Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

8th Power Electronics, Drive Systems & Technologies Conference (PEDSTC 2017)

14-16 Feb. 2017, Ferdowsi University of Mashhad, Mashhad, Iran 2017

Application of improved Hilbert-Huang and wavelet


packet transforms in broken rotor bar fault detection

Farzaneh Sabbaghian Bidgoli Javad Poshtan


Faculty of Electrical Engineering Faculty of Electrical Engineering
Iran University of Science and technology Iran University of Science and technology
Tehran, Iran Tehran, Iran
sabbaghianf@yahoo.com jposhtan@iust.ac.ir

Abstract— one of the common techniques of rotary machinery nonstationary signals is very crucial in machinery fault
fault diagnosis is the signal based fault diagnosis, in which the diagnosis[1].
signal processing is one of its integral part. Signal processing Vibration signals are one of the signals that are used in
converts the raw data into useful features to make the diagnostic signal-based fault diagnosis. These vibration signals contain
operations. These features should be independent from the normal very useful information about mechanical equipment condition,
working conditions of the machine and the external noise. They and we can reveal different kinds of faults using them. However,
should be sensitive only to faults in the machine. Therefore, the development of effective signal processing techniques that
providing more efficient processing techniques in order to achieve can extract key fault information from the vibration signals is a
more useful features of the signal and faster and more accurate
challenge[2].
fault detection have been considered by researchers. This project
applies the improved Hilbert-Huang Transform to decompose the
Among rotating machines, induction motors by high
signal into narrow frequency bands and extract instantaneous reliability and performance play an important role in many
frequency and wavelet packet transform to remove the initial industries. Despite the high reliability of these motors, they are
signal noise in vibration signal due to the broken rotor bars fault subject to some defects. Applying suitable condition monitoring
to achieve more useful features of vibration signals for the next techniques, we can detect these defects at early stages and
stages of diagnosis. Comparison of Hilbert transform amplitude prevent their development. Common faults in induction motors
spectrum and detected instantaneous frequency by the Hilbert- are rotor faults, stator faults, bearing and gear faults. This paper
Huang Transform and the improved Hilbert-Huang transform investigates rotor bar broken fault, so that vibration signal due
techniques and combining the improved Hilbert-Huang transform to this fault is modeled and then feature extraction is applied.
and wavelet packet transform indicate the superiority of the In 2010 [3] has used the Stationary Wavelet Packet
combined technique to detect frequencies of the fault. Transform (SWPT) for extracting features in Broken Rotor Bar
(BRB) diagnostics of induction motor under the condition that
Keywords— Hilbet-Huang Transform, wavelet packet the sampling rate is low. They concluded that using SWPT not
transform, Signal-based fault detection only doesn’t have a negative impact on the performance of fault
diagnosis, but also requires less cost in terms of calculation and
I. INTRODUCTION
implementation. In addition, [4] has evaluated BRB fault of the
During the operation of rotary machines, inherent dynamic electro-pump induction motor using the steady state vibration
properties and external excitations, often leads to nonstationary signal when the engine is operating at low speed. This case uses
signals, especially under varying operating conditions over time. the frequency domain features that highly depend on motor slip.
This nonstationary signals contain information about the In this method, the fault diagnosis is failed in cases where the
health of machinery. So if appropriate processing techniques motor load and the motor slip are low, in addition ignores the
applied to these signals, the important feature of fault detection non-stationary nature of the vibration signal. In the same year
and diagnosis could be extracted. [5] evaluated the BRB fault of the induction motor using
However, many signal processing techniques are for the vibration signals from MEMS sensors. Signals derived from the
stationary signals. These methods can only analyze based on rotor were evaluated at no-load and under-load states using the
statistical averages in either time or frequency domain, and vibration signals spectrum and Fast Fourier Transform analysis
cannot reveal local features in the time domain and frequency and the features such as torque, speed, noise, and vibration under
domain simultaneously. On the other hand we know that the various loads were analyzed according to the motor slip changes.
statistical properties of nonstationary signals, changes with time. The previous work’s shortcomings were also observed in this
So in engineering applications, such methods are unsuitable for case. In 2014 [6] provided a time-frequency analysis of vibration
the analysis of nonstationary signals. The development of signal signals for fault diagnosis of BRB in the axial and radial
processing methods to extract the time-variant features from direction of the core. In this case, the frequency domain analysis
was used initially to separate the defective state and then, the operating mechanical system by hitting the shaft with a hammer.
wavelet transform was used for a closer investigation of the The resulting vibration was recorded by a sensor mounted on the
suspicious case. However, there was a challenge that the ridge of drive-end bearing bore. This was repeated several times,
frequency component of the BRB fault was close to the main and the recorded data were averaged to achieve the desired
frequency component of the motor and this caused difficulty in transfer function. In[9], a 2hp, three-phase induction motor and
fault diagnosis during the first phase that was done in the two accelerometers with a bandwidth of 5-5,000Hz and a lv/g
frequency domain. In the same year [7] has used the vibration output, as shown in Fig.1, was used. One is mounted on the
signals and Hilbert transform and Continuous Wavelet motor housing at the drive-end to measure the signal and the
Transform (CWT) common techniques to detect different faults other on an impact hammer to measure the impact strength.
and compared the results to find relevant fault frequencies. The
comparison results showed that although CWT gives a better
time resolution at high frequency areas, but it cannot separate
the impacts whose time interval is very low. In the field, HT is
better than CWT. On the other hand, CWT needs more
computing time and the Hilbert transform requires a pre-
processing technique such as the Empirical Mode
Decomposition (EMD). Therefore, each of them has advantages
and disadvantages than others cannot be preferred over another.
In 2016 [8] applied Discrete Wavelet Transform (DWT)
Fig. 1. A Schematic of the Experimental System [9]
technique on the transient vibration of induction motor in the
case of a broken rotor bars under different loads. The difference The estimated transfer function of the mechanical path is:
with previous works was broken bars fault diagnosis with
different intensities and in different mechanical loads using 2.283 + 642.6 +
vibrations in three directions of parallel to the axis of the rotor, ( )=
+ 619.3 + 4.42 × 10 +
perpendicular to the rotor axis, and radial direction. The 4.457 × 10
directions of parallel to the axis of the rotor achieved the best (1)
results using the Nearest Neighborhood classification technique. 4.234 × 10 s + 5.697 × 10
This reference chooses the signal related to the frequency band
This system is of fourth order and has two resonant
0 to 200Hz after applying DWT and then, performs subsequent
frequencies at 3.29 kHz and 580 Hz. A more detailed
diagnostic stages. While frequency components can have up to
explanation about this transfer function can be found in Ref [9].
seven times of multiplying engine speed by the slip frequency
according to engine speed for the broken bars fault, especially in B. Fault modeling
low-intensity and newborn faults. Therefore, using only this Vibration is the response of a mechanical system to the
frequency band will interfere the quality and accuracy of forces produced by defects. Forces that create vibration can
diagnostics. come from different sources such as defects. Defects are
In this paper, the vibration signals of the BRB fault were modeled as input forces to mechanical systems that we
modeled and the combining of the two techniques of improved introduced in Section 1-2. Fig.2 shows the vibration model that
Hilbert-Huang transform and wavelet packet transform was was used for simulation. n(t) is assumed to be white Gaussian
used. Initially, the noise removal operation was done by wavelet noise[10].
package transform on the vibration signal. Then, the De-noised
signal decomposed into Intrinsic Mode Functions (IMFs) during
applying the improved Hilbert-Huang transform, which is itself
produced by a combination of the Ensemble Empirical Mode
Decomposition and Hilbert transform, this decomposition leads
to narrow frequency bands to calculate their instantaneous
frequency more accurately. With this method, the frequency
components related to the incipient faults that cannot be detected Fig. 2. A simplified model for vibration[10]
due to their low amplitude, and frequency components due to
BRB that is lost in main component of the rotational property of according to [10] and [11] broken rotor bar produces a
the system in low load conditions, can be detected. As well as, sinusoidal force given by:
the non-stationary property of the vibration signal is included in
the analysis and more reliable and more robust results can be ( )= a cos(2 + ) (2)
achieved.
Where ai is the appropriate amplitude determined by the size
II. BROKEN ROTOR BARS FAULT MODELING of the defect, fi is the i-th frequency component due to defect,
is the appropriate phase angle and N is the number of harmonics.
A. System modeling
According to [12], frequency components due to broken rotor
To achieve vibration signal, we need a transfer function with bar are =1X and = 1 − . Also
vibration signal as output, and effects of defect as input.
According to [9], The transfer function, H(s), for the model was =( − ) × ( ) (3)
estimated as follows: an impulsive force was applied to the 60
Where P is the number of poles, is rotor speed, 1X is rotating stationary signal processing. In EEMD, the added white noise
frequency in r.p.s and is pole frequency. populates the time–frequency space uniformly, helping a natural
So input force due to broken rotor bar is modeled as separation of the frequency scales, which decreases the
occurrence of mode mixing.
= a cos(2 + )+ (2 + ) (4) The EEMD algorithm Adds a white noise ni(t) with the given
amplitude to the original signal x(t) to generate a new signal:
After all, assuming a 2-pole motor, with = 2926 and x (t) = x(t) + n (t) (5)
= 3000, frequency components are obtained as f =122 and where ni(t) denotes the i-th added white noise series, and x(t)
f =47.56 .With respect to the previous two sections and is the noise-added signal of the i-th trial, while i=1,2,…,M.
assuming sampling frequency 12 kHz, the vibration signal is Then Uses the original EMD algorithm to decompose xi(t)
obtained as shown in Fig.3 and Fig.4. into IMFs.
convolution of[(F(t)+n(t)),h(t)]#### y(t)
amplitude

50 ( )= , ( )+ , ( ) (6)
0
-50 Where S is the number of IMFs, , ( )is the final residue,
0 0.05 0.1 0.15 0.2 0.25 0.3
which is the mean trend of the signal, and Ci,s(t) represents the
time
IMFs (Ci,1, Ci,2, Ci,3,…, Cs) which include different frequency
FFT of faulty signal
bands ranging from high to low. A more detailed explanation
amplitude

2
1
about EMD can be found in Ref.[14].
0
0 1000 2000 3000 4000 5000 6000
IV. HILBERT-HUANG AND IMPROVED HILBERT-
frequency HUANGTRANSFORM
Fig. 3. Simulated signal due to broken rotor bar Hilbert transform, is defined based on the convolution of
convolution of[F(t),h(t)]#### y(t) signal x(t) with 1/t and can emphasize the local properties of x(t),
amplitude

50
as follows:
0
( )
-50 ( )= (7)
0 0.1 0.2 0.3 −
time
FFT of conv[F(t),h(t)]= y(t) Where P is the Cauchy principal value. Coupling the x(t) and
amplitude

1 y(t), we can have the analytic signal z(t) of x(t); as


0.5
( )= ( )+ ( )= ( ) ( )
0
0 2000 4000 6000
(8)
frequency Where
Fig. 4. Simulated signal due to broken rotor bar (without noise)
/ ( )
( )=[ ( ) + ( ) ] , ( ) = arctan (9)
( )
III. EMD AND ENSEMBLE EMD METHODS
The EMD method was raised by Huang et al. [13] and is able a(t) is the instantaneous amplitude of x(t); which can show
to decompose a signal into some Intrinsic Mode Functions how the energy of the x(t) changes with time, and φ( ) is the
(IMFs). An IMF represents a simple oscillatory mode imbedded instantaneous phase of x(t). An important feature of the Hilbert
in the signal. Based on this assumption that any signal consist transform is that if the signal x(t) is mono-component, then the
different simple IMFs, the EMD method was introduced to time derivative of instantaneous phase φ( ) will be the physical
decompose a signal into IMF components. Applying EMD, we
obtain a residue rI and a collection of I IMFs ci (i=1, 2,..., I). meaning of instantaneous frequency ω( ) of signal x(t), as the
Summing up all IMFs and the final residue rl, we get: following:
( )
( )=∑ + . ω( ) = (10)

The IMFs, c1, c2,..., cI, cover different frequency bands Unfortunately, in many practical applications, the signal to
ranging from high to low. The frequency components included be treated is multicomponent, and it prevents the wide
in each frequency band are different and they change with the application of instantaneous frequency concept. To apply the
variation of the signal x(t), while rI represents the central instantaneous frequency, the Hilbert Transform is applied on
tendency of the signal x(t). A more detailed explanation about IMFs resulting from EMD method. The EMD preprocessor-
EMD can be found in Ref. [13]. based Hilbert transform is called Hilbert–Huang Transform
To reduce the disadvantage of the mode mixing, Wu and (HHT)[15].
Huang proposed an effective noise-assisted method named Furthermore Improved Hilbert-Huang is based on EEMD, so
EEMD which significantly reduces the chance of inadmissible that Hilbert Transform is applied on IMFs resulting from EEMD
mode mixing and preserves the dyadic property of the to produce Hilbert-Huang spectrum.
decomposition for any data. EEMD is an improved version of
original EMD and a better tool for a non-linear and non-
V. WAVELET PACKET TRANSFORM A. Application of Hilbert-Huang Transform
WT multi-scale analysis can be useful for a signal through In TableⅠ, there are 12 instantaneous frequencies resulting
dilation and translation, so it can extract effective features in from Hilbert-Huang Transform. You can see that detected
time domain and frequency domain. Consequently, WT is frequency components are far from frequency components due
successful for the condition monitoring and fault diagnosis of to defect.
electromechanical equipment. A disadvantage of WT is that the
frequency resolution in the high-frequency sub-band where the TABLE I. INSTANTANEOUS FREQUENCY OF EACH IMF (FROM HHT)
characteristics due to fault always exist in, is weak. The Wavelet
IMF6 IMF5 IMF4 IMF3 IMF2 IMF1
Packet Transform (WPT), a development of wavelet bases, is
alternative bases formed by linear combinations of usual wavelet 143.34 326.49 603.00 1572.63 3255.79 2245.95
functions. WPT divides the frequency space into different parts IMF12 IMF11 IMF10 IMF9 IMF8 IMF7
and allows a better time–frequency localization of signals[16].
The WPT can decompose detail the information in the high- 3.12 3.33 9.67 19.17 39.54 76.36
frequency area. To perform WPT of a signal at a certain level
(e.g. level 3), the functions in Eq. (12) are unified as In Fig.6 we have Hilbert-Huang Amplitude spectrum. It
shows many frequency components that cannot be separated
( ) = √2 ∑ ℎ( ) (2 − ) from each other.
(12)
( ) = √2 ∑ ( ) (2 − ) Hilbert Amplitude Spectrum- for EMD
140

Where ( ) = ( ), and ( ) = ( ). Correspondingly, 120

the signal is decomposed as 100

Freq(hz)
80
, = ℎ( −2 ) ,
60
(13)
, = ( −2 ) ,
40

20

Where , denotes the wavelet coefficients at the j level, n


sub-band, , and , denotes the wavelet coefficients 0 1 2 3 4 5
Time(sec)
6 7 8 9

at the j+1 level, 2n and 2n+1 sub-bands, respectively, and m is Fig. 6. Hilbert-Huang Amplitude Spectrum
the number of the wavelet coefficients. As illustrated in Fig. 5, a
3-level WPT generates a total of eight sub-bands, and each sub- B. Application of Improved Hilbert-Huang Transform
band covers one eighth of the frequency information consecutive In Table Ⅱ, there are 12 instantaneous frequencies resulting
[17]. from Improved Hilbert-Huang Transform. It shows that
detection of frequency components due to defect, are done better
than previous section. (See frequency components of IMF8 and
IMF9)

TABLE II. INSTANTANEOUS FREQUENCY OF EACH IMF (FROM IHHT)

IMF6 IMF5 IMF4 IMF3 IMF2 IMF1


608.72 572.64 119.42 2245.95 3339.32 2042.27
Fig. 5. Illustration of wavelet packet transform[17]. IMF12 IMF11 IMF10 IMF9 IMF8 IMF7
13.01 2.99 3.12 116.25 49.51 23.00
VI. ANALYSIS OF SIMULATED BROKEN ROTOR BAR SIGNAL
In simulated signal, in addition to 3290 Hz and 580 Hz In Fig.7 we have Improved Hilbert-Huang Amplitude
frequency components, there are 47Hz and 122Hz frequency spectrum. It shows that decomposition of frequency components
components due to broken rotor bar. Each of applied algorithms is much better than Hilbert-Huang Transform and they are more
that can more accurately detect these components, will be more detectable.
efficient. According to Fig.4, without noise, the amplitude of
122Hz frequency component is 0.12 and of 47Hz frequency
component is 0.015, also 3300Hz has 0.1 (approximately equal
to that of 122Hz) amplitude and 580Hz has 0.5 amplitude. On
the other hand, in Fig.3 there are more powerful peaks that
detection of this two components (47Hz and 122Hz) is
impossible. So we are looking for the solution to separate the
components due fault.
Hilbert Amplitude Spectrum- for EEMD Hilbert Amplitude Spectrum- for EEMD

140 140

120 120

100 100
Freq(hz)

Freq(hz)
80 80

60
60

40
40

20
20

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35


Time(sec) 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35
Time(sec)

Fig. 7. Improved Hilbert-HuangAmplitude Spectrum Fig. 9. Improved Hilbert-Huang and Wavelet Packet Amplitude Spectrum

Closest to the 47Hz and 122Hz frequency components,


C. Application of Hilbert-Huang Transform and Wavelet which were detected by the algorithm in the previous sections,
packet transform are summarized in Table Ⅳ.
According to [18], although HHT is a powerful method to TABLE IV. DETECTED FREQUENCY COMPONENTS DUE TO FAULT
detect incipient faults, its performance is reduced, at presence of
WPT+EEMD+HT EEMD+HT EMD+HT algorithm
noise. So first we remove the noise and then apply HHT. To
remove the noise, we apply WPT, thresholding and then reverse IMF4 119.44 IMF9 116.25 IMF6 146.34 Frequency
WPT. In fact, we remove the noise by thresholding on wavelet IMF8 46.42 IMF8 49.51 IMF7 76.36 components
packet coefficients. Fig.8 shows de-noised vibration signal. In
WPT, we use dmey as mother wavelet and SURE entropy value VII. CONCLUSION
4.61 and continue decomposition to 4-th level. Among the applied algorithms in this paper, combination of
De-noised signal wavelet packet algorithm and Hilbert-Huang transform in
50 detection of frequency components to be more successful and
0
have more accuracy and less errors, afterward the algorithm
EEMD + HT gained most success. This accuracy is especially
-50
0 0.05 0.1 0.15 0.2 0.25 0.3 more important in cracked rotor bar defect that leads to weak
Fig. 8. De-noised signal frequency components. as well as according to the Hilbert
transform amplitude spectrum can be observed that the hybrid
In Table Ⅲ, there are 12 instantaneous frequencies resulting algorithm WPT + EEMD + HT is more capable in the frequency
from combination of Improved Hilbert-Huang Transform and separation and detection of frequency components that are close
wavelet packet transform. together. We should consider that the more accurate algorithm
leads to a richer input for classification and diagnosis.
TABLE III. INSTANTANEOUS FREQUENCY OF EACH IMF (FROM
HHT+WPT)
REFERENCES
IMF6 IMF5 IMF4 IMF3 IMF2 IMF1 [1] Z. Feng, M. Liang, and F. Chu, “Recent advances in time-frequency
analysis methods for machinery fault diagnosis: A review with application
179.33 396.13 639.13 1922.9 3018.9 1809.3 examples,” Mech. Syst. Signal Process., vol. 38, no. 1, pp. 165–205, 2013.
IMF12 IMF11 IMF10 IMF9 IMF8 IMF7 [2] Y. Lei, Z. He, and Y. Zi, “Application of the EEMD method to rotor fault
diagnosis of rotating machinery,” Mech. Syst. Signal Process., vol. 23, no.
3.10 5.90 9.64 19.55 46.47 119.44 4, pp. 1327–1338, 2009.
[3] H. Keskes, A. Braham, and Z. Lachiri, “Broken rotor bar diagnosis in
According to TableⅢ, frequency components due to defect induction machines through stationary wavelet packet transform and
multiclass wavelet SVM,” Electr. Power Syst. Res., vol. 97, pp. 151–157,
have been detected in IMF7 and IMF8. So we can detect low- 2013.
energy components (insipient faults). As you can see in Fig.9,
[4] Ž. Kanovi, D. Mati, Z. Jeli, M. Rapai, B. Jakovljevi, and M. Kapetina,
122Hz frequency component appears around 100Hz, and 47Hz “Induction Motor Broken Rotor Bar Detection Using Vibration Analysis
weak frequency component appears in Hilbert amplitude – A Case Study,” pp. 64–68, 2013.
spectrum with low amplitude (light blue). [5] V. H. Maruthi.G.S, “An Experimental Investigation on Broken Rotor Bar
in Three Phase Induction Motor by Vibration Signature Analysis using
MEMS Accelerometer,” IJETAE, vol. 3, no. 4, pp. 1–7, 2013.
[6] A. D. T. Rosario Miceli, Yasser Gritli, “Vibration signature analysis for
monitoring rotor broken bar in double squirrel cage induction motors
based on wavelet analysis,” COMPEL Int. J. Comput. Math. Electr.
Electron. Eng., vol. 33, no. 5, pp. 1625–1641, 2014.
[7] P. Konar and P. Chattopadhyay, “Multi-class fault diagnosis of induction
motor using Hilbert and Wavelet Transform,” Appl. Soft Comput., vol. 30,
pp. 341–352, 2015.
[8] J. Rangel-magdaleno, H. Peregrina-barreto, J. Ramirez-cortes, R.
Morales-caporal, and I. Cruz-vega, “Vibration Analysis of Partially
Damaged Rotor Bar in Induction Motor under Different Load Condition
Using DWT,” vol. 2016, 2016.
[9] X. Lou, “Fault Detection and diagnosis for rolling element bearings,”
Case Western Reserve University, 2000.
[10] H. Ocak and K. A. Loparo, “Estimation of the running speed and bearing
defect frequencies of an induction motor from vibration data,” Mech. Syst.
Signal Process., vol. 18, no. 3, pp. 515–533, 2004.
[11] K. Agoston, “Fault Detection of the Electrical Motors Based on Vibration
Analysis,” in 8th international Conference Interdisciplinarity in
Engineering, INTER-ENG 2014,9-10 october, 2015, pp. 547–553.
[12] G. K. Singh, A. Saleh, and A. Kazzaz, “Induction machine dri v e
condition monitoring and diagnostic research * a survey,” vol. 64, p.
7796, 2003.
[13] N. Huang, Z. Shen, S. Long, M. Wu, H. SHIH, Q. ZHENG, N. Yen, C.
Tung, and H. Liu, “The empirical mode decomposition and the Hilbert
spectrum for nonlinear and non-stationary time series analysis,” Proc. R.
Soc. A Math. Phys. Eng. Sci., vol. 454, no. 1971, pp. 995, 903, 1998.
[14] X. Zhang and J. Zhou, “Multi-fault diagnosis for rolling element bearings
based on ensemble empirical mode decomposition and optimized support
vector machines,” Mech. Syst. Signal Process., vol. 41, no. 1–2, pp. 127–
140, 2013.
[15] Z. K. Peng, P. W. Tse, and F. L. Chu, “A comparison study of improved
Hilbert-Huang transform and wavelet transform: Application to fault
diagnosis for rolling bearing,” Mech. Syst. Signal Process., vol. 19, no. 5,
pp. 974–988, 2005.
[16] R. Zhou, W. Bao, N. Li, X. Huang, and D. R. Yu, “Mechanical equipment
fault diagnosis based on redundant second generation wavelet packet
transform,” Digit. Signal Process. A Rev. J., vol. 20, no. 1, pp. 276–288,
2010.
[17] R. Yan, R. X. Gao, and X. Chen, “Wavelets for fault diagnosis of rotary
machines: A review with applications,” Signal Processing, vol. 96, no.
PART A, pp. 1–15, 2014.
[18] N. H. Chandra and A. S. Sekhar, “Fault detection in rotor bearing systems
using time frequency techniques,” Mech. Syst. Signal Process., vol. 72–
73, pp. 105–133, 2016.

You might also like