion binary values to decimal.

urn

As shown in Figure 8 (A), PLEFT = {P0, P1, P2, and P3} from the left and PRIGHT = {P4, P5, P6,
and P7} from the right of the Pc, has total of 8 neighbor values. The parameter P denotes the
total number of neighbor to be considered. In this study, the P-value is taken 8. The parameter
Jo

(β) in the developed method is assigned by the user. The comparison of the center value with
the neighbor values is carried out with the following Equation (Figure 8 (B)).

(1)
 Journal Pre-proof

Ternary patterns are obtained depending on the parameter (β). The value of this parameter is
determined by the user according to the amplitude of the signal. In this study, the value of this
parameter is determined by the user according to the standard deviation of neighboring
values. The local variation changes with ±β. Therefore, different values of β should be tried
for different signals. The β parameter is vital for obtaining different patterns. Besides, this

of
parameter can be considered as bandwidth on the signals. With this parameter, the center
value is fixed with ± β. Therefore, as seen in Figure 9, neighbors for each Pc value are filtered
according to the β parameter.

pro
re-
a lP
urn

Figure 9. Filtering by different β values

Jo

After the filtering by β parameter, two decimal values are generated with comparisons for the
Pc value. While the sub-features are produced by the negative ones ( ), the upper
properties are obtained by the positive ones ( ). Different values obtained from
the lower and upper values are stored as two different values of the center value (Figure 8
(C)). After applying these operations to the whole signal, two different signals are obtained.
 Journal Pre-proof

Two different sets of features are obtained from the histograms of these two different signal
groups. In the case of P = 8, 28 = P = 256 features are obtained from each signal. Histograms in
the form of different lengths can be obtained with P. In the case where P is 6, 8, 10,
respectively 64 (26), 256 (28) and 1024 (210) features can be produced. In this proposed
method, P and β parameters play an important role in obtaining different patterns. The values
of these two parameters should be decided by trial and error. The 1D-TP signals obtained by
applying the 1D-TP method to the sample vibration sets and the histograms of the sub-

of
features extracted from these signals are shown in Figure 10.

pro
re-
lP

Figure 10 (A) Sample signal groups of fault types (B) Transformation of sub-signals by 1D-
TP method, (C) Histograms of sub-signals, (D) Transformation of upper-signals by 1D-TP
a

method, (E) Histogram of upper-signals

urn

3.2. Proposed Method

In this study, a completely different approach was proposed from previous studies for the
classification of bearing vibration signals. The block diagram of the classification steps of the
vibration signals is given in Figure 11.
Jo

Figure 11: Block diagram of the proposed method.

 Journal Pre-proof

Block 1: This block indicates the vibration signals obtained from the raw vibration data of
bearing for the transformation of 1D ternary patterns.

Block 2: The block where the 1D-TP method is applied. This process was explained in
section 3.1. After applying this method to the signals, 1D-TP signals are obtained. The values
of these signals vary between 0-255. After applying this transformation to the signals, two
different signals are obtained in the form of upper and lower. These signals are shown as in

of
Figure 12.

pro
re-
Figure 12. Illustration of the 1D-TP method. (A) A sample raw vibration signal, (B, C) Up
and low signals obtained after 1D- TP transformation.

Block 3: At this stage, the histograms of the up and low marks are generated after applying
lP

the 1D-TP method to the vibration signals. Each value in the histogram indicates a different
pattern. If P = 8, then 256 patterns are obtained. The histograms obtained for the example
markings in Figure 12 (A) are shown in Figure 13.
a
urn
Jo

Figure 13. Low (A) and up (B) feature histograms obtained for the sample signal

Block 4: This block refers to the classification stage. In this study, different classification
methods have been used. Classification methods, namely, Support Vector Machine (SVM),
 Journal Pre-proof

decision trees such as Random Forest (RF) and Functional Tree (FT), Bayes Net (BayesNet)
and Artificial Neural Networks (ANN) methods were used as a classifier. The classification
process was performed according to the 10-fold cross-validation test.

3.3. Performance metrics

A confusion matrix points out the performance of a classifier on a test data for which
corresponding the valid values are determined in advance. To understand the confusion

of
matrix, some definitions must be known.

 True Positives (TPs): Actual status is positive, but the prediction is positive.

pro
 True Negatives (TNs): Actual status is negative, but the prediction is negative.
 False Positives (FPs): Actual status is negative, but the prediction is positive.
 False Negatives (FNs): Actual status is positive, but the prediction is negative.

According to the confusion matrix, some performance metrics can be obtained. The most
re-
popular and straightforward method for determining the performance of the proposed model
was the accuracy rate of the model. The accuracy is defined as the ratio of the (TP + TN) to
the (TP + TN + FP + FN) [27-29]. Accuracy, error-rate, precision, and f-measure metrics
were used in this study. The metrics values are calculated by the following equations [30].
lP

(2)

(3)
a

(4)
urn

(5)

(6)
Jo

4. RESULTS

4.1. Ternary Patterns

A new feature extraction approach is proposed for the classification of bearing vibration
signals. The proposed method is based on statistical information obtained as a result of
comparisons between the neighbors of each value on one-dimensional vibration signals. Two
 Journal Pre-proof

different sets of features were obtained from the proposed ternary patterns method. The most
important feature of the 1D-TP method is the precision and ease of calculation for the smallest
changes in the signals. The β threshold parameter can be used at different scales to obtain
different features from the signals [31]. This user-defined parameter takes into account
different amplitude ranges for feature extraction. Therefore, this parameter provides an
important essential advantage for capturing patterns separating the signals from each other.
The effect of the β parameter for the same sign is shown in Figure 14.

of
When the figures are taken into consideration, it is seen that different patterns are obtained
from the same signal with the β parameter. This change makes it possible to differ for both the

pro
low and up features. Therefore, low and up features can be recognized from each other. In
addition, if the β value increases, the distribution of the obtained features shows skewness to
the left. Increasing bandwidth may cause some patterns to disappear. Therefore, β should be
decided as a result of the tests according to the signals.
re-
a lP
urn
Jo
 Journal Pre-proof

of
pro
re-
a lP

Figure 14. Histograms of patterns obtained from the same raw signal for different β values
urn

4.2. Classification of Vibration Signals

The other parameter of the 1D-TP method is P, which indicates the neighbor number. 1D-TP
analysis can be performed using different P values. When generating the 1D-TP plane, the
Jo

large values of P increase both the feature histogram and also increase the cost of the
operation. Small values of P can cause significant loss of information. Appropriate P and β
values should be decided upon as a result of trials. Combinations of vibration signals obtained
in different situations were used to test the proposed method. The obtained ternary patterns
were used as features, and different machine learning methods with 10-fold cross-validation
test were used for classification. The classification success rates for the patterns obtained with
 Journal Pre-proof

P = 8 and β = 0.5xSD (standard deviation) are given in Table 1. WEKA was used as open-
source software for the classification process.

Table 4. Comparative classification success rates.

Features Dataset RF k-NN SVM BayesNet ANN

Low Features Dataset 1 91.25 84.583 78.33 89.58 76.66
Dataset 2 100 100 100 100 100

of
Dataset 3 100 100 100 100 100
Up Features Dataset 1 87.5 81.66 79.58 88.75 76.66
Dataset 2 100 100 100 100 100

pro
Dataset 3 100 100 100 100 100

Table 4 shows that high success rates are obtained. High classification success rates have been
achieved with both lower and upper pattern features. Achieving high success rates shows that
the original features characterize the type and size of the bearing fault very well.
re-
Dataset_1 and dataset_2, in other words, a 100% success rate in the classification of signals of
fault types and different fault sized signal were observed. However, lower success rates were
observed in the classification of the signals in different rpm (Dataset_3). Bearing fault
lP

analysis tests were performed for five different speed values (1500 - 1740 - 1800 - 1860 -
2100 rpm). The performance values of the original feature method were investigated by
selecting lower ranges. As the values of rpm are very close to each other (1740 - 1800 -
1860), there are deviations in the disturbance matrix, as seen in Table 5. The most significant
a

factor in the occurrence of these deviations is that the vibration signals obtained from the
experimental setup are very similar. Therefore, the highest success rate for the classification
urn

of speed signals was 91.25%. The highest success rate for different classification procedures
was observed with RF. Overall, the success rates obtained with the sub-feature groups were
more successful than the upper feature groups. The confusion matrix obtained for the
classification of different rpm signals by RF is given in Table 5.
Jo

Table 5. RF disturbance matrix of vibration signals in different rpm

Low Features Up Features

H1 H2 H3 H4 H5 H1 H2 H3 H4 H5
H1 48 48
H2 46 2 41 7
H3 6 36 6 7 28 13
H4 7 41 3 45
 Journal Pre-proof

H5 48 48

As seen in table five, all samples for H1 and H5 are categorized correctly with both lower and
upper feature groups. However, it is seen that all samples for H2, H3, and H4 are not
classified correctly. The success rate of H2, H3, and H4 signals are very close to each other
because the test conditions are very close to each other.

of
Other performance measures obtained by RF using the upper and lower feature sets obtained
from the vibration signals at different rpm speeds are shown in Table 6. When comparing the
tables, one can notice that other performance criteria are also observed as high. Hence, it can

pro
be seen that the original features developed for the study and the classification models were
very well adapted to each other.

Table 6. Performance Metrics

Features Class Recall(%) Precision Accuracy f-measure

re-
(%) (%)
H1 1,000 1,000 1,000 1,000
Low Features

H2 0,958 0,885 0,958 0,920

H3 0,750 0,800 0,750 0,774
H4 0,854 0,872 0,854 0,863
lP

H5 1,000 1,000 1,000 1,000

Mean 0,913 0,911 0,913 0,911
H1 1,000 1,000 1,000 1,000
H2 0,854 0,854 0,854 0,854
Up Features

H3 0,583 0,737 0,583 0,651

H4 0,938 0,776 0,938 0,849
a

H5 1,000 1,000 1,000 1,000

Mean 0,875 0,873 0,875 0,871
urn

To test the effectiveness of the proposed 1D-TP feature extraction method for fearing fault
classification, it is compared with the feature groups obtained from the same signals both in
time and frequency domain. From the frequency and time domain, the statistical features
Jo

(min, max, average, median, energy, kurtosis, distortion, entropy, correlation, and variation
coefficient) were obtained. Classification by RF was performed using these features. The
success rates are shown in Table 7.

Table 7. Identification of Time Domain and Frequency Domain

Datasets 1D-TP Time Domain Frequency Domain
 Journal Pre-proof

Dataset 1 91.25 86.66 87.08

Dataset 2 100 97.22 98.61
Dataset 3 100 99.30 100

From Table 7, it is observed that the features obtained by the proposed feature extraction
methods have higher success than those obtained from the time and frequency domain. It has

of
also been observed that the features obtained from the signals in the frequency domain are
more successful than the same features obtained from the same signals in the time domain in
fearing fault classification.

pro
The most important advantages of the proposed method are that it is fast and real-time
applications. The feature extraction and classification process were performed on a computer
with Intel Core i5 processor, 8 GB RAM and a windows operating system with Matlab
R2017b. Feature vectors were obtained from 1D-TP, time, and frequency domains. Table 8
re-
shows the time is taken (sec.) for the creation of feature vectors.

Table 8. Time complexity values (sec.)

Datasets 1D-TP (sec.) Time Domain (sec.) Frequency Domain (sec.)
lP

Dataset 1 0.935783 1.158888 4.389985

Dataset 2 0.352553 0.767146 0.912881
Dataset 3 0.656576 0.976005 1.293314
a

When looking at Table 8, feature vectors from bearing vibration signals were obtained faster
urn

than other methods by the 1D-TP method. It is also observed that obtaining the feature vectors
from the time domain is faster than the frequency domain.

Gaussian noises are added to the bearing vibration signals to test the proposed method against
noise. 10 dB signal-to-noise ratios (SNR) noise was added to the signals for three datasets.
Jo

Features were obtained with 1D-TP from noisy signals, and the classification process was
performed with different machine learning methods. The results obtained are given in Table 9.

Table 9. Comparative classification success rates for noisy signals.

Features Dataset RF k-NN SVM BayesNet ANN

Low Features Dataset 1 87.50 85.83 82.50 83.75 78.75
Dataset 2 97.22 97.22 97.22 95.83 95.83
 Journal Pre-proof

Dataset 3 98.95 97.91 98.95 98.95 97.91

Up Features Dataset 1 89.58 85.83 87.50 82.50 79.16
Dataset 2 97.22 97.22 97.22 97.22 95.83
Dataset 3 98.95 97.91 98.95 98.95 98.95

When Table 9 is examined, more unsuccessful results were obtained with the feature groups
formed with 1D-TP from noisy signals. However, acceptable results were observed.

of
4.3. Effect of β Parameter

In the study, different values for β were tested to show the effect of β parameter. For the

pro
dataset_1 signals, the classification was performed by using the low and up feature sets. The
success rates are given in Table 10. It can be seen from Table 10 that different patterns are
obtained with the β parameter as seen from the success rates. The value of the parameter β
should be decided upon by trials. In the study, the value of β parameter is taken as the
multiples of the standard deviation of the values on the signals. As the β value increases, the
re-
power of the ternary patterns is reduced due to the decrease in the signals above and below the
bandwidth. When β = 0.5xSD was taken for dataset_1, the high accuracy was obtained with
both low and up feature vectors. However, as the β value increases, for example, β = 3xSD,
lP

this success rate decreases.

Table 10: Success rates obtained by RF according to β parameter

Feature Vectors β=0.5xSD β=1xSD β=1.5xSD β=2xSD β=2.5xSD β=3xSD

Low Features 0,9125 87.920 86.67 87.5 85.41 82.08
a

Up Features 0,8750 87.5 87.91 83.75 82.5 84.166

urn

4.4. Comparison of the Obtained Results with Literature

In Table 11, we provide a comprehensive list of the Bearing Fault method in the literature. As
seen in the literature, success rates between 75% and 100% have been reported. A new
approach was tested for three different sets created in this study. According to the obtained
Jo

results, success rates between 91.25% and 100% were obtained in the classification of
vibration signals by the proposed feature extraction method with RF classifier.

Table 11: The Reported Studies On the Bearing Fault

Author(s) Model Dataset Fault Type Accuracy

Hoang &Kang [32] - Convolutional Neural - Case Western Reserve - Normal State - 97.74%
Network (CNN) University Bearing (NS), Inner Ring
 Journal Pre-proof

Fault Database Fault (IRF),

(CWRU) Outer Ring Fault
(ORF), and Ball
Fault (BF)

Pan et al. [33] - Symplectic Geometry - CWRU - NS, IRF ORF, - 100 %
Matrix Machine - The experimental and BF,
(SGMM) roller bearing data of - Normal, IRF - 99.83%
Hunan University ORF, and BF
Yang et al. [34] - Deep Stacking Least - CWRU - NS, ORF, IRF, - 99.90%
Squares Support Vector and BF
Machine (DSLS-SVM)
Li et al. [35] - Improved Support - CWRU - NS, ORF, IRF, - 100%

of
Vector Machine based and BF
Binary Tree (ISVM-
BT)
Saidi et al. [36] - Support Vector - CWRU - NS, ORF, IRF, - 100%
Machines (SVM) and BF

pro
Zhang et al. [37] - Deep Residual - CWRU - NS, ORF, IRF, - 99.99%
Learning (DRL) and BF
Liu et al. [38] - Gated Recurrent Unit - CWRU - NS, ORF, IRF, - 99.75%
Based Non-Linear and BF
Predictive
Autoencoders (GRU-
NP-DAEs)
Zhao et al. [39] - Synchro squeezing - Experimental setup of - NS, IR, OR - 98.3%
Transform (SST) and authors
re-
Deep Convolutional
Neural Network
(DCNN)

Kavathekar et al. [40] - Rotation Forest - CWRU - NS, ORF, IRF, - 75%
and BF
lP

Seera et al. [41] - Fuzzy Min-Max - CWRU - NS, ORF, IRF, - 99.9%
Neural Network - - Experimental setup of and BF - 99.8%
Random Forest FMM- authors
RF
Zarei et al. [42] - Removing the non- - Experimental setup of - NS, ORF, IRF - 100%
bearing fault authors
component (RNFC)
filter based on artificial
a

neural networks
(ANNs)
Yang et al. [43] - Deep Neural Network - CWRU - NS, ORF, IRF, - 100%
urn

(DNN) and BF
Ahmed et al. [44] - compressive - CWRU - NS, ORF, IRF, - 99.9%
Sampling- Laplacian and BF
Score Logistic
Regression Classifier
(CS-LS-LRC)
He & He [45] - Discrete Fourier - Experimental setup of - NS, ORF, IRF, - 99.92%
Transform -Inverse authors BF, Cage
Discrete Fourier
Transform Autoencoder
Jo

(DFT-IDFT
autoencoders)
Ding et al. [46] - Local and Global - CWRU - NS, ORF, IRF, - 99.6%
Principal Component BF and fault size
Analysis (LGPCA) - (10 class)
based fusion feature
extraction methods by
the nearest mean
classifier
Authors of this article - One Dimensional - Experimental setup of - NS, ORF, IRF, - 100%
Ternary Pattern (1D- authors BF,
TP) with RF (Random
 Journal Pre-proof

Forest), k-NN (k-

nearest neighbor), SVM - Fault size, - 100%
(Support Vector
Machine), BayesNet,
ANN (Artificial Neural - Motor speed - 91.25%
Networks)

5. DISCUSSION

In today's manufacturing industry, machine automation systems are widely used. When we

of
examine these systems, machine movement is provided through bearings. The bearings are
used as a transition element between rotary and fixed parts in machines. Even if the costs of
the bearings are too low compared to the machines, the malfunctions on the bearing can cause

pro
malfunctions in the whole mechanism and disruptions in the system.

When the literature is examined, many research studies are using different feature extraction
schemes and benchmark datasets for fault classification techniques. In this study, the models
re-
are generated in order to predict the fault type, size, and the rotational speed of the motor. For
this purpose, firstly, bearing test setup has been established. Afterward, sensitive artificial
errors of specific sizes were created by a laser beam to the bearings to be used in the
experimental setup. Artificial faults created by the laser beam were examined under a
lP

microscope, and their sensitivities were examined. Since detection and diagnosis of bearing
failures is a complex process, a new feature extraction method is proposed for bearing
vibration signals. 1D-TP is a statistical approach that uses binary comparisons with neighbors
of each value on signals. Therefore, this method is efficient and robust feature extraction
a

procedure. With 1D-TP, two different sets of feature vectors are obtained from each signal.
Since these features are the best representations of vibration signals, it has been observed that
urn

performance rates are increased. These datasets were used for different machine learning
methods. According to the obtained results, 91.25% for dataset1 (different speed), 100% for
dataset2 (fault type - inner ring, outer ring, ball) and 100% for dataset3 (fault size (mm))
success rate were obtained in the classification of vibration signals by the proposed method
Jo

with RF classifier. The most important advantages of the proposed method are that it is fast
and real-time applicable.

Besides, the proposed 1D-TP feature extraction method is compared with the feature groups
obtained from the same signals both in time and frequency domain in terms of time-
complexity and accuracy results. It is observed that the features obtained by the proposed
 Journal Pre-proof

feature extraction methods have higher success rates and better time complexity than those
obtained from the time and frequency domain.

REFERENCES

1) Yu, G. (2015). Fault feature extraction using independent component analysis with
reference and its application on fault diagnosis of rotating machinery. Neural computing and
applications, 26(1), 187-198.

of
2) Huang, W., Sun, H., Luo, J., & Wang, W. (2019). Periodic feature-oriented adapted
dictionary free OMP for rolling element bearing incipient fault diagnosis. Mechanical

pro
Systems and Signal Processing, 126, 137-160.
3) Zhang, Z., Li, S., Wang, J., Xin, Y., & An, Z. (2019). General normalized sparse
filtering: A novel unsupervised learning method for rotating machinery fault
diagnosis. Mechanical Systems and Signal Processing, 124, 596-612.
4) Sun, R. B., Yang, Z. B., Zhai, Z., & Chen, X. F. (2019). Sparse representation based
re-
on parametric impulsive dictionary design for bearing fault diagnosis. Mechanical Systems
and Signal Processing, 122, 737-753.
5) Wang, Y., Peter, W. T., Tang, B., Qin, Y., Deng, L., Huang, T., & Xu, G. (2019).
lP

Order spectrogram visualization for rolling bearing fault detection under-speed variation
conditions. Mechanical Systems and Signal Processing, 122, 580-596.
6) Zhao, D., Wang, T., & Chu, F. (2019). Deep convolutional neural network-based
planet bearing fault classification. Computers in Industry, 107, 59-66.
a

7) Cheng, Y., Wang, Z., Zhang, W., & Huang, G. (2019). Particle swarm optimization
algorithm to solve the deconvolution problem for rolling element bearing fault diagnosis. ISA
urn

Transactions.
8) Karaçay, T., & Aktürk, N. (2009). Modeling of Vibrations Caused by Localized
Defects in Ball Bearings. Journal of the Faculty of Engineering and Architecture of Gazi
University, 24(2).
Jo

9) Arslan, H., Aslan, E., & Aktürk, N. (2006). Investigation of Ball Bearing Defects by
Vibration Analysis Method. Journal of the Faculty of Engineering and Architecture of Gazi
University, 21(3).
10) Tahir, M. M., Khan, A. Q., Iqbal, N., Hussain, A., & Badshah, S. (2017). Enhancing
fault classification accuracy of ball bearing using central tendency based time-domain
features. IEEE Access, 5, 72-83.
 Journal Pre-proof

11) Wang, Z., Zhang, Q., Xiong, J., Xiao, M., Sun, G., & He, J. (2017). Fault diagnosis of
a rolling bearing using wavelet packet denoising and random forests. IEEE Sensors Journal,
17(17), 5581-5588.
12) Ding, X., He, Q., & Luo, N. (2015). A fusion feature and its improvement based on
locality preserving projections for rolling element bearing fault classification. Journal of
Sound and Vibration, 335, 367-383.
13) Yan, X., & Jia, M. (2018). A novel optimized SVM classification algorithm with

of
multi-domain feature and its application to fault diagnosis of rolling
bearing. Neurocomputing, 313, 47-64.

pro
14) Ertunc, H. M., Ocak, H., & Aliustaoglu, C. (2013). ANN-and ANFIS-based multi-
staged decision algorithm for the detection and diagnosis of bearing faults. Neural Computing
and Applications, 22(1), 435-446.
15) Kavathekar, S., Upadhyay, N., & Kankar, P. K. (2016). Fault classification of ball
bearing by rotation forest technique. Procedia Technology, 23, 187-192.
re-
16) Zhang, Y., Duan, L., & Duan, M. (2019). A new feature extraction approach using an
improved symbolic aggregate approximation for machinery intelligent diagnosis.
Measurement, 133, 468-478.
lP

17) Attoui, I., Fergani, N., Boutasseta, N., Oudjani, B., & Deliou, A. (2017). A new time-
frequency method for identification and classification of ball bearing faults. Journal of Sound
and Vibration, 397, 241-265.
18) Singh, M., & Shaik, A. G. (2019). Faulty bearing detection, classification, and location
a

in a three-phase induction motor based on Stockwell transform and support vector machine.
Measurement, 131, 524-533.
urn

19) Jiang, L., Xuan, J., & Shi, T. (2013). Feature extraction based on semi-supervised
kernel Marginal Fisher analysis and its application in bearing fault diagnosis. Mechanical
Systems and Signal Processing, 41(1-2), pp. 113-126.
20) Bayram, S., Kaplan Kaplan, M. K., & Ertunç, H. M. (2013). Bilyeli Rulmanlarda
Zaman Uzayında İstatistiksel Öznitelik Çıkarımı ve Yapay Sinir Ağları Metodu ile Hata
Jo

Boyutunun Kestirimi, Otomatik Kontrol Ulusal Toplantısı, TOK2013, 26-28 Eylül 2013,
Malatya, pp. 986-991.
21) Kaplan, K., Kuncan, M., & Ertunc, H. M. (2015, May). Prediction of bearing fault size
by using the model of adaptive neuro-fuzzy inference system, IEEE 23nd Signal Processing
and Communications Applications Conference (SIU), pp. 1925-1928.
 Journal Pre-proof

22) Kaplan, K., Kuncan, M., & Ertunc, H. M. Classification of Bearing Fault Size by
Using Support Vector Machines, International Conference on Advances and Innovations in
Engineering (ICAIE)
23) Kaplan, K., Bayram, S., Kuncan, M., & Ertunç, H. M. (2014). Feature Extraction of
Ball Bearings in Time-Space and Estimation of Fault Size with Method of ANN, Proceedings
of the 16th Mechatronica, 2014.
24) Bayram, S., Kaplan, K., Kuncan, M., & Ertunç, H. M., The Effect Of Bearings Faults

of
To Coefficients Obtained By Using Wavelet Transform, IEEE 22nd Signal Processing and
Communications Applications Conference (SIU 2014).

pro
25) Ertunç, H. M. (2018). A Combined Decision Algorithm For Diagnosing Bearing
Faults Using Artificial Intelligent Techniques. Sigma: Journal of Engineering & Natural
Sciences/Mühendislik ve Fen Bilimleri Dergisi, 36(4).
26) Kaya, Y. (2015). Hidden pattern discovery on epileptic EEG with 1-D local binary
patterns and epileptic seizures detection by grey relational analysis. Australasian Physical &
re-
Engineering Sciences in Medicine, 38(3), 435-446.
27) Kaya, Y., Uyar, M., Tekin, R., & Yıldırım, S. (2014). 1D-local binary pattern-based
feature extraction for the classification of epileptic EEG signals. Applied Mathematics and
lP

Computation, 243, 209-219.

28) Kuncan, F., Kaya, Y., Kuncan, M. (2019). A Novel Approach for Activity
Recognition with Down-Sampling 1D Local Binary Pattern Features. Advances in Electrical
and Computer Engineering, 19(1), 35-44.
Kaya, Y., & Ertuğrul, Ö. F. (2016). A novel feature extraction approach for text-based
a

29)
language identification: Binary patterns — Journal of the Faculty of Engineering and
urn

Architecture of Gazi University, 31(4).

30) Kuncan, F., Kaya, Y., Kuncan, M., (2019). New approaches based on local binary
patterns for gender identification from sensor signals. Journal of the Faculty of Engineering
and Architecture of Gazi University, 34(4).
31) Kaya, Y., & Ertuğrul, Ö. F. (2016). A novel feature extraction approach in SMS spam
Jo

filtering for mobile communication: one‐dimensional ternary patterns. Security and

Communication Networks, 9(17), 4680-4690.
32) Hoang, D. T., & Kang, H. J. (2019). Rolling element bearing fault diagnosis using
convolutional neural network and vibration image. Cognitive Systems Research, 53, 42-50.
 Journal Pre-proof

33) Pan, H., Yang, Y., Zheng, J., Li, X., & Cheng, J. (2019). A fault diagnosis approach
for roller bearing based on symplectic geometry matrix machine. Mechanism and Machine
Theory, 140, 31-43.
34) Li, X., Yang, Y., Pan, H., Cheng, J., & Cheng, J. (2019). A novel deep stacking least
squares support vector machine for rolling bearing fault diagnosis. Computers in
Industry, 110, 36-47.
35) Li, Y., Xu, M., Wei, Y., & Huang, W. (2016). A new rolling bearing fault diagnosis

of
method based on multiscale permutation entropy and improved support vector machine-based
binary tree. Measurement, 77, 80-94.

pro
36) Saidi, L., Ali, J. B., & Fnaiech, F. (2015). Application of higher-order spectral features
and support vector machines for bearing faults classification. ISA Transactions, 54, 193-206.
37) Zhang, W., Li, X., & Ding, Q. (2018). Deep residual learning-based fault diagnosis
method for rotating machinery. ISA Transactions.
38) Liu, H., Zhou, J., Zheng, Y., Jiang, W., & Zhang, Y. (2018). Fault diagnosis of rolling
re-
bearings with recurrent neural network-based autoencoders. ISA Transactions, 77, 167-178.
39) Zhao, D., Wang, T., & Chu, F. (2019). Deep convolutional neural network-based
planet bearing fault classification. Computers in Industry, 107, 59-66.
lP

40) Kavathekar, S., Upadhyay, N., & Kankar, P. K. (2016). Fault classification of ball
bearing by rotation forest technique. Procedia Technology, 23, 187-192.
41) Seera, M., Wong, M. D., & Nandi, A. K. (2017). Classification of ball bearing faults
using a hybrid intelligent model. Applied Soft Computing, 57, 427-435.
a

42) Zarei, J., Tajeddini, M. A., & Karimi, H. R. (2014). Vibration analysis for bearing
fault detection and classification using an intelligent filter. Mechatronics, 24(2), 151-157.
urn

43) Yang, Y., Fu, P., & He, Y. (2018). Bearing fault automatic classification based on
deep learning. IEEE Access, 6, 71540-71554.
44) Ahmed, H., & Nandi, A. K. (2018). Compressive sampling and feature ranking
framework for bearing fault classification with vibration signals. IEEE Access, 6, 44731-
44746.
Jo

45) He, M., & He, D. (2019). A new hybrid deep signal processing approach for bearing
fault diagnosis using vibration signals. Neurocomputing.
46) Ding, X., He, Q., & Luo, N. (2015). A fusion feature and its improvement based on
locality preserving projections for rolling element bearing fault classification. Journal of
Sound and Vibration, 335, 367-383.
*Conflict of Interest

Journal Pre-proof

There is no ‘Conflict of Interest’ in the publication of the manuscript “A Novel

Feature Extraction Method for Bearing Fault Classification with One
Dimensional Ternary Patterns”.

of
pro
re-
a lP
urn
Jo