1558-1748 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2018.2866708, IEEE
Sensors Journal
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
3
input vectors to ELM, as well as the sensitive dimensionless
parameters based on WPD and KPCA with ANNs and SVMs.
The experimental results indicate that our proposed method can
effectively improve the accuracy and quickly diagnose the
faults. The flow chart of the fault diagnosis method based on
WELM with WPD and KPCA is shown in Fig. 1.
.
Fig. 1. Flow chart of fault diagnosis for rotating machinery
The proposed method can be implemented according to the
following steps:
Step 1. The vibration acceleration signal of each fault type from
the sensor of the rotating machinery is acquired using the data
collection system.
Step 2. Signal preprocessing. The original signals are
decomposed into three layers using a db10 wavelet;
time-domain signals of 8 frequency components from low
frequency to high frequency are extracted.
Step 3. Feature extraction. Five dimensionless parameters as
fault features, namely waveform indicator, impulse indicator,
margin indicator, peak indicator, and kurtosis indicator, are
calculated for the original vibration signal including
time-domain signals of 8 frequency components respectively.
For convenience,
S
,
,
CL
,
C
, and
are used to represent
the waveform indicator, impulse indicator, margin indicator,
peak indicator, and kurtosis indicator respectively. Thus, 45
dimensionless parameters as a high-dimensionality feature set
are obtained, i.e.,
i
S
,
i
,
i
CL
,
i
C
, and
i
(
0,1, 2,....8i
),
where
i
represents the
ith
original vibration signal and
time-domain signals of 8 frequency components from the low
frequency to the high frequency.
Step 4. Feature evaluation and dimensionality reduction. The
KPCA is employed to evaluate the 45 dimensionless
parameters by feature contribution rate. The contribution rate of
each dimensionless parameter is sorted from high to low. The
dimension of excellent dimensionless parameters can be
determined preliminarily according to the principle that the
cumulative contribution rate of principal components is not less
than 95%. The rest of the dimensionless parameters are
eliminated.
Step 5. Training the WELM model. The samples of the
sensitive features are divided into training samples and testing
samples; training samples are used as inputs of WELM to train
the networks. The number of nodes in hidden layers is set as 30.
Through training the networks, the final input features can be
determined according to the classification accuracy of WELM.
When the classification accuracy reaches the maximum values,
the corresponding input features as the WELM input features
are the most sensitive ones. They are adopted to identify fault
patterns of the rotating machinery. We finally obtain the best
model of fault classification.
Step 6. Testing the WELM model. The accuracy and speed
of the WELM model are calculated using the test samples.
III. R
ELATED THEORIES
A. Wavelet packet decomposition
WPD is a time–frequency analysis method, which has the
advantage of being able to provide more accurate
decomposition in the high frequency part of the signal without
redundancy or omissions. Due to the noise in the industrial
environment and the non-stationary mechanical vibration
signal with high frequency characteristic, WPD can deal with
the high-frequency part with thresholds using a variable
spectral window. Thus, WPD can perform a better time-domain
analysis of the signals, which can extract the fault feature well
and improve the accuracy of diagnosis. A more detailed and
comprehensive description concerning WPD can be found in
[4].
B. Dimensionless parameters
Vibration signals in the time domain are beneficial to keep the
basic feature of signals. In practice, dimensional parameters vary
in different work conditions, and they are easily influenced by
many disturbances (e.g., speed, load, and equipment sensitivity)
that cause data deviation. Dimensionless parameters are sensitive
to faults instead of work conditions; thus, they are employed
widely in more complex industrial conditions [32-33]. A
dimensionless parameter is defined as follows:
(1)
where
denotes the vibration amplitude, and ()px the
probability of a density function of vibration amplitude.
1
1
()
=
()
l
l
x
m
m
xpxdx
xpxdx