IEEE Computational Intelligence Magazine - August 2021 - 86

View1
GP
Raw
Signal
Features
View2
Features
View3
Features
Construct Features
GP
Construct Features
GP
Construct Features
KNN
KNN
KNN
FIGURE 5 Architecture of ensemble fault diagnosis system.
FIGURE 6 NCEPU test rig.
construct features from different views.
Since the high-level features are constructed
from different views, they are
diverse and effective. To effectively use
the constructed features, an ensemble is
created for fault diagnosis. Consistent
with the fitness evaluation process, the
ensemble uses three KNNs as the base
classifiers, and each KNN uses the features
constructed from a single view.
The overall fault diagnosis system using
ensemble is illustrated in Figure 5.
The training set (including the transformed
feature sets from three different
views and the corresponding class labels)
are fed into three KNNs, and these
three KNNs are used as base classifiers
of an ensemble. This ensemble will have
high generalization performance since
the inputs are the features constructed
TABLE IV Description of the NCEPU dataset.
CLASS
LABEL
1
2
3
4
5
6
RUNNING CONDITION
Normal
Inner ring fault
Outer ring fault
Rolling element fault
Inner and outer ring compound fault
Rolling element and outer ring
compound fault
TRAINING
SAMPLES
5
5
5
5
5
5
TEST
SAMPLES
45
45
45
45
45
45
from different views. This constructed
ensemble is applied to predict the class
labels for the unseen samples in the test
set. In the process, the majority voting
algorithm is employed because it is a
simple and common method to comprehensively
consider the results
obtained by the multiple base classifiers
[53]. For a test sample, among the class
labels/fault types obtained by three
KNNs, the fault type with the largest
number of votes is chosen as the final
fault type.
IV. Experimental Design
A. Dataset Description
The mechanical equipment typically
operates normally, and the failure data
that can be collected is limited. Collecting
a large number of labeled samples
requires high cost and a long period of
time. Therefore, the goal of this paper is
to deal with fault diagnosis using the
limited number of training samples. In
the experiments, three datasets with a
small number of training samples are
employed to evaluate the effectiveness
of MFCGPE. In each dataset, only five
samples of each running condition are
Majority
Voting
Final
Diagnosis
randomly selected as the training set,
and the remaining samples are used
as the test set. The three datasets are
described as follows.
1) NCEPU [25]: It is a bearing fault
dataset collected by North China
Electric Power University (NCEPU)
and contains vibration signals of six
fault types. Figure 6 shows the used
test rig of NCEPU. The six rolling
bearing running conditions are the
normal state (NOR), inner ring fault
(IRF), outer ring fault (ORF), rolling
element fault (REF), inner & outer
ring compound fault (IOCF), and
rolling element & outer ring compound
fault (ROCF), which are simulated
by manufacturing defects on
normal bearings using electrical discharge
machining (EDM) technology.
All vibration signals are collected
with a sampling frequency of 12,000
Hz. The first 102,400 data points of
vibration signals under each running
condition are divided into 50 samples
on average and there is no overlap
between each sample. Table IV
lists the detailed information of the
NCEPU dataset. Figure 7 shows the
time domain waveform of vibration
signals under six running conditions
in NCEPU.
2) CWRU [54]: It is a bearing fault
dataset collected by Case Western
Reserve University (CWRU) and
contains vibration signals of ten
fault types and fault degrees. Figure
8 shows the used test rig of CWRU.
All faults are generated by manufacturing
defects on the inner ring,
outer ring, and rolling element surfaces
using EDM, and the same fault
category with three defect diameters
(0.007, 0.014, and 0.021 inches,
respectively). The ten different rolling
bearing running conditions are
the normal state (NOR), three kinds
of inner ring fault (IRF07, IRF14,
and IRF21), three kinds of outer
ring fault (ORF07, ORF14, and
ORF21), and three kinds of rolling
element fault (REF07, REF14, and
REF21), respectively. IRF, ORF, and
REF indicate inner ring fault, outer
ring fault, and rolling element fault,
86 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | AUGUST 2021
IEEE Computational Intelligence Magazine - August 2021

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - August 2021
