ISA Transactions 49 (2010) 189195

Contents lists available at ScienceDirect

ISA Transactions
journal homepage: www.elsevier.com/locate/isatrans

Detection of broken rotor bars in induction motors using nonlinear Kalman filters
Farzaneh Karami a , Javad Poshtan a, , Majid Poshtan b
a b

Department of Electrical Engineering, Iran University of Science and Technology, Tehran, Iran The Petroleum Institute, Abu Dhabi, UAE

article

info

abstract
This paper presents a model-based fault detection approach for induction motors. A new filtering technique using Unscented Kalman Filter (UKF) and Extended Kalman Filter (EKF) is utilized as a state estimation tool for on-line detection of broken bars in induction motors based on rotor parameter value estimation from stator current and voltage processing. The hypothesis on which the detection is based is that the failure events are detected by jumps in the estimated parameter values of the model. Both UKF and EKF are used to estimate the value of rotor resistance. Upon breaking a bar the estimated rotor resistance is increased instantly, thus providing two values of resistance after and before bar breakage. In order to compare the estimation performance of the EKF and UKF, both observers are designed for the same motor model and run with the same covariance matrices under the same conditions. Computer simulations are carried out for a squirrel cage induction motor. The results show the superiority of UKF over EKF in nonlinear system (such as induction motors) as it provides better estimates for rotor fault detection. 2010 Published by Elsevier Ltd on behalf of ISA.

Article history: Received 8 June 2009 Received in revised form 15 November 2009 Accepted 18 November 2009 Available online 4 March 2010 Keywords: Model-based fault detection Induction motors Nonlinear Kalman filters

1. Introduction The use of induction motors in todays industry is extensive. This requires monitoring their operation for detection of abnormal electrical and mechanical conditions that indicate or may lead to a failure of the system of which the motor is a part. An induction motor can fail due to a combination of a variety of fault mechanisms [1]. Internal motor faults (e.g. short circuit of motor leads, inter-turn short circuits, ground faults, bearing and gearbox failures, broken rotor bar and cracked rotor end rings) as well as external motor faults (e.g. phase failure, asymmetry of main supply and mechanical overload) are expected to happen sooner or later [2]. Many researchers have focused their attention on incipient fault detection and preventive maintenance. There are invasive and noninvasive methods for machine fault detection. In practice, there are a variety of measurement sensors to collect signals from various parts of on electrical machine. This sensor data can be used for failure detection and monitoring. Typical sensors might measure stator voltage and current, air-gap and magnetic flux density, rotor position and speed, and motor vibrations. It is clear that a failure monitoring system should be able to extract the signs of many possible failures by analyzing several measurements provided by many physically different sensors [35]. Ideally, monitoring and diagnostics of induction motors should use passive electronic sensors and not any thermal or mechanical

Corresponding author. E-mail addresses: fkarami@ee.iust.ac.ir (F. Karami), jposhtan@iust.ac.ir (J. Poshtan), mposhtan@pi.ac.ae (M. Poshtan). 0019-0578/$ see front matter 2010 Published by Elsevier Ltd on behalf of ISA. doi:10.1016/j.isatra.2009.11.005

sensors [6,7]. In this way, in [8] a failure monitoring system was proposed that combines the induction motor physical model with the data provided by a few number of sensors. In this approach, by combining the failure mechanisms with the symptoms faults are detected through the state and parameter value evaluations within the models. Deviations in the states and parameter values from their predefined norms are used as a tool for monitoring failures and suggesting the necessary maintenance. The general structure of a model-based approach is shown in Fig. 1 [9]. Most estimation methods which have been recently developed are computationally unmanageable, and require special assumptions on the form of the process and observer models which should be satisfied in practice. For these reasons, the families of Kalman Filter (KF), and its extensions are still the most widely used estimation algorithms [10]. In [11] Kalman filtering is proposed for estimating the state vectors of a discrete-time nonlinear system. In addition to the EKF, the stochastic observer theory contains Unscented Kalman Filter [12,13], and the stochastic sampling-based algorithms such as the recently proposed Particle Filters (see [14] and the references therein, and In [15] authors have tried to compare Particle and Kalman filtering for state estimation and as a case study of the state vector of a DC motor is utilized. In simulation tests it was observed that for a large number of particles the Particle Filter could succeed in accurately estimating the motors state vector, but at the same time it required higher computational effort.) Although the stochastic sampling-based algorithms are very close to optimality even when the system has uncertainty, they are far from being useful in practical applications due to their high computation load. The UKF is a novel estimation tool to replace EKF in nonlinear filtering problems. The EKF is a simple solution which

190

F. Karami et al. / ISA Transactions 49 (2010) 189195

Nomenclature xk f (., .) uk the state vector the machine dynamics the known input the process noise the observation matrix the measurement noise the estimate of the state the covariance matrix state prediction innovation Kalman gain state prediction covariance innovation covariance EKF state prediction covariance EKF state prediction rotor resistance of phase C the rotor resistances the resistance of each rotor bar the rotor resistance by l broken bars the Jacobean matrix of the function f d/q-axis stator current d/q-axis stator voltage d/q-axis stator flux Stator (Rotor) self-inductance Stator(Rotor) resistance Pole pair Mutual inductance. total leakage coefficient friction coefficient rotor inertia number of rotor bars number of broken bars rotor resistance of phase A rotor resistance of phase B the change of rotor resistance resistance at reference temperature reference temperature resistance temperature coefficient temperature increase
Fig. 1. Structure of a model-based fault detection approach [9].

wk
H

vk xk|k
Pk|k xk+1|k

vk+1

Kk+1 Pk+1|k v Pk+1|k

EKF Pk+1|k

xEKF1|k k+

Rrc Rr Rbar Rlr fx ia /ib ua /ub

a /b Ls (Lr ) Rs (Rr )
np Lm

Ff J N n Rra Rrb R r R0 T0

Fig. 2. Linear and nonlinear calculation of probability.

and compared with its nominal value to detect broken bars. It is estimated that when one bar is broken the rotor resistance increases only between 3% and 4% which is very small. Therefore, the rotor resistance estimation requires a high amount of accuracy, which can be provided by the proposed methods as will be shown later. This paper presents the comparison of performances of UKF and EKF in the estimation of parameter values in a fault detection problem. 1.1. Nonlinear Kalman filtering approach

is derived by direct linearization of the state equations to extend the (linear) Kalman filter (KF) into the nonlinear filtering domain. Although it is straightforward and simple, it has some disadvantages [16]. Amongst the most important are: 1. Costly calculation of Jacobean matrices. 2. Instability due to linearization and erroneous parameters. 3. White Gaussian noise assumption. To overcome these problems UKF is proposed. The main advantage of UKF is that it does not use any linearization for calculating the state prediction covariance. Consequently, it has more accurate covariance and Kalman gain estimates, and this leads to a more exact state estimation. Broken rotor bar detection is based on the rotor resistance of an induction motor that will increase when a rotor bar breaks [8,17]. In particular, rotor resistance is estimated and compared with its nominal value, at the same temperature and saturation conditions of the machine [18]. In fact, during the online monitoring of rotor resistance, a rapid step-like increase of this resistance relative to its initial value indicates that a rotor bar has broken. To detect broken bars, the measurements obtained from stator voltages and currents are processed for estimation of rotor resistance by EKF/UKF. In particular, rotor resistance is estimated

The KF uses only the first two moments of the state (mean and covariance) in its update rule. Although this is a relatively simple state representation, it has a number of important practical merits and demerits [10]: (1) The mean and covariance of an unknown distribution requires the maintenance of only a small and constant amount of information, but that information is sufficient to support most kinds of operational activities. (2) The mean and covariance are linearly transformable quantities. As shown in Fig. 2(a), the estimated means and covariances estimates can be maintained effectively when subjected to linear and quasi-linear transformations. Meanwhile the nonlinear transformation of means and covariances, as shown in Fig. 2(b) can be obtained by EKF and UKF. (3) The sets of estimated means and covariances can be used to characterize the additional features of distribution [10]. The most common application of the KF to nonlinear systems is in the form of EKF [19,20]. Exploiting the assumption that all transformations are quasi-linear, the EKF simply linearizes all nonlinear transformations, and substitutes Jacobean matrices for the linear transformations in the KF equations. Although the EKF maintains

F. Karami et al. / ISA Transactions 49 (2010) 189195

191

Fig. 3. A cloud sample is propagated through an arbitrary highly nonlinear function, mean and covariances are calculated, which can be regarded as the two approaches, EKF and UKF.

the elegant and computationally efficient recursive update form of the KF, it suffers from a number of serious limitations [10]. These can be listed as follows: (1) Linearized transformations are only reliable if the error propagation can be well approximated by a linear function. If this condition does not hold, the linearized approximate will undermine the performance of the filter or even cause the divergence of the estimates altogether [10,2125]. (2) Linearization can be applied only if the Jacobean matrix exists. However, this is not always the case. Some systems contain discontinuities, and in some cases the states themselves are inherently discrete (for more details see [10] and the references therein). (3) Calculating Jacobean matrices can be a very difficult task. The Jacobean equations frequently produce many pages of dense algebra [10]. It is worth noting that linearization errors in this case arise from an implicit truncation of the Taylor series description of the true transformed estimate. Therefore, keeping higher order terms may lead to better results. The unscented transformation (UT) was developed to address the deficiencies of linearization by providing a more direct and explicit mechanism for transforming mean and covariance information. This is illustrated in Fig. 3. The UT is founded on the intuition that it is easier to approximate a probability distribution function than an arbitrary nonlinear function or transformation [22]. As can be seen in Fig. 4, by choosing a suitable set of points (sigma points) with known mean x and covariance x , and applying a nonlinear function to each point, a cloud of transformed points will be achieved. The statistics of the transformed points can then be calculated to form an estimate of the nonlinearly transformed mean and covariance [10]. 2. Theoretical background of UKF and EKF Assume the state vector xk of a nonlinear dynamic evolves according to the discrete-time nonlinear state transition equation as follows: xk+1 = f (xk , uk ) + wk . (1)

Fig. 4. The principle of the Unscented Transformation [10].

This can be the model of induction motor where f (., .) is the machine dynamics, xk is the state of the machine at the sampling instant k, uk is the known input to the machine at time k, and wk is the process noise term representing modeling error. Also, it is assumed that we have a set of noisy measurements zk that are related to the state vector of the machine by the linear relationship zk = Hxk + vk (2) where H is the properly sized observation matrix, and vk is the white measurement noise related to the measuring device. The added white noise vectors wk , and vk are Gaussian and uncorrelated from each other with zero mean and covariances Q and R respectively. The states of the system are assumed to be unknown, and the aim is to find an estimate of the state vector xk|k , which is optimal in the sense of mean square error (MSE) [16]: xk|k = E xk |Z k
k

(3)

where Z = {z1 , z2 , . . . , zk }, and E {x|y} denotes the expected value of the quantity x, given the information y. Also, traditionally, the error estimates can be calculated by the covariance matrix Pk|k defined as:

Pk|k = E [xk xk|k ][xk xk|k ]T |Z k .

(4)

Since calculating the covariance matrix in the batch form and using these direct definitions are too cumbersome, recursive forms are adopted for both the state and the covariance estimates. These recursive update equations can be stated as follows:

xk+1|k+1 = xk+1|k + Kk+1 vk+1

(5)

192

F. Karami et al. / ISA Transactions 49 (2010) 189195

v T Pk+1|k+1 = Pk+1|k Kk+1 Pk+1|k Kk+1

(6)

where the state prediction vector xk+1|k , innovation vector vk+1 , Kalman gain matrix Kk+1 , state prediction covariance matrix Pk+1|k , v and innovation covariance matrix Pk+1|k may be defined as below: xk+1|k = E {f (xk , uk )|Z k } Pk+1|k = E {[xk+1 xk+1|k ][xk+1 xk+1|k ] |Z } Zk+1|k = H xk+1|k vk+1 = zk+1 zk+1
v Pk+1|k
T k

(7) (8) (9) (10) (11) (12) (13)

Eqs. (20) and (21) replace (7) and (8). The other UKF operations are the same as (9)(13). Note that operations in the new set of Eqs. (20), (21) and (9)(13) together with measurement updates given in (5) and (6), use only standard vector and matrix operations and need no derivative and Jacobean approximations. Also, the order of calculation is the same as that of EKF. 3. Induction motor model Induction motors can be described by fifth order nonlinear differential equations with four electrical variables (currents and fluxes), a mechanical variable (rotor speed), and two control variables (stator voltages) [26]. In ab axes fixed in the stator, the stator voltages and the states are: X T = [ ia U T = [ ua Y = [ ia ib

= HPk+1|k H + R

xz v Kk+1 = Pk+1|k (Pk+1|k )1 xz Pk+1|k

(22) (23) (24)

= Pk+1|k H .
T

ub ] ib ].

The quantities xk+1|k and Pk+1|k are called the state prediction and the prediction covariance of the state, respectively. They are vital for the overall filter performance. Eqs. (7) and (8) do not specify how these quantities are calculated. EKF assumes that the errors in the state estimates are small enough to approximate (7) and (8) to their first-order Taylor series. As a result, xk+1|k and Pk+1|k are calculated in EKF as:
EKF Pk+1|k =

Then, the nonlinear dynamic system of the induction motor can be described by the following equations

X = f (x(t ), U (t ), t )
Y = H (x(t ), t ) where

(25)

fx Pk|k fxT + Q

(14) (15)

xEKF1|k k+

= f ( k|k , uk ) x

where fx denotes the Jacobean matrix of the function f with respect to the state x. This linearization in EKF frequently yields wrong results in the estimates of the covariance, and thus the state. UKF solves the prediction problem by sampling the distribution of the state in a deterministic manner and then transforming each sample using the nonlinear state transition equation shown in Fig. 4 [10]. The n-dimensional random variable xk with mean xk|k and covariance Pk|k is approximated by 2n + 1 weighted samples or sigma points selected by the algorithm:

x1 x2 f (X , U ) = x3 x4 x5 =
Tr = Rs

= x1 + = x2 + = =
Lm Tr Lm Tr

KL Tr KL Tr

x3 + np KL x4 x5 + ua x4 np KL x3 x5 + ub x3 np x4 x5 x4 np x3 x5 Tl J 1

x1 x2

Tr 1 Tr

(26)

= np

Lm JLr

(x2 x3 x1 x4 )
Lm & =

Ff J

x5

Ls
Lr Rr

RrL2 m

Ls L2 r

& KL = L2 m Ls Lr

Ls Lr

Ls

(27)

X0 (k|k) = xk|k ,

W0 =

k k+n

(16) (17)

& = 1

Xi (k|k) = xk|k + [ (n + k)(Pk|k + Q )]i 1 Wi = 2(n + k) Xi+n (k|k) = xk|k + [ (n + k)(Pk|k + Q )]i 1 Wi+n = for i = 1, . . . , n 2(n + k)

The measurement matrix is thus given by H = 1 0 0 1 0 0 0 0 0 . 0 (28)

(18)

4. Application to the induction motor The induction motor model described above is used to apply the extended Kalman filter. By selecting the augmented state variables as X T = [ x1 x2 ib x3 x4 x5 x6 ] Rr ] (29)

where k is a free real number such that n|k = 0, and [ (n + k)(Pk|k + Q )]i is the ith column of the matrix of square root of (n + k)(Pk|k + Q ), and Wi is the weight associated with the ith point. Given these sets of samples, the prediction process is as follows: 1. Transforming each sigma point through the process dynamic: Xi (k + 1|k) = f (Xi (k|k), uk ). 2. Computing the state prediction:
2n

= [ ia

(19)

xk+1|k =
i=0

Wi Xi (k + 1|k).

(20)

3. Calculating the prediction covariance:

2n

Pk+1|k =
i =0

Wi [Xi (k + 1|k) xk+1|k ] [Xi (k + 1|k) xk+1|k ] .

T

the dynamic equation of induction motor will be modified (26) to: K x1 = 1 x1 + 2 x1 x6 + L x3 x6 + np KL x4 x5 + ua Lr x2 = 1 x2 + 2 x2 x6 + KL x4 x6 np KL x3 x5 + ub Lr Lm 1 x3 = x1 x6 x3 x6 np x4 x5 f (X , U ) = (30) Lr Lr x = L m x x 1 x x n x x 4 2 6 4 6 p 3 5 Lr Lr x5 = np Lm (x2 x3 x1 x4 ) Tl Ff x5 JLr J J x6 = 0

(21)

1 =

Rs

Ls

2 =

L2 m

Ls L2 r

(31)

F. Karami et al. / ISA Transactions 49 (2010) 189195

193

Fig. 5. Unloaded induction motor rotor resistance estimation with 74% R stepwise at 1 s.

4.1. Modeling of broken rotor bars It is assumed that the motor has a total number of N rotor bars, n bars of which are broken per phase. Here a symmetric broken bar condition at the rotor is considered in which the magnetic flux in the rotor and stator remains symmetric. Let Rra , Rrb , and Rrc be the resistances of rotor in phase A, phase B, and phase C, Rr the rotor resistances, Rbar the resistance of each rotor bar, Rlr the rotor resistance when there are broken bars, and Rr the change of rotor resistance. Besides, it is common to assume that: Rr = Rra = Rrb = Rrc . (32)

Table 1 Rated data of this simulated induction motor Rated values Power Frequency Voltage (/Y ) Current (/Y ) Speed Pole pair (np) Rr Rs l0 M 4 60 220/380 15/8.6 1440 2 1.15 1.1 0.013 0.203 0.004 0.042 0.032 KW Hz V A Rpm

Rated parameters

H H I/ C kg m2 IS

Since rotors are in parallel connection within one phase, it can be written [27]: Rr = Rlr = Rbar
N 3

Constants

J Ff

3Rbar N

(33) mask the effect of, a broken bar [26]. Therefore, as a part of the EKF/UKF estimation process, the thermal variation in the rotor resistance is compensated by using the same approach developed in [6,26]. In fact, it is easy to monitor the rotor temperature by estimating its resistance and then its temperature using R = R0 (1 + T ). (36)

Rbar
N 3

3Rbar N 3n 3Rbar N

(34) 3Rbar N 3n 3Rbar N 3n n.

Rr = Rr Rlr =

(35)

For example, Rr 3.7%for one broken rotor bar, andRr 7.7% for two broken rotor bars. 5. Simulation results The proposed estimation techniques, UKF and EKF, have been simulated on a squirrel cage induction motor whose ratings are summarized in Table 1. In Fig. 5, the state estimation performance of EKF and UKF are simulated. The resistance of rotor under the broken bar condition is estimated. This corresponds to an abrupt stepwise change on rotor resistant. In fact, apart from the thermal compensation problem, Figs. 5 and 6 show that a broken bar is responsible for the increase in the rotor resistance, and the proposed EKF and UKF techniques provide estimates before and after bar breakage. 5.1. Influence of temperature on the rotor resistance estimation One difficulty with the rotor resistance estimation and the broken bars detection is that a variation in the rotor temperature can cause a significant variation in Rr . A thermal variation in Rr can

Here R0 is the resistance at the reference temperature, and T is the temperature increase from 25 C. Eq. (36) provides a mean for thermal compensation of Rr. Assuming that this compensation works as planned, variations in the thermally compensated rotor resistance estimation can be attributed to broken bars alone. For illustration, the Kalman filter ability to track exponential profile for thermal effect of the rotor resistance has been tested. This is shown by Fig. 7. It can be noticed that the estimation accuracy is quite satisfactory for monitoring purpose with exponential thermal profile over time. Therefore, any changes in Rr is handled by UKF better than by EKF. Fig. 8 shows the estimation of Rr when the real value increases with exponential thermal profile over time, and one and two bars breakage at 0.5 s and 1 s. In practice, a change in rotor resistance with temperature occurs very slowly the thermal constant time is about one hour [18,28]. We underline that the effect of broken bar on apparent rotor resistance is detected quickly with the rapid change of the value of this resistance; however, the thermal effect happens very slowly in time. Consequently, these two effects can be distinguished easily because the estimation process of the rotor resistance is done in real time.

194

F. Karami et al. / ISA Transactions 49 (2010) 189195

Fig. 6. Unloaded induction motor rotor resistant estimation with 3.7% R at 0.5 s (one broken bar), 7.7% R at 1 sec (two broken bars) rotor resistance increase.

Fig. 7. Temperature effect on Rr estimation.

Fig. 8. Estimation of Rr with increase 3.3% R at 0.5 s (one broken rotor bar), 7.7% R at 1 s (two broken rotor bars) and resistance change with temperature.

6. Conclusion This paper adapts the derivative-free nonlinear Kalman filtering technique to the state estimation problem of detecting broken rotor bars in induction motors. The detection is based on the hypothesis that the rotor resistance of the induction motor will

increase apparently when a rotor bar breaks. Using this existing observer/state estimator, the implementation of a model-based fault detection scheme for induction motors can become more efficient and economical. It has been shown that UKF has a satisfactory state observation capability to detect broken bars. In particular, rotor resistance is estimated and compared with its nominal value to

F. Karami et al. / ISA Transactions 49 (2010) 189195

195

detect broken bars. Furthermore, two different Kalman filtering techniques are compared with each other. Although it is shown that both observers are promising estimation tools, each observer has specific advantages and disadvantages depending on the system requirements and operating conditions. First of all, UKF is far easier to implement because it does not require any linearizations, and eliminates the derivation and evaluation of Jacobean matrices. UKF recovers quickly, while EKF has a slower convergence. The error in the EKF estimate is biased. Linearization in EKF can produce highly unstable filter performance in low-cost applications with a slow micro processing unit, if the time-step intervals are not sufficiently small. As a result, UKF is recommended as a better alternative, especially for low-cost applications with limited computation ability. References
[1] Das S, Purkait P, Chakravorti S. Space-vector characterization of induction motor operating conditions. In: Proc.2008 national power systems conf., p. 5127. [2] Aroui T, Koubaa Y, Toumi A. Modeling and diagnostics of inductions machines under rotor failures. ICGST-ACSE Journal 2007;7(Nov):918. [3] Bornnett H, Soukup GC. Cause and analysis of stator and rotor failures in three-phase squirrel-cage induction motors. IEEE Transactions on Industry Applications 1992;28(4):92137. [4] Vas P. Parameter estimation, condition monitoring, and diagnosis of electrical machines. Oxford: Clarendon Press; 1993. [5] Kliman GB, Stein J. Methods of motor current signature analysis. Electric Machines & Power Systems 1992;20(5):46374. [6] Beguenane R, Benbouzid MEH. Induction motors thermal monitoring by means of rotor resistance identification. IEEE Transactions on Energy Conversion 1999;14(3):56670. [7] Benbouzid MEH, Vieira M, Theys C. Induction motors faults detection and localization using stator current advanced signal processing techniques. IEEE Transactions on Power Electronics 1999;14(1):1422. [8] Cho KR, Lang JH, Umans SD. Detection of broken rotor bars in induction motors using state and parameter estimation. IEEE Transactions on Industry Applications 1992;28(3):7029. [9] Liu K. Model-based failure detection in induction motors using nonlinear filtering. Ph.D. dissertation, Dept Systems. Control and Industrial Engineering Case Western Reserve Uni; 1995. [10] Julier SJ, Uhlmann J. Unscented filtering and nonlinear estimation. Proceedings of the IEEE 2004;92(3):40122.

[11] Giovaninia L, Dondo R. A fault detection and isolation filter for discrete linear systems. ISA Transactions 2003;42:6439. [12] Julier SJ, Uhlmann JK. A new extension of the Kalman filter to non linear systems. In: Proc SPIEInt soc opt eng. 1997. 3068, p. 18293. [13] Julier SJ, Uhlmann JK, Durrant-Whyte HF. A new method for the nonlinear transformation of means and covariances in filters and estimators. IEEE Transactions on Autom Control 2000;45(3):47782. [14] Doucet A, de Freitas N, Gordon N, editors. Sequential Monte Carlo methods in practice. New York: Springer-Verlag; 2001. [15] Rigatos Gerasimos G. Particle and Kalman filtering for state estimation and control of DC motors. ISA Transactions 2009;6272. [16] Akn B, Orguner U, Ersak A, Ehsani M. Simple derivative-free nonlinear state observer for sensor less AC drives. IEEE/ASME Transactions on Mechatronics 2006;11(5):63443. [17] Chan C, Wang H. An effective method for rotor resistance identification for high-performances induction motor vector control. IEEE Transactions on Industrial Electronics 1990;37(6):47782. [18] Nat Sad N, Nat Sad MS, Bouguechal N. Rotor resistance estimation of an induction motor to detect broken bars fault using HH method. Electric Power Components and Systems 2004;32:14961. [19] Jazwinski AH. Stochastic processes and filtering theory. San Diego, CA: Academic; 1970. [20] Sorenson HW, editor. Kalman filtering: Theory and application Piscataway. NJ: IEEE; 1985. [21] Costa PJ. Adaptive model architecture and extended KalmanBucy filters. IEEE Transactions on Aerospace and Electronic Systems 1994;30:52533. [22] Athans M, Wishner RP, Bertolini A. Suboptimal state estimation for continuous-time nonlinear systems from discrete noisy measurements. IEEE Transactions on Automatic Control 1968;AC-13:50418. [23] Austin JW, Leondes CT. Statistically linearized estimation of reentry trajectories. IEEE Transactions on Aerospace and Electronic Systems 1981;AES-17: 5461. [24] Mehra RK. A comparison of several nonlinear filters for reentry vehicle tracking. IEEE Transactions on Automatic Control 1971;AC-16:30719. [25] Lerro D, Bar-Shalom YK. Tracking with debiased consistent converted measurements vs. EKF. IEEE Transactions on Aerospace and Electronic Systems 1993;AES-29:101522. [26] Sad Nat Sad M, Benbouzid MEH, Benchaib andA. Detection of broken bars in induction motors using an extended kalman filter for rotor resistance sensorless estimation. IEEE Transactions on Energy conversion 2000;15(1): 6670. [27] Yu G. Observer-based fault detection for induction motors, MS dissertation, Dept Electrical and Computer Engineering, Alberta; 2006. [28] Nat-Sad M-S, Benbouzid MEH. H-G diagram based rotor parameters identification for induction motors thermal monitoring. IEEE Transactions on Energy Conversion 2000;15(1):148.