Professional Documents
Culture Documents
PR 4
PR 4
Hyun Cheol Cho1, Kwang Su Kim2, Chang Hwan Song3, Young Jin Lee4, and Kwon Soon Lee5
Dept. of Electrical Engineering, Dong-A Univ., 840 Hadan 2-dong, Saha-gu, Busan, S. Korea 604-714
1
hyunccho@gmail.com, 2 kim-kwang-su-2@hanmail.net, 3 hwanny77@hotmail.com
4
airlee011@hanmail.net, 5 kslee@dau.ac.kr
Abstract: This paper presents stochastic fault detection algorithm for induction motor systems. We measure the current
of a healthy induction motor by means of a Hall sensor and then determine its probability distribution. We propose a
recursive probability density estimation algorithm suitable for real-time experimental implementation due to its
simplicity and low computational load. We apply our fault detection approach to three-phase induction motors and
obtain real-time experimental results that demonstrate its reliability and practicability.
Keywords: fault detection/diagnosis, induction motor, probabilistic model, probability density estimation
- 1380 -
PR0001/08/0000-1380 ¥400 © 2008 SICE
networks [20], self organizing maps [21], and modified probability density estimation of the system state
back-propagation neural model [22]. sequence. Fig. 2 illustrates the concept of this fault
Since FDD is based on the predicted behavior of the detection configuration. Here, a probabilistic model
system based on its mathematical model, the model p(x*) constructed from healthy system is compared to
must accurately represent the behavior of the system. In that of an actual system p(x) to generate the residual
most model based FDD applications, the measurements signal ]p similarly. We modify the hypothesis testing in
are treated as deterministic signals. In practice, (1) to accommodate this stochastic fault detection
measurements typically include additive noise, which is scheme as
non-Gaussian and nonstationary for nonlinear ° H 0 : no fault , if ] p rp
time-varying systems. Hence, deterministic FDD ® (2)
methods are ineffective for such applications. °̄ H1 : fault , otherwise
This paper presents a stochastic FDD methodology where rp (0,1) and ]p = p(x*) p(x). To apply the
for induction motor systems. We construct a stochastic hypothesis testing of (2) we must first determine the
model using the motor current measured by means of a probability density model from system state sequence
Hall sensor. This model is used to generate residue by by means of online estimation.
comparing to probability distribution of an actual
system and decision making for FDD is carried out x* (t )
through binary hypothesis testing. We propose a System model
recursive probability density estimation algorithm
input
suitable for real-time implementation due to its low ] (t )
computational load. In addition, we prove the
convergence and stability of the proposed estimation Actual system
x (t )
algorithm. We apply our FDD method to three-phase
induction motors and test its reliability and
practicability through our real-time experiment. Fig. 1 A block diagram of model based fault detection
This paper is organized as following: We provide a method.
review of the literature on FDD methods for induction
p ( x* )
motor systems in Section 2. We present probabilistic
Probabilistic model
model based FDD algorithm and recursive probability
density estimation in Section 3. Real-time experiment
]p
and its result for induction motor FDD are presented in
Section 4. Conclusions and future work are given in input Probability density
Actual system
Section 5. x(t ) estimation p (x)
2. PROBABILISTIC MODEL BASED FDD Fig. 2 Stochastic model based fault detection scheme.
Fig. 1 shows a block diagram of a typical model 3. RECURSIVE PROBABILITY DENSITY
based fault detection scheme. Here, the modeling
ESTIMATION
deviation ] is obtained from the difference between the
model state x*(t) and the actual system state x(t), i.e. ] = Probability density estimation is an important signal
x* x. A simple decision making rule for fault detection processing problem [23]. Although several estimation
is the binary hypothesis testing as algorithms have been proposed, the most popular
° H 0 : no fault , if ] r algorithm is the relative-frequency approach because it
® (1) can be easily implemented. In this scheme, the
°̄ H 1 : fault , otherwise
frequency of occurrence for any given state is assigned
where r > 0. In (1), hypothesis H0 involves healthy a probability which is then estimated from experimental
system state when the residual is less than a predefined data. Consider a discrete state vector x(k) given as a set
threshold r. By contrast, hypothesis H1 is regarded that with n elements as
possible fault is occurred in systems when it is greater x(k ) ^x1 , x2 , , xn ` (3)
than or equal to threshold.
and its probability distribution is expressed by
In general, practical systems include inherently
random signals due to noise in real-time p( x(k )) ^p( x1 ), p ( x2 ),, p ( xn )` (4)
implementations. It is often reported that a deterministic where according to probability axiom this random
fault detection scheme is unsatisfactory or unreliable in vector is constraint as
practical applications. Moreover, since the complexity n
- 1381 -
initial and current time points respectively. Hence, we The first is a healthy motor while the other two motors
have the probability estimate are faulty in the stator winding and the bearing
N ( xi ) respectively. The stator winding fault is an electric
p( xi ) , i [1, n] (6) fault and the bearing fault is a mechanical fault.
k k0
Prior to estimating the probability distribution of the
where N(xi) is number of times xi occurs in [kk0] and k0 state, we first obtain the motor current sequence through
< k. The computation of (6) is quite simple. However, a digital data acquisition system (0.01 sec sampling
large quantities of data are required to obtain a reliable time) during 60 sec. Fig. 4 shows plots of the current
estimate. In addition, the computation of p(xi) becomes measurements for each motor where we observe that the
unreliable as the time interval increases and overflow healthy motor current is almost sinusoidal while the
errors may result. Thus, computational errors eventually others are not. Next, we discretize the continuous
make the fault detection becomes unreliable. To current signal as
overcome this problem we modify the estimation x(k ) ^x1 , x2 , , x8 ` (10)
algorithm to a recursive framework. First, we define any
probability as with
pi (k ) Umi (k ), i [1, n] (7) x1 ^x | x ! 3` , x2 ^x | 2 x d 3`
where pi(k) { p(xi) for simplicity, mi(k) is likelihood of x3 ^x | 1 x d 2` , x4 ^x | 0 x d 1`
xi(k) and U is given as a normalizing coefficient to x5 ^x | 1 x d 0` , x6 ^x | 2 x d 1`
satisfy the constraint in (6). A likelihood variable mi(k) x7 ^x | 3 x d 2` , x8 ^x | x d 3`
in (7) is dependent on the current observation, expressed First, we utilize our probability density estimator to
in recursive form as construct a stochastic model for the current sequence of
§ k 1 · 1 the healthy motor in Fig. 4(a). Fig. 5 shows the
mi (k ) ¨ ¸mi (k ) ] i (k ) (8)
© k ¹ k probability distributions for each phase of the motor.
where ]i(k) is an optional parameter based on This result demonstrates that the distributions are almost
observation sequence. Explicitly, ]i(k) is selected as a identical since each current waveform is the same as
non-negative constant c if the current state is observed shown in Fig. 4(a). Next, Figs. 6 and 7 are plots of the
histories of the residual for each phase of the two faulty
as xi, otherwise ]i(k) is zero. This rule is expressed
motors. In Fig. 6 a large transient error is followed by
mathematically as
convergence with steady-state error of about 0.72, 0.58,
c, if x(k ) xi and 0.18% for phases A, B, and C, respectively.
] i (k ) ® (9)
¯0, otherwise Similarly, in Fig. 7 the residual signals have about 0.87,
In summary, the likelihood in (8) is updated by the 0.81, and 0.78% in the steady-state for phases A, B, and
previous likelihood and the current observation, and C, respectively. The latter is larger on the average than
then normalized to calculate the probability distribution. the former, which indicates that the bearing fault of the
Obviously, this algorithm is easily implemented and its motor is more easily detected than the stator winding
reduced computational load makes is suitable for fault.
on-line computation.
5. CONCLUSIONS
4. REAL-TIME EXPERIMENT OF THE
This paper presents a FDD methodology for
INDUCTION MOTOR FDD
induction motor systems based on online probability
density estimation. To construct a probabilistic model of
We apply the proposed FDD algorithm to three-phase
the healthy motor, we use a Hall sensor to measure the
induction motor systems to demonstrate its reliability.
current and estimate its probability distributions using
The specifications of the applied motor are as follows:
our proposed estimation algorithm. The estimation
rated power is 10 hp, rated stator voltage is 220 voltages
algorithm is recursive making it suitable for real-time
per phase, rated stator current is 32 Ampere per phase,
implementation. The current signal of the faulty motor
rate frequency is 60 Hz, and rate speed is 1760 rpm.
is propagated to a PC where our FDD algorithm is
Fig. 3 shows the FDD test-bed for induction motor
programmed in Matlab© code, and then its discrete
systems. The induction motor is operated and controlled
probability distribution is estimated. By comparing the
by inverter systems, and each phase current of the motor
actual distribution to the stochastic model of the healthy
is measured by means of a Hall sensor. This signal is
motor, we experimentally obtain the residual signal in
propagated to a PC through an interface board. The
real time. Future work includes a microprocessor based
proposed algorithm is programmed with Matlab©
FDD test-bed of this mechanism for practical
software on a PC where the actual probability
implementation in industrial application.
distribution for the state signal is calculated online and
compared to a pre-selected probability distribution to
create the residual signal.
We use three induction motors with the same
specifications under identical experimental conditions.
- 1382 -
2
phase a
1.8 phase b
phase c
1.6
1.4
Residual signal
1.2
0.8
0.6
0.4
0.2
0
0 10 20 30 40 50 60
time(sec)
2
phase a
phase b
Fig. 3 Appearance of the test-bed for induction motor
1.8
phase c
1.4
(a)
Residual signal
1.2
4
2 phase a
current
phase b 1
0
phase c
-2 0.8
-4
0 10 20 30 40 50 60
0.6
time(sec)
(b)
5 0.4
phase a
current
0
phase b 0.2
phase c
0
-5 0 10 20 30 40 50 60
0 10 20 30 40 50 60
time(sec)
time(sec)
(c)
4
phase a
Fig. 7 The residual signal for faulty bearing.
2
current
phase b
0
phase c
-2
-4
0 10 20 30 40 50 60
ACKNOWLEDGEMENT
time(sec)
(a) Healthy motor (b) Faulty stator (c) Faulty bearing This research was supported by KEPCO grant
Fig. 4 Current waveform of the induction motors. 2006-0273.
REFERENCES
WUZ
[1] S. Nandi, H. A. Toliyat, and X. Li, 䙼Condition
OXP
WUY\ monitoring and fault diagnosis of electrical motors
OYP
WUY - A review,䙽 IEEE Trans. on energy Conversion,
probability
OZP
vol. 20, no. 4, pp. 719-729, 2005.
O[P
WUX\ [2] W. Li and C. K. Mechefske, 䙼 Detection of
O\P
WUX induction motor faults: A comparison of stator
O]P
current, vibration and acoustic methods, 䙽 J. of
WUW\ O^P
Vibration & Control, vol. 12, no. 2, pp. 165-188,
O_P
W 2006.
wGh wGi wGj
[3] M. Benbouzid, H. Nejjari, R. Beguenane, and M.
Vieira, 䙼 Induction motor asymmetrical faults
Fig. 5 Probability density estimation of the healthy detection using advanced signal processing
motor. techniques,䙽 IEEE Trans. on Energy Conversion,
vol. 14, no. 2, pp. 147-152, 1999.
[4] J.-H. Jung, J.-J. Lee, and B.-H. Kwon, 䙼Online
- 1383 -
diagnosis of induction motors using MCSA, 䙽 1112-1120, 2008.
IEEE Trans. on Industrial Electronics, vol. 53, no. [19] H. Su and K. T. Chong, 䙼 Induction machine
6, pp. 1842-1852, 2006. condition monitoring using neural network
[5] A. Boggess and F. J. Narcowich, A first course in modeling,䙽 IEEE Trans. on Industrial Electronics,
wavelets with Fourier analysis, Prentice Hall, vol. 54, no. 1, pp. 241-249, 2007.
Upper Saddle River, 2001. [20] L. Eren, A. Karahoca, and M. J. Devaney, 䙼Neural
[6] B. Boashash, Time frequency analysis, Elsevier network based motor bearing fault detection, 䙽
Science, 2003. Proc. of IEEE Instrumentation & Measurement
[7] C. Nikias and A. P. Petropuou, Higher-order Tech. Conf., Como, Italy, pp. 1657-1660, 2004.
spectra analysis, Prentice Hall, Upper Saddle
[21] N. Li and C. K. Mechefske, 䙼 Induction motor
River, 1993.
fault detection and diagnosis using artificial neural
[8] S.-H. Lee, S.-P. Cheon, Y. Kim, and S. Kim,
networks,䙽 Int. J of COMADEM, vol. 9, no. 3, pp.
䙼Fourier and wavelet transformations for the fault
15-23, 2006.
detection of induction motor with stator current,䙽 [22] V. Duraisamy, N. Devarajan, D.
Int. Conf. on Intelligent Computing, Kunming, Somesundareswary, and S. N. Sivanandam,
China, pp. 557-569, 2006.
䙼 Modified back propagation algorithm for
[9] G. A. Jimenez, A. O. Munoz, and M. A.
incipient fault detection in three-phase induction
Duarte-Mermoud, 䙼 Fault detection in induction
motor, 䙽 Modelling, Measurement & Control A,
motors using Hilbert and wavelet transforms, 䙽 vol.79, no.1-2, pp. 15-25, 2006.
Electrical Engineering, vol. 89, pp. 205-220, 2007. [23] R. O. Duda, P. E. Hart, and D. G. Stork, Pattern
[10] M. Blodt, D. Bonacci, J. Regnier, M. Chabert, and classification, Wiley-Interscience, 2000.
J. Faucher, 䙼 On-line monitoring of mechanical
faults in variable-0speed induction motor drives
using the Wigner distribution,䙽 IEEE Trans. on
Industrial Electronics, vol. 55, no. 2, pp. 522-533,
2008.
[11] N. Arthur and J. Penman, 䙼 Induction machine
condition monitoring with higher order spectra,䙽
IEEE Trans. on Industrial Electronics, vol. 47, no.
5, pp. 1031-1041, 2000.
[12] C. Combastel, S. Lesecq, S. Petropol, and S. Gentil,
䙼Model-based and wavelet approaches to induction
motor on-line fault detection, 䙽 Control
Engineering Practice, vol.10, pp.493-509, 2002.
[13] S. Haykin, Neural networks - A comprehensive
foundation, Prentice Hall, Upper Saddle River,
1998.
[14] L.-X. Wang, A course in fuzzy systems and control,
Prentice Hall, Upper Saddle River, 1996.
[15] D. F. Leite, M. B. Hell, P. H. Diez, B. S. Ganglio,
L. O. Nascimento, and P. Costa Jr., 䙼Real-time
model-based fault detection and diagnosis for
alternators and induction motors,䙽 Proc. of IEEE
Int. Electric Machines & Drives Conf., Antalya,
Turkey, pp. 202-207, 2007.
[16] P. V. J. Rodriguez and A. Arkkio, 䙼 Induction
motor stator fault detection using fuzzy logic, 䙽
Proc. of the IASTED Int. Conf. on Circuits, Signals,
& Systems, Cancun, Mexico, pp. 96-100, 2003.
[17] F. Zidani, D. Diallo, M. E. H. Benbouzid, and R.
Nait-Said, 䙼 A fuzzy based approach for the
diagnosis of fault modes in a voltage-fed PWM
inverter induction motor drive,䙽 IEEE Trans. on
Industrial Electronics, vol. 55, no. 2, pp. 586-593,
2008.
[18] P. V. J. Rodriguez and A. Arkkio, 䙼Detection of
stator winding fault in induction motor using fuzzy
logic, 䙽 Applied Soft Computing, vol. 8, pp.
- 1384 -