machines
Article
Rotor Asymmetry Detection in Wound Rotor Induction Motor
Using Kalman Filter Variants and Investigations on Their
Robustness: An Experimental Implementation
Furzana John Basha * and Kumar Somasundaram
Citation: John Basha, F.;
Somasundaram, K. Rotor Asymmetry
Detection in Wound Rotor Induction
Motor Using Kalman Filter Variants
and Investigations on Their
Robustness: An Experimental
Implementation. Machines 2023, 11,
910. https://doi.org/10.3390/
machines11090910
Academic Editor: Ahmed
Abu-Siada
Received: 10 August 2023
Revised: 8 September 2023
Accepted: 11 September 2023
Published: 14 September 2023
Copyright: © 2023 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
Machines 2023, 11, 910. https://doi.org/10.3390/machines11090910
Department of Instrumentation Engineering, Madras Institute of Technology Campus, Anna University,
Chromepet, Chennai 600044, India; skumar@mitindia.edu
* Correspondence: furzana.rizwan@gmail.com
Abstract: This paper analyzes the performance of Kalman filter-based estimators for robust filter-
ing and rotor asymmetry detection in wound rotor induction machines (WRIMs) using real-time
data. Filter models were designed based on an extended model of WRIMs. The detection of rotor
asymmetry was achieved by estimating the states of rotor resistance and speed using four filters. The
sensitivity of the parameters under healthy and asymmetry conditions was thoroughly analyzed and
categorized as low, medium, and high sensitivity parameters. Robust model-based estimators were
designed to minimize the probability of false alarms. The performance analysis demonstrated that
the dual unscented Kalman filter (DUKF) outperformed other Kalman filters such as the extended
Kalman filter (EKF), dual extended Kalman filter (DEKF), and unscented Kalman filter (UKF) for
state estimation of WRIM.
Keywords: wound rotor induction machine; rotor asymmetry detection; diagnosis; robust filtering;
parameter sensitivity analysis; model-based approach
1. Introduction
Wound rotor induction machines (WRIMs) have been widely employed in various
industrial applications due to their robustness and ability to handle high starting torque re-
quirements. In recent times, WRIMs have received special attention due to their wide utiliza-
tion in electrical power generation, particularly as double-fed induction generators (DFIGs)
in variable speed wind turbines [1]. However, despite their widespread use, WRIMs are
prone to various operational challenges, one of which is electrical rotor asymmetry.
Rotor faults are the most significant faults as they cause secondary failures that can
lead to serious motor malfunctions. The effect of rotor faults is associated with broken
rotor bars [2–4] and cracked end rings [5] in squirrel cage induction motors, as well as rotor
winding impedances [6–8] in wound rotor induction motors. Electrical rotor asymmetry
refers to an imbalance in the electrical characteristics of the rotor windings. Electrical
rotor asymmetry can arise from several factors, including variations in the resistance
and inductance of the rotor windings. These variations can be caused by manufacturing
defects, winding damage, or changes in the motor’s operating conditions [9,10]. Electrical
rotor asymmetry can have significant consequences on the motor’s performance and
reliability. It may lead to uneven torque production, reduced efficiency, increased losses,
and variations in motor speed [11]. Furthermore, electrical rotor asymmetry can cause
abnormal heating and stress on motor components, potentially resulting in premature
failures [12]. Consequently, developing efficient and reliable techniques for detecting rotor
asymmetry has become a subject of utmost importance in the field of motor diagnostics
and maintenance.
Techniques for detecting rotor asymmetry in WRIMs can be achieved using either
model-based or signal-based approaches. Signal-based approaches are widely available
https://www.mdpi.com/journal/machines
Machines 2023, 11, 910 2 of 21

in the literature. Stator current is directly influenced by rotor asymmetry fault (RAF),
and rotor current is directly influenced by stator asymmetry fault (SAF) [13]. Initially,
the motor current signature analysis (MCSA) method has been used to diagnose faults
in electrical motors using current analysis. More recently, diagnostic techniques based
on the analysis of stator current have been developed using methods such as continuous
wavelet transform (CWT) and iterative localized discrete Fourier transform (IDFT) [14],
instantaneous frequency analysis [15], stator current space vector magnitude (SCSVM),
and instantaneous magnitude of the stator current (IMSC) [13], which are good indicators
for RAF monitoring. To extract the defective signature component from current signals,
the extended Kalman filter (EKF) provided better results than CWT and IDFT [14]. The
RAF index is calculated by multiplying the original stator current signal by a function of
the dominant frequency of the original signal [15]. A low level of asymmetry was reliably
diagnosed during non-stationary operation by extracting the instantaneous frequency of
faulty stator currents [7].
On the other hand, model-based approaches are widely used to detect rotor faults in
squirrel cage induction motors (SCIMs). Rotor asymmetry faults were detected in a wind
turbine by [16], where there is more scope for detecting rotor asymmetry faults (RAFs) in
WRIMs using the model-based approach.
An observer constitutes a dynamic mechanism that provides estimations of the current
state of a system, utilizing prior information about its inputs and outputs. The Kalman
filter, an extensively employed optimal observer, estimates the state of a dynamic system
by considering both the system model and the measurements from sensors. From the
available literature, apart from the Kalman filter, high-gain and adaptive observers have
been utilized. A high-gain observer amplifies estimation sensitivity, rendering it effective
in scenarios with limited sensor measurements or significant noise. The adaptive observer
adjusts its parameters based on real-time data, ensuring reliable estimation even when the
system dynamics change over time.
The Kalman filter is used in the proposed work. The extended Kalman filter (EKF) is
useful for nonlinear systems and is easy to implement. However, it might not work well in
situations with a lot of uncertainty. The unscented Kalman filter (UKF) is more accurate for
nonlinear systems and handles uncertainty more effectively, but it can become complex in
high-dimensional situations.
Since rotor asymmetry is detected by observing variations in rotor resistance [17],
estimating rotor resistance through model-based approaches is performed. The extended
Kalman filter (EKF) is employed in model-based methods for detecting broken rotor bars in
induction motors [4,17]. Ref. [18] estimated speed using EKF, UKF, and NSF (neural state
filter) for squirrel cage induction motors (SCIMs). Additionally, model-based approaches
demonstrated the capability to overcome performance issues at extremely low and zero
speeds [9].
Based on the literature, the detection of rotor electrical asymmetry using model-based
approaches is described by [16]. The author chose the unscented Kalman filter (UKF) to
detect the fault. Numerous works on signal-based methods to detect rotor asymmetry
have been proposed. As demonstrated in this work, there is a considerable scope in the
model-based approach owing to its capability to accurately capture the motor’s dynamics.
The model-based approach utilizes an extended state space model with load torque and
rotor resistance as additional states. The motor’s parameters include stator resistance (Rs ),
rotor resistance (Rr ), stator inductance (Ls ), rotor inductance (Lr ), and mutual inductance
(Lm ).
To summarize, the proposed work comprises four key aspects:
1. Analytically calculating parameters for the real-time mathematical model of a WRIM
through standardized tests.
2. Rotor asymmetry fault resulting from changes in Rr also affects the motor’s speed.
Therefore, both Rr and speed are estimated using various variants of Kalman filters
(EKF, UKF, DEKF, DUKF). The developed Kalman filters for state estimation are
Machines 2023, 10, x FOR PEER REVIEW 3 of 22

2. Rotor asymmetry fault resulting from changes in Rr also affects the motor’s speed.
Machines 2023, 11, 910 Therefore, both Rr and speed are estimated using various variants of Kalman filters 3 of 21
(EKF, UKF, DEKF, DUKF). The developed Kalman filters for state estimation are
applied to real-time data, tested, and validated through filter convergence using the
kurtosis method.
applied to real-time data, tested, and validated through filter convergence using the
3. Qualitative analysis of the robustness of different filter variants is conducted by
kurtosis method.
performing state estimation on real-time data under both healthy and asymmetric
3. Qualitative analysis of the robustness of different filter variants is conducted by
conditions. The estimation is carried out with consistent initial conditions and
performing state estimation on real-time data under both healthy and asymmetric con-
different initial conditions.
ditions. The estimation is carried out with consistent initial conditions and different
4. A initial
significant contribution of this work lies in the parameter sensitivity analysis used
conditions.
4. to minimize the probability of
A significant contribution of false alarms.
this work lies in the parameter sensitivity analysis used
The
to paper is organized
minimize as follows:
the probability Section
of false 2 presents the mathematical modeling of
alarms.
WRIMThe andpaper
the methodology
is organizedfor rotor asymmetry
as follows: Section 2detection using
presents the EKF, UKF, DEKF,
mathematical modeling andof
DUKF. Section 3 presents the results and discussion on state estimation
WRIM and the methodology for rotor asymmetry detection using EKF, UKF, DEKF, and and analysis of
healthy and asymmetric rotor detection, robustness analysis, parameter sensitivity
DUKF. Section 3 presents the results and discussion on state estimation and analysis of anal-
ysis, filterand
healthy convergence,
asymmetricand accuracy
rotor under
detection, variousanalysis,
robustness operating conditions.
parameter Finally,analysis,
sensitivity infer-
ences are presented in the concluding section.
filter convergence, and accuracy under various operating conditions. Finally, inferences are
presented in the concluding section.
2. Materials and Methods
2. Materials
The chosen andWRIM
Methods is a versatile three-phase multi-functional machine equipped
with aTheslip-ring
chosenrotor.
WRIM Thisis ainherent
versatileflexibility
three-phaseenables the machinemachine
multi-functional to serveequipped
as either with
an
a slip-ring rotor.
asynchronous This inherentmotor
or synchronous flexibility enables the
or generator, asmachine
shown in to Figure
serve as1.either an asyn-
The distinct
chronousoforthe
terminals synchronous motor as
rotor designated or K,generator,
L, and Masare shown
clearlyin Figure 1. The
identified distinct1b.
in Figure terminals
These
of the rotor
terminals designated
possess uniqueas K, L, anddue
attributes M to
arethe
clearly
varyingidentified
winding in configurations
Figure 1b. These terminals
associated
possess
with each.unique attributes due
This adaptability to the the
enables varying
WRIM winding configurations
to operate effectively associated
in synchronouswith each.
or
This adaptability enables the WRIM to operate effectively in synchronous
asynchronous modes. The experimental investigations described in the study are con- or asynchronous
modes.while
ducted The experimental investigations
the machine operates in its described
asynchronous in themotor
studymode.
are conducted while the
This operational
machine
mode operates
denotes in its asynchronous
the machine performing motor as an mode. This operational
asynchronous motor during modethe denotes
experi-the
machine
ments. performing
To replicate as an
a state of aasynchronous
healthy rotor,motorthe K-L-Mduring the experiments.
terminals are shorted.To replicate
This action a
state symbolizes
likely of a healthyarotor, the K-L-M
balanced rotor terminals
condition are whereshorted. Thisterminals
all three action likelyare symbolizes
electrically a
linked, representing a symmetric setup. Conversely, to imitate an asymmetrical condi- a
balanced rotor condition where all three terminals are electrically linked, representing
symmetric
tion, the K-Msetup. Conversely,
terminals to imitate
are shorted. Thisan asymmetrical
configuration condition,
implies that the
the K-M
K andterminals
M termi- are
shorted. This configuration implies that the K and M terminals are
nals are electrically connected. This arrangement serves to mimic a situation where rotor electrically connected.
This arrangement
asymmetry serves toTable
is introduced. mimic a situation
1 presents where rotor asymmetry
specifications for the WRIM is introduced.
operating in Table
an 1
presents specifications for the WRIM operating in an asynchronous
asynchronous motor mode, outlining the motor’s nameplate information for this partic- motor mode, outlining
themode.
ular motor’s nameplate information for this particular mode.

(a) (b)

1.1.(a)(a)
Figure
Figure Front
Front view
view ofof WRIM,(b)
WRIM, (b)zoomed
zoomedview
viewofofrotor
rotorterminal
terminalpanel.
panel.

Table 1. Specification of WRIM when operating as an asynchronous motor.

Table 1. Specification of WRIM when operating as an asynchronous motor.
Power Voltage Current Frequency Power Speed
Power Voltage Current Frequency Power Speed
(KW) (V) (A) (Hz) Factor (RPM)
(KW) (V) (A) (Hz) Factor (RPM)
1.0 220/380 ∆/Y 4.32/2.5 ∆/Y 50 0.82 1385
1.0 220/380 ∆/𝑌 4.32/2.5 ∆/𝑌 50 0.82 1385

2.1. Mathematical Modelling of WRIM

In general, for state observation using EKF and UKF, the augmented state space model
of the induction motor with five state variables is considered in an arbitrary reference
frame [19]. However, in this work, an extended state space model with seven state variables
Machines 2023, 11, 910 4 of 21

is employed for offline estimation. Here, the sixth state variable represents the motor load
torque (TL ), and the seventh state variable is the rotor resistance (Rr ). The extended state
space model and the output equation of the three-phase wound rotor induction motor are
provided in Equations (1) and (2), respectively.

  0

2 Rr0 Lm

 .  − LRσs + Rr02Lm 0 Lm Pωm
Lσ Lr0
0 0 0   
1

ids Lr L σ Lr02 Lσ  ids 0
  Lσ
  
 .   Rr0 Lm 2 0

− R0r2Lm

Rs
i   0 − + − LmL Pω m 0 0 0 iqs   0 1
   
 qs   Lσ Lr02 Lσ σ rL0 Lr L σ Lσ 
 .      
Ψ   0
Ψdr 
 0 0

Rr0 LLm0 − RL0r − Pωm
  
 dr   0 0 0 0  Vds

 .   r r
(1)
  
Ψ  =   Ψqr 
0
+ 0 0

Rr0 LLm0 − RL0r
 
qr

0 − Pωm 0 0 0 

 Vqs
  
 .   r r   
ω   
 ω  0 0 
 m  −3P Lm 3P Lm
0 0 0 − J1 0  m 
2JL Lr0 ψqr 2J L Lr0 ψdr

 .   L    
T    TL   0 0
  
 L 
0 0 0 0 0 0 0

.

Rr Rr 0 0

0 0 0 0 0 0 0

ids
 
 iqs 
 
   ψdr 
ids 1 0 0 0 0 0 0  
=  ψqr  (2)
iqs 0 1 0 0 0 0 0 
 ωm 

 
T 
L
Rr
The notations are defined as follows:
ids and iqs : Stator current components in stationary axis
Rs and Ls : Stator resistance and inductance, respectively
Vds and Vqs : Stator voltage components in stationary axis
ψdr and ψqr : Rotor flux linkage components in stationary axis
Rr ’ and Lr ’: Rotor resistance and inductance referred to stator, respectively
P: No. of pole pairs
Lm : Mutual inductance
ω m : Rotor angular velocity
σ: Leakage factor
Lm 2

σ = 1− 0
L s Lr
Lσ = σLs ;
JL : Total inertia of the induction motor
In order to determine the exact parameter values, a series of tests, including the DC
resistance test, no-load test, and blocked rotor test [20], are undertaken. The following are
the procedures to conduct those tests.
1. No-load test: The motor is run through this test at its rated voltage without any load.
The purpose of this test is to determine the magnetizing inductance (Lm ). By applying
the rated voltage Vo , corresponding current Io and the power Wo is noted, the value of
Lm is calculated using the following formula.

Lm = 2πVf ss Im
Im = Io cos∅o
2. Blocked rotor test: The blocked rotor test is conducted to understand the performance
of an induction motor when it is operating under full load conditions, similar to the
short-circuit test. The objective of this test is to ascertain the values of stator inductance
(Ls ), rotor inductance (Lr ), and rotor resistance (Rr ). By applying the rated current (Isc )
and recording the corresponding voltage (Vsc ) and power (Wsc ), the values of Ls , Lr ,
and Rr are calculated using the following formulas.
Machines 2023, 11, 910 5 of 21

Zsc sin∅sc
Lls + Llr =
2π f s
Assuming, Lls = Llr , Ls = Lls + Lm ; Lr = Llr + Lm

Rr = Zsc cos∅sc − Rs

3. DC resistance test: By applying a DC voltage across two motor terminals and measur-
ing the resulting current, it is possible to determine the stator resistance (Rs ).
These tests are conducted for both healthy and asymmetrical rotor conditions. The
healthy rotor emulation is achieved by shorting the K-L-M terminals, as previously men-
tioned. The motor’s current, voltage, and power are recorded during the execution of the
aforementioned three tests. Similarly, to emulate an asymmetric condition, accomplished
by shorting the K-M terminals, the same tests are carried out and the motor’s current,
voltage, and power are recorded. Utilizing the equations provided below, equivalent circuit
values are calculated, revealing variations in rotor resistances that confirm the asymmetry
in the K-M connection. The parameter values resulting from these tests for both the healthy
and asymmetry scenarios are outlined in Table 2.

Table 2. Parameters calculated for WRIM.

Case Rs (Ω) Rr (Ω) Ls (H) Lr (H) Lm (H)

Healthy 8.8 7.768 0.863 0.863 0.831
Asymmetry 8.8 15.8501 0.4613 0.4613 0.4042

2.2. Methodology
This work aims to identify rotor asymmetries through a model-based approach. In this
method, the extended Kalman filter, unscented Kalman filter, dual extended Kalman filter,
and dual unscented Kalman filter are employed to estimate the state of rotor resistance,
enabling the detection of rotor asymmetry.

2.2.1. Extended Kalman Filter

The extended Kalman filter (EKF) is a recursive estimation method that expands upon
the fundamental principles of the Kalman filter. Unlike its linear counterpart, the EKF is
designed to accommodate non-linear systems by approximating them as linear around the
current estimated state. This crucial adaptation allows the EKF to effectively estimate the
true state of dynamic systems even in the presence of noisy measurements. At its core, the
EKF operates through a state-space representation, capturing system dynamics in terms of
first-order ordinary differential equations
The state vector x, control input u, system dynamics function f(x,u), and process noise
Q constitute the components of this representation. Furthermore, the relationship between
the state and measurements is encapsulated in the measurement equation z = h(x) + R,
where h(x) characterizes the non-linear connection between the state and the measurements,
and R signifies measurement noise.
The EKF algorithm unfolds through two primary steps: prediction and update. In
the prediction step, the state estimate is projected ahead using the system dynamics, and
the error covariance matrix is updated to account for process noise. This matrix involves
the Jacobian matrix Fk and the process noise covariance matrix Qk . In the subsequent
update step, the Kalman gain is computed, which involves the Jacobian matrix Hk and
the measurement noise covariance matrix Rk . The state estimate is then adjusted using
measurements, and the error covariance matrix is updated accordingly.
In this work, acquired voltage data are given as input and acquired current data are
given as measurement, and the residual acquired with this measurement is updated to the
next step. The Flowchart of Extended Kalman Filter is shown in Figure 2.
 update step, the Kalman gain is computed, which involves the Jacobian matrix Hk and
the measurement noise covariance matrix Rk. The state estimate is then adjusted using
measurements, and the error covariance matrix is updated accordingly.
In this work, acquired voltage data are given as input and acquired current data are
Machines 2023, 11, 910 given as measurement, and the residual acquired with this measurement is updated6 of to21
the next step. The Flowchart of Extended Kalman Filter is shown in Figure 2.

Figure 2. Flowchart of extended Kalman filter.

Figure 2. Flowchart of extended Kalman filter.
2.2.2. Unscented Kalman Filter
2.2.2. Unscented
The unscented Kalman Filter filter is an alternative to the extended Kalman filter (EKF)
Kalman
for The unscented
handling Kalman
non-linear filter is an
systems. alternative
The UKF avoids to thethe
extended Kalman
linearization filter
step (EKF) forin
employed
handling
the EKF, non-linear
which can systems. The UKF avoids
lead to inaccuracies the linearization
in non-linear step employed
systems. Instead in the
of linearizing the
system
EKF, dynamics
which andto
can lead measurement
inaccuraciesfunctions, the UKF
in non-linear employs
systems. a deterministic
Instead sampling
of linearizing the
technique
system knownand
dynamics as the “unscented functions,
measurement transformation”.
the UKF This technique
employs effectively captures
a deterministic sam-
the mean and covariance of a set of carefully selected sigma points
pling technique known as the “unscented transformation”. This technique effectively through non-linear
transformations.
captures the mean These sigma points
and covariance of a capture the distribution
set of carefully selected of the state
sigma pointsspace more
through
accuratelytransformations.
non-linear than linearized These
pointssigma
used in the EKF.
points captureThetheequations involved
distribution of theinstate
Unscented
space
Kalman
more Filter isthan
accurately shown in Figure
linearized 3.
points used in the EKF. The equations involved in Un-
scented InKalman
the UKF,Filter
the filter’s
is shown performance
in Figure 3.was enhanced through the appropriate selection
of scaling parameters α (alpha), β (beta), and κ (kappa) [21]. The spread of sigma points is
illustrated in Figure 4. By calculating the square root of the sigma points’ covariance, we
obtain the spread value. By adjusting the values of the scaling parameters, the distribution
of sigma points varies, and its corresponding value is also determined. The scaling factors
associated with the smallest obtained value are adopted for the UKF and DUKF algorithms.
A subset of the scaling factors is presented in Figure 4. It can be inferred from Figure 4f that
α = 0.1, β = 2, and κ = −3 yield the smallest value of 0.00027. Throughout this work, we
utilized the values α = 0.1, β = 2, and κ = −3.
Machines 2023, 11, 910
Machines 2023, 10, x FOR PEER REVIEW 7 of 22
7 of 21

Machines 2023, 10, x FOR PEER REVIEW 8 of 22

Figure3. 3.
Figure Unscented
Unscented Kalman
Kalman filter. filter.

In the UKF, the filter’s performance was enhanced through the appropriate selection
of scaling parameters α (alpha), β (beta), and κ (kappa) [21]. The spread of sigma points is
illustrated in Figure 4. By calculating the square root of the sigma points’ covariance, we
obtain the spread value. By adjusting the values of the scaling parameters, the distribu-
tion of sigma points varies, and its corresponding value is also determined. The scaling
factors associated with the smallest obtained value are adopted for the UKF and DUKF
algorithms. A subset of the scaling factors is presented in Figure 4. It can be inferred from
Figure 4f that α = 0.1, β = 2, and κ = −3 yield the smallest value of 0.00027. Throughout this
work, we utilized the values α = 0.1, β = 2, and κ = −3.

(a) (b)

(c) (d)

(e) (f)

Figure
Figure 4. Selection
4. Selection of appropriatescaling
of appropriatescaling parameter parameter
(a) α = 1, β (a)
= 2,α = 1,
and κ =β3 =(b)2,αand
= 0.1,κβ==32,(b)
andα = 0.1, β = 2, and
κ =κ3 =
(c)3α(c)
= 1,αβ =
= 2,
1,and
β =κ2,= 0and
(d) ακ==0.1, β = 2,
0 (d) α and κ =β
= 0.1, 0 (e)
= 2,α =and
1, β κ
= 2,
= and
0 (e)κ =α−3= (f)
1, αβ ==0.1, = 2, κ = −3 (f) α = 0.1,
2, βand
and κ = −3.
β = 2, and κ = −3.
2.2.3. Dual Extended Kalman Filter
Dual nonlinear Kalman filters involve the simultaneous operation of two filters,
wherein one is dedicated to estimating the state, and the other focuses on estimating pa-
rameters as shown in Figure 5. These filters operate concurrently to estimate both states
and parameters from observational data. While the state filter treats parameters as con-
 Machines 2023, 11, 910 8 of 21

2.2.3. Dual Extended Kalman Filter

Dual nonlinear Kalman filters involve the simultaneous operation of two filters,
wherein
Machines 2023, 10, x FOR PEER REVIEW
one is dedicated to estimating the state, and the other focuses on estimating 9 of 22
parameters as shown in Figure 5. These filters operate concurrently to estimate both
states and parameters from observational data. While the state filter treats parameters as
constants, the parallel parameter filter considers states as known entities. The equations
involved {y(k)
in −Dual xˆ (k | k − 1)} Kalman Filter is as follows:
C(k)Extended
C (k) =
 γ(k) = y(k) − yˆ (k | k − 1)
ParameterPrediction:
xˆ(k | k − 1) State Prediction:
ρ̂(k=|kC−(k)
1) = ρ̂(k − 1||k −ˆ(k|k 1) x̂ (k|k=−C1)(k) = x̂P(k (k−| k1|k−− 1)1C )+ (k)T + R
 = −1) Pyy (k)
Pρ (k|k − 1) = Pρ (k − 1|k − 1) + Qρ Rk
 
K  (k) = Pxy (k | k − 1)Pyy (k)−1 τ), ρ̂(k − 1|k − 1), u(k − 1),]dτ
F[x(H
k−1=
C(k)  x1)=
P (k | k − 1) = P (k | k − 1)C (k)T
xy   P(k|k − Φ−1),
xˆ (k|k (k(k|k
)P(−1),
k u−(k −11)|k − 1) Φ (k)T + Q

ˆ (k |Parameter 1) + K  (k)  y (k) − H   xˆ (k | k − 1), ˆ (k | k − 1) State Update:

k) = ˆ (k | k −Update:
Pρ ( k ) = C ( k ) P ( k | k − 1 ) C ( k ) T + R (k|k=−P
Kŷ(k) H[x̂(−k1|k − 1)]
xy (k)Pyy (k)
1) =
P (kyy| k) = [I −
ρ ρ
K∂{(k) CC((k)]
y(k)− k)x̂(kP
ρ
1)}| k − 1)
|k−(k
ρ
γ(k) = y(k) − ŷ(k|k − 1)
Cρ ( k ) =
Pyy (k) = C(k)P(k|k − 1)C(k)T + R
∂ρ
|k−1)
= C(k) ∂x̂(k∂ρ |ρ=ρ̂(k|k−1)
xˆ (k ) == x
ˆ∂x(k | k − 1) + K(k) γ(k)
h i
−1 C(|kk) ∂H
Kρ ( k ) = P ρ
xy ( k | k − 1 ) Pρ
yy k)
( [x̂(k|k−1),ρ(k|k−1),u(k−1)]
T K(k) = Pxy (k)Pyy−1 (k )
Pρ
xy (k|k − 1) = Pρ (k|k − 1)Cρ (k)

 
x̂(k|k) = x̂(k|k − 1) + K(k)γ(k)
|k)==I[I−−KK(k) |k |−k1−) 1)
ρ̂(k|k) = ρ̂(k|k − 1) + P(k
Kρ (k)[y(k) − Hρ [x̂(k|k − 1), ρ̂(k|k − 1)]] P(|kk) (k)CC((k)
k)]PP
(k(k
Pρ (k|k) = [I − Kρ (k)Cρ (k)]Pρ (k|k − 1)

Figure 5. Flowchart of dual Kalman filter.

Figure
2.2.4. 5. Flowchart
Dual of dual
Unscented Kalman
Kalman filter.
Filter
The equations involved in Dual Unscented Kalman Filter is as follows:
2.2.4. Dual Unscented Kalman Filter

Parameter Prediction: State Prediction:

2L
ˆ (k | k − 1) = ˆ (k − 1 | k − 1) xˆ (k | k − 1) = W m (i)  (k | k − 1, i)
i =0
 2Lρ

ŷρ (k | k − 1) =  Wm (i).Hρ[x̂(k | k − 1),χρ (k | k − 1,i)] 2L

i =0 yˆ ( k | k − 1) =  W m (i ).H [ χ * (k | k − 1, i )]
2Lρ i =0
Machines 2023, 11, 910
Pyyρ (k) = [W c (i){Hρ [χρ (k | k −1, i)]− yˆρ (k | k −1)}. 9 of 21
2L
Pyy(k) = [Wc (i){H[χ*(k | k −1,i)]− yˆ(k | k −1)}.
i=0

{Hρ [χρ (k | k −1, i)]− yˆρ (k | k −1)}T ] + Rρ i =0

Parameter Prediction:
2Lρ
State Prediction:
{H[χ*(k | k −1,i)]− yˆ(k | k −1)}T ] + R
2L

k|k − 
m
Pρ̂Pxyρ(k((k)
|k − 1) = ρ̂(ck − 1|k − 1)
= 1) =WP (i){ χ (k | k −1, i) − ρ̂(k | k −1)}. L ∑ W (i )χ(k |k − 1, i )
x̂ (k |k − 1)2=
(k − 1ρ|k − 1) + Q Pxy (k ) = W c (2L
i){χc* (k | k −1, i) − xˆ(k | k −1)}.
i =0
ρ ρ ρ
i =0
P(k|ki−
=0
1) = ∑ W (i ){χ(k |k − 1, i )− x̂ (k|k − 1)}.
{Hρ [χ ρ (k | k −1, i)]− ŷρ (k | k −1)}T i =0
T

Parameter Update: State Update:

{H{[χχ(k*|k(k− |1,ki)−−1,x̂i()] − y(k | k −1)}T
k|k −ˆ1)} + Q
2Lρ 2L

γŷρρ((kk|)k =− y1)(k=)i∑=−0 H ˆ(k | k − 1), ρ),(χkρ |(kk|k−−1)]

W (i).Hρ [x̂(k|k − 1ˆ
γŷ((kk|k )− = y ( k ) − yˆ ( k | k − 11)
m m ∗
1) = ∑ W (i ).H [χ (k|k − 1, i )]
ρ [x
1, i)]
i =0
2Lρ 2L
Pyy (k ) =ρ ∑ [W c (ρi−){ Pyy (k ) = ∑ [W c (i ){ H [χ∗ (k|k − 1, i )] − ŷ(k |k − 1)}.
K ( k ) = Pxy ( k{)HP[χyy−(1k|(kk−)1, i)] − ŷ(k|k − 1)}
ρ
1 Hρ [χρ (k |k − 1, i )] − ŷρ (k |k − 1)}.
Kρ (k ) = P (k )P
xy
i =0
yy (k ) i =0
∗ T
{ Hρ [χρ (k|k − 1, i )] − ŷρ (k|k − 1)}T ] + Rρ ]+R
2Lρ 2L
Pρ c
= ∑ W (i){χρ (k|k − 1, i) − ρ̂(k|k − 1)}. Pxy (k ) = ∑ W c (i ){χ∗ (k|k − 1, i ) − x̂ (k|k − 1)}.
(kγ
xy (k)
ρˆ ( k | k ) = ρ=ˆ ( k {|Hk [χ− 1()k|+k −K1,ρi)]( k−).ŷ γ(kρ|k( k−)1)}
i 0
ρ ρ ρ
T xˆ(k | k) = xˆ(k{|Hk[χ−(k1|k)−+1,K
i =0
∗ (kŷ).
i )] − (k1)})
|k − T

γ(k) = y(k ) − ŷ(k|k − 11)

γρ (k) = y(k) − Hρ [ x̂ (k |k − 1), ρ̂(k|k − 1)] −1 ( k )
K (k ) = Pxy (k ) Pyy
Kρ (k) = Pxy (k) Pyy −1 (k )
ρ ρ

P (k | k) = P (k | k −1) − K (k).P (k).KρT (k) P(k | k) = P(k | k −1) − K(k).Pyy (k).KT (k)
ρ̂ρ(k|k ) = ρ̂(k|k −ρ1) + Kρ (k).γρ (k ) ρ
ρ
ρ
yy
x̂ (k |k) = x̂ (k|k − 1) + K (k).γ(k)
P(k|k ) = P(k|k − 1) − K (k).Pyy (k ).K T (k)
T
Pρ (k |k) = Pρ (k|k − 1) − Kρ (k).Pyy (k ).Kρ (k)

2.2.5.
2.2.5. Real-Time
Real-TimeExperimental
ExperimentalSetup Setup
The
Thereal-time
real-timeexperimental
experimental setup setup consists
consists ofof aa1.0
1.0kWkWWRIM
WRIMconnected
connectedto tothe
theDAQ
DAQ
USB-6351 for data acquisition, as illustrated in Figure 6. Approximately
USB-6351 for data acquisition, as illustrated in Figure 6. Approximately 5000 sampling 5000 sampling
instances
instancesof ofstator
statorcurrent
currentand andvoltage
voltagedatadatawere
werecollected
collectedfromfromthetheWRIM
WRIMunder
underno-load
no-load
conditions
conditions at at aa sampling
sampling rate
rate ofof 11 ms
ms using
using the
theNI NIUSB-6351DAQ
USB-6351DAQ and and LabVIEW
LabVIEW 20182018
software. The acquired
software. The acquired voltage
voltage datadata are
are used
used as as input
input for
for the
the extended
extended Kalman
Kalman filter,
filter,
unscented
unscented Kalman
Kalman filter,
filter, and
and their
their respective
respective extensions
extensions for for state
state estimation,
estimation, while
while the
the
acquired current data serve as measured values used for residuals. The
acquired current data serve as measured values used for residuals. The four Kalman fil- four Kalman filters
are
tersestimated offline
are estimated with with
offline the assistance of MATLAB
the assistance of MATLAB R2021a software.
R2021a The assumptions
software. The assump-
made throughout the analysis are presented in
tions made throughout the analysis are presented in Table 3.Table 3.

Figure6.6.Real-time
Figure Real-timeexperimental
experimentalsetup.
setup.

3. Assumptions
Table The for EKF,data
acquired voltage UKF,and
DEKF, and DUKF.
current data at a sampling rate of 1 ms are shown in
Figures 7 and 8, respectively.
P(0|0) = 1 × diag(7)
Q = (5.3) × 10−5 ; (4.82) × 10−5 ; (1.5) × 10−6 ; (1.5) × 10−6 ; 155 × 10−7 ; 1 × 10−4 ; 8 × 10−6
R = (5.3) × 10−5 ; (4.82) × 10−5
α = 0.1, β = 2, κ = −3

The acquired voltage data and current data at a sampling rate of 1 ms are shown in
Figures 7 and 8, respectively.
 Figure 6. Real-time experimental setup.

The acquired voltage data and current data at a sampling rate of 1 ms are shown in
Machines 2023, 10, x FOR PEER REVIEW
Figures
7 and 8, respectively. 11 of 22
Machines 2023, 11, 910 10 of 21

(a) (b)
(a) (b)
Figure 7. Enlarged view of real-time acquired voltage data. (a) Healthy rotor, (b) Asymmetry rotor.
FigureFigure 7. Enlarged
7. Enlarged view
view of of real-time
real-time acquired
acquired voltage
voltage data.
data. (a) (a) Healthy
Healthy rotor, rotor, (b) Asymmetry
(b) Asymmetry rotor. rotor.

(a) (b)
(a) (b)
Figure 8. Enlarged view of real-time acquired current data. (a) Healthy rotor,(b) Asymmetry rotor.
Figure 8. Enlarged view of real-time acquired current data. (a) Healthy rotor, (b) Asymmetry rotor.
Figure 8. Enlarged view of real-time acquired current data. (a) Healthy rotor,(b) Asymmetry rotor.
The estimators employed in this study include the EKF, UKF, DEKF, and DUKF. The
real-time acquired current data is treated as the residual measured current. During the
The estimators employed in this study include the EKF, UKF, DEKF, and DUKF. The
filter evaluation process, noise covariance values P and Q were determined through a
real-time acquired current data is treated as the residual measured current. During the
trial-and-error approach to achieve optimal estimation, and the measured noise R was
filter evaluation process, noise covariance values P and Q were determined through a
quantified using the uncertainty formula as presented in Equation (3), derived from the
trial-and-error approach to achieve optimal estimation, and the measured noise R was
collected real-time data.
quantified using the uncertainty formula rh
as presented in Equation (3), derived from the
collected real-time data. i
Uncertainity = sum( xi − mean )2 /(n × 1(n − 1)) (3)
𝑈𝑛𝑐𝑒𝑟𝑡𝑎𝑖𝑛𝑖𝑡𝑦 = [𝑠𝑢𝑚(𝑥 − 𝑚𝑒𝑎𝑛 ) /(𝑛 × 1(𝑛 − 1))] (3)
where xi − readings, n-no of readings in the dataset.
where 𝑥 − 𝑟𝑒𝑎𝑑𝑖𝑛𝑔𝑠,
Motor speedn-no
and of readings
rotor in the
resistance (Rrdataset.
) are chosen as the reference states for evaluating
Motor speed and rotor resistance (R r) are chosen as the reference states for evaluat-
the filter performance. The average value of the calculated mean squared error (MSE) for
ing the filter
each performance.
instance The average
is employed value of the
as a performance calculated
metric. mean squared
Throughout errorthe
the analysis, (MSE)
values in
for each instance is employed as a performance
Table 3 are assumed for all the filters. metric. Throughout the analysis, the
values in Table 3 are assumed for all the filters.
3. Results and Discussion
Table 3. Assumptions for EKF, UKF, DEKF, and DUKF.
Motor speed and Rr are estimated by all four Kalman filters. To assess the effectiveness
𝑃(0|0)of = 1 ×filters,
these 𝑑𝑖𝑎𝑔(7) various analyses are conducted on their estimates under different operating
conditions,
Q = (5.3) namely
× 10−5; (4.82) × 10healthy and
−5; (1.5) × 10−6asymmetry
; (1.5) × 10−6cases.
; 155 × 10−7; 1 × 10−4; 8 × 10−6
The following
R = (5.3) × 10 ; (4.82) × 10
−5 analyses
−5 were conducted:
• β =State
α = 0.1, 2, κ =estimation
−3
• State estimation with different initial conditions
•
3. Results Analytical
and Discussionsolution of the model
• Parameter
Motor speed and sensitivity analysisby all four Kalman filters. To assess the effective-
Rr are estimated
• these
ness of Kurtosis
filters, various analyses are conducted on their estimates under different
operating conditions, namely healthy and asymmetry cases.
3.1. State Estimation
The following analyses were conducted:
In this analysis, acquired voltage and current data were utilized. The calculated
• State estimation
parameter values for the measured speed are 1475 rpm, and calculated rotor resistance is
 • • Kurtosis
Parameter sensitivity analysis
• Kurtosis
3.1. State Estimation
3.1. State Estimation
In this analysis, acquired voltage and current data were utilized. The calculated pa-
Machines 2023, 11, 910 rameterIn this analysis,
values acquired voltage
for the measured speed are and1475current
rpm,dataandwere utilized.
calculated Theresistance
rotor calculatedis11
pa-of 21
rameter values for the measured speed are 1475 rpm,
7.768 ohms. Similarly, in the asymmetry rotor condition, where terminals K-M areand calculated rotor resistance is
7.768 the
shorted, ohms. Similarly,
measured speed inisthe
757 asymmetry
rpm, and therotor condition,
calculated where terminals
rotor resistance K-M are
is 15.85 ohms.
shorted,
Using7.768 the
acquired measured speed is 757 rpm, and the calculated rotor resistance is 15.85 ohms.
ohms.voltage
Similarly,data
in as
theinput and acquired
asymmetry current data
rotor condition, where asterminals
residual,K-M speedareand
shorted,
Using
rotor the acquired
resistance voltage
values data
areisobtainedas input
from and acquired current data as residual, speed and
measured speed 757 rpm, anddifferent variants
the calculated of the
rotor Kalmanisfilter.
resistance 15.85 ohms. Using
rotor
The resistance values areemployed
initialvoltage
conditions obtained from different these
to acquired
estimate variants of the Kalman filter. and
acquired data as input and current states
data asfor both healthy
residual, speed and rotor
asymmetryThe initial
rotor conditions
casesareareobtainedemployed
set as {1.5; 1; 0.4; to estimate
0.3; 200 these
× (2 × pi/60); states for both
filter healthy
estimatesand
resistance values from different variants of the 0; 2}. Thefilter.
Kalman
asymmetry
for both healthy rotor cases are
andconditions
asymmetry set as {1.5; 1; 0.4;
rotor conditions 0.3; 200 ×
arethese(2 × pi/60);
illustrated 0; 2}.
in Figures The filter
9 andand estimates
10,asymme-
re-
The initial employed to estimate states for both healthy
fortry rotor cases are set as {1.5; 1; 0.4; 0.3; 200 × (2 × pi/60); 0; 2}. The filter estimates10,
both
spectively. healthy and asymmetry rotor conditions are illustrated in Figures 9 and forre-
both
spectively.
healthy and asymmetry rotor conditions are illustrated in Figures 9 and 10, respectively.

(a) (b)
(a) (b)
FigureFigure
9. Evolution of trueofand
9. Evolution trueestimated (a) Speed,
and estimated (b) Rotor
(a) Speed, resistance
(b) Rotor of healthy
resistance rotor rotor
of healthy winding.
winding.
Figure 9. Evolution of true and estimated (a) Speed, (b) Rotor resistance of healthy rotor winding.

(a) (b)
(a) (b)
Figure 10. Evolution of true and estimated (a) Speed, (b) Rotor resistance of asymmetry rotor
Figure 10. Evolution of true and estimated (a) Speed, (b) Rotor resistance of asymmetry rotor
winding.
Figure 10. Evolution of true and estimated (a) Speed, (b) Rotor resistance of asymmetry rotor
winding.
winding.
Based on observations from Figure 9a,b and findings provided in Table 4, it can be
Based onthat
inferred observations from of
in the scenario Figure 9a,b rotor
aFigure
healthy and findings provided
condition, specific in Table 4, are
behaviors it can be
noticeable.
inferred that in the scenario of a healthy rotor condition, specific behaviors are noticeable. be
Based on observations from 9a,b and findings provided in Table 4, it can
inferred that in the scenario of a healthy rotor condition, specific behaviors are noticeable.
Table 4. Estimated states of steady state data for healthy rotor winding.

STATES TRUE EKF UKF DEKF DUKF

SPEED 1475.45 1475.6 1475.59 1475.595 1475.48
Rr 7.768 7.758 7.842 7.7635 7.7695

During the initial transient state, both UKF and DUKF provided minor biases in
their speed and rotor resistance estimates. As the system reached a stable state, EKF
provided a bias of +0.15% for speed estimation, while UKF and DEKF demonstrated biases
of +0.14% and +0.145%, respectively. For DUKF, a minimal bias of +0.03% was observed in
speed estimation.
In terms of rotor resistance estimation, EKF provided a bias of −0.01%, UKF presented
a bias of +0.074%, DEKF displayed a bias of −0.0045%, and DUKF revealed a bias of
 Machines 2023, 11, 910 12 of 21

+0.0015%. Considering both speed and rotor resistance estimation, it is evident that DUKF
provided the least biased estimates.
From Figure 10a,b and Table 5, in the case of an asymmetry condition, it is observed
that during the initial transient state, all filters exhibit biases away from the true values.
Machines 2023, 11, 910 13 of 22

Table 5. Estimated states of steady state data for asymmetry rotor winding.

STATES TRUE EKF UKF DEKF DUKF

For rotor resistance estimation, EKF provided a +0.03% bias, UKF provided a +0.45%
SPEED
bias, DEKF provided757.78 757.77
a +0.25% bias, and DUKF756 757.15 bias. In both
presented a +0.02% 757.781
estima-
tionRrscenarios, DUKF
15.85consistently provided the16.3
15.88 least biased estimates.
16.1 15.87

Table 5. Estimated states of steady state data for asymmetry rotor winding.
For speed estimation, EKF provided a −0.01% bias, UKF provided a −1.78% bias,
STATES
DEKF provided a −0.63% bias,
TRUE EKFand DUKF presented
UKF DEKFbias.
a +0.001% DUKF
SPEED For rotor resistance
757.78 estimation,
757.77 EKF provided
756 a +0.03% bias,
757.15 UKF provided
757.781 a +0.45%
bias,
Rr DEKF provided
15.85 a +0.25% bias,
15.88 and DUKF
16.3presented a +0.02%
16.1 bias. In both estimation
15.87
scenarios, DUKF consistently provided the least biased estimates.
3.2. State Estimation with Different Initial Conditions
3.2. State Estimation with Different Initial Conditions
In this analysis, the states of speed and rotor resistance for healthy rotor data and
In this analysis, the states of speed and rotor resistance for healthy rotor data and
asymmetry rotor data are estimated under different initial conditions for both speed and
asymmetry rotor data are estimated under different initial conditions for both speed and
rotor resistance. These initial conditions include speeds of 100, 200, 300, and 400, each
rotor resistance. These initial conditions include speeds of 100, 200, 300, and 400, each
paired with corresponding resistances of 1, 2, 3, and 4.
paired with corresponding resistances of 1, 2, 3, and 4.
The utilized initial conditions are as follows: {1.5; 1; 0.4; 0.3; 100 × (2 × pi/60); 0; 1},
The utilized initial conditions are as follows: {1.5; 1; 0.4; 0.3; 100 × (2 × pi/60); 0; 1},
{1.5;
{1.5;1;1;0.4;
0.4;0.3;
0.3;200
200× ×(2 (2
× pi)/60); 0; 2},{1.5;
× pi)/60); 1; 0.4;
0; 2},{1.5; 0.3; 0.3;
1; 0.4; 300 300
× (2×× pi)/60); 0; 3},{1.5;
(2 × pi)/60); 1; 0.4; 1;
0; 3},{1.5; 0.3;
0.4;
400 × (2 × pi)/60); 0;4}.
0.3; 400 × (2 × pi)/60); 0;4}.
From
FromFigure
Figure11a,b,
11a,b, using all initial
using all initialconditions,
conditions,speed
speed and
and rotor
rotor resistance
resistance are are esti-
estimated
mated and the steady state remains
and the steady state remains the same. the same.

(a) (b)
Figure 11. Evolution of true and estimated states. (a) Speed, (b) Rotor resistance of healthy rotor.
Figure 11. Evolution of true and estimated states. (a) Speed, (b) Rotor resistance of healthy rotor.
Figure 12a,b demonstrate the estimation of speed and rotor resistance using all initial
Figure 12a,b
conditions, demonstrate
and the theremains
steady state estimation of speed and rotor resistance using all ini-
the same.
tial conditions, and the steady state remains the
Notably, when estimating states with various same.
initial conditions, the resulting estimates
converge to the same values. Among all the filters, DUKF consistently yields the least
biased estimates.
Inference 1:
The inference from the above analysis has two aspects:
1. Based on the acquired voltage and current data and the estimation of both states
for both the healthy and asymmetry rotor, it is inferred that DUKF is robust and
outperforms all other filter estimates.
2. All filter estimates remain consistent across various initial conditions.
Machines 2023, 10, x FOR PEER REVIEW 14 of 22

Machines 2023, 11, 910 Figure 12a,b demonstrate the estimation of speed and rotor resistance using all13ini-
of 21
tial conditions, and the steady state remains the same.
Machines 2023, 11, 910 14 of 22

Notably, when estimating states with various initial conditions, the resulting esti-
mates converge to the same values. Among all the filters, DUKF consistently yields the
least biased estimates.
Inference 1:
The inference from the above analysis has two aspects:
1. Based on the acquired voltage and current data and the estimation of both states for
both the healthy and asymmetry rotor, it is inferred that DUKF is robust and outperforms
all other filter estimates.
2. All filter estimates remain consistent across various initial conditions.

3.3. Analytical Solution(a) (b)

Figure 12.12.
Figure
The Evolution
analytical ofsolution
Evolution true
of and
true estimated
and
of modelstates.
estimated (a)(a)
istates.
ds and Speed, (b)(b)
iqsSpeed, Rotor
measured resistance
Rotor
and asymmetry
resistance rotor.
asymmetry
estimated results rotor.
are
shown below.
3.3.Notably,
Analyticalwhen
Solution
estimating states with
Figure 13a,b show the output ids and iqs various initial conditions,
for the healthy the resulting
rotor, displaying esti-
both meas-
mates The analytical
converge to thesolution
ured and estimated currents.same of model
values. Amongi ds and
all i
the qs measured
filters, DUKF and estimated
consistently results
yields are
the
shown
least below.
biased estimates.
Similarly, Figure 13c,d present the output ids and iqs for the asymmetry rotor, with
Figure 13a,b
Inference
both measured 1: show
and the output
estimated ids and iqs for the healthy rotor, displaying both measured
currents.
and estimated currents.
The inference from the above analysis has two aspects:
1. Based on the acquired voltage and current data and the estimation of both states for
both the healthy and asymmetry rotor, it is inferred that DUKF is robust and outperforms
all other filter estimates.
Healthy Rotor

2. All filter estimates remain consistent across various initial conditions.

3.3. Analytical Solution

The analytical solution of model ids and iqs measured and estimated results are
shown below.
Figure 13a,b show the output ids and iqs for the healthy rotor, displaying both meas-
ured and estimated currents.
Similarly, Figure 13c,d present the output ids and iqs for the asymmetry rotor, with
(a) (b)
both measured and estimated currents.
Rotor Rotor
Asymmetry
Healthy

(c) (d)
(a) (b)
Figure 13.13.
Analytical solution
Analytical of iof
solution ds iand iqs.
Figure ds and iqs .

3.4. Parameter Sensitivity

Similarly, Figure Analysis
13c,d present the output ids and iqs for the asymmetry rotor, with
both measured and estimated currents.
The parameter sensitivity analysis enables the study of how parameters are sensitive
towards the estimated states when parameters are increased or decreased. In this work,
3.4. Parameter Sensitivity Analysis
we are estimating rotor resistance as one of the states. Therefore, remaining parameters
such asThe parameter
stator sensitivity
resistance, analysis enables
stator inductance, the study of and
rotor inductance, howmutual
parameters are sensitive
inductance are
towards the estimated
considered for analysis. states when parameters are increased or decreased. In this work,
weInare estimating
this analysis, rotor resistance
we increased theasstator
one of the states.toTherefore,
resistances remaining
10%, decreased parameters
the stator and
rotor inductances to 10%, reduced mutual inductance to 10%, and explored the effects of
 Machines 2023, 11, 910 14 of 21

Machines 2023, 11, 910 such as stator resistance, stator inductance, rotor inductance, and mutual inductance are
15 of 22
considered for analysis.
In this analysis, we increased the stator resistances to 10%, decreased the stator and
rotor inductances to 10%, reduced mutual inductance to 10%, and explored the effects of
increasing resistances while decreasing all inductances on how the estimates respond.
increasing resistances while decreasing all inductances on how the estimates respond. The
The range
range of variation
of variation is calculated
is calculated by by subtracting
subtracting maximum-minimum.
maximum-minimum.
The
The results of the parameter sensitivity analysisare
results of the parameter sensitivity analysis arepresented
presentedfor
forboth
bothhealthy
healthyand
and
asymmetry
asymmetry conditions,
conditions, considering
consideringa parameter mismatch
a parameter of aof
mismatch 10% increase
a 10% in stator
increase re-
in stator
sistance, as shown
resistance, in Figure
as shown 14. 14.
in Figure
Healthy Rotor

(a) (b)
Asymmetry Rotor

(c) (d)

Figure 14.14.
Figure Filter estimates
Filter with
estimates a mismatch
with ofof
a mismatch 10% increase
10% ininRsR. s .
increase

FromFigure
From Figure 14a,
14a, for
for speed
speedestimation
estimationin in thethe
healthy condition,
healthy EKF EKF
condition, provided a +1.13%
provided a
bias, DUKF
+1.13% provided
bias, DUKF a −0.01%
provided bias,bias,
a −0.01% DEKF provided
DEKF a +0.1%
provided bias,bias,
a +0.1% andandUKF provided
UKF pro-
a +1.075%
vided a +1.075%biased estimate.
biased The range
estimate. of variation
The range is 1.14.
of variation FromFrom
is 1.14. Figure 14b, regarding
Figure 14b, re-
rotor resistance estimation under healthy conditions, EKF provided
garding rotor resistance estimation under healthy conditions, EKF provided a −0.65% a − 0.65% bias, UKF
provided a − 0.615% bias, DEKF provided a +0.005% bias, and
bias, UKF provided a −0.615% bias, DEKF provided a +0.005% bias, and DUKF provided DUKF provided a +0.07%
a biased
+0.07%estimate. The range
biased estimate. Theofrange
variation is 0.72. is 0.72.
of variation
InInFigure
Figure14c,14c,ininthe
thecase
caseofofspeed
speedestimation
estimationunder
underasymmetry
asymmetryconditions,
conditions,EKF EKF
provideda a+1.9%
provided +1.9%bias,
bias,UKF
UKFprovided
provideda a+1.6%
+1.6%bias,
bias,DEKF
DEKFprovided −0.61%bias,
providedaa−0.61% bias,and
and
provided a −
DUKF provided a −1.6% biased estimate. The range of variation is 2.51. Finally, in Figurein
DUKF 1.6% biased estimate. The range of variation is 2.51. Finally,
Figure
14d, 14d, for
for rotor rotor resistance
resistance estimation estimation in the asymmetry
in the asymmetry condition,condition,
EKF provided EKF aprovided
−1.1%
a − 1.1% bias, UKF provided a − 0.9% bias, DEKF provided
bias, UKF provided a −0.9% bias, DEKF provided a +0.25% bias, and DUKF provided a +0.25% bias, and DUKF a
provided
+0.45% a +0.45%
biased biased
estimate. Theestimate.
range ofThe range of
variation variation
is 1.55. is 1.55.inHowever,
However, both cases, in both cases,
the dual
the dual
filters filterslower
provided provided lower
biased biased estimates.
estimates.
Theresults
The resultsofofthe
theparameter
parametersensitivity
sensitivityanalysis
analysisare
arepresented
presentedfor forboth
bothhealthy
healthyandand
asymmetry conditions, considering a parameter mismatch of a 10%
asymmetry conditions, considering a parameter mismatch of a 10% decrease in stator and decrease in stator and
rotor inductance, as shown in
rotor inductance, as shown in Figure 15. Figure 15.
 bias, UKF provided a −0.9% bias, DEKF provided a +0.25% bias, and DUKF provided a
+0.45% biased estimate. The range of variation is 1.55. However, in both cases, the dual
filters provided lower biased estimates.
The results of the parameter sensitivity analysis are presented for both healthy and
Machines 2023, 11, 910 asymmetry conditions, considering a parameter mismatch of a 10% decrease in stator15and of 21
rotor inductance, as shown in Figure 15.

Healthy Rotor

(a) (b)
Asymmetry Rotor

(c) (d)

Figure 15. Filter estimates with a mismatch of 10% decrease in Ls and Lr .

Figure 15. Filter estimates with a mismatch of 10% decrease in Ls and Lr.
From Figure 15a, for speed estimation in the healthy condition, EKF and UKF provided
From Figure
a −9.86% 15a, provided
bias, DEKF for speedaestimation
-0.04% bias,inand
theDUKF
healthy condition,
provided EKF and
a +0.09% UKF
biased pro-
estimate.
vided a −9.86% bias, DEKF provided a -0.04% bias, and DUKF provided a +0.09% biased
The range of variation is 9.95. In Figure 15b, for rotor resistance estimation in the healthy
estimate.
condition,The
EKFrange of variation
provided is 9.95.
a −32.232% In UKF
bias, Figure 15b, fora rotor
provided resistance
−37.232% estimation
bias, DEKF in
provided
the healthy condition, EKF provided a −32.232% bias, UKF provided a
a +0.005% bias, and DUKF provided a +0.072% biased estimate. The range of variation−37.232% bias,
DEKF provided a +0.005% bias, and DUKF provided a +0.072% biased estimate. The
is 37.304.
range In
of Figure
variation
15c,isfor
37.304.
speed estimation in the asymmetry condition, EKF failed to estimate,
UKF provided a −45% bias, DEKF provided a −0.5% bias, and DUKF provided a −1.5%
biased estimate. The range of variation is 44.5. Finally, in Figure 15d, for rotor resistance
estimation in the asymmetry condition, EKF failed to estimate, UKF provided a +40.85%
bias, DEKF provided a +0.25% bias, and DUKF provided a +0.45% biased estimate. The
range of variation is 40.6. However, in both conditions, dual filters proved to be robust and
provided low biased estimates.
The results of the parameter sensitivity analysis are presented for both healthy and
asymmetry conditions, considering a parameter mismatch of 10% decrease in mutual
inductance, as shown in Figure 16.
From Figure 16a, speed estimation in the healthy condition, EKF provided a −115.12%
bias, UKF provided a −105.12% bias, DEKF provided a +0.06% bias, and DUKF provided
a +0.07% biased estimate. The range of variation is 115.27. From Figure 16b, regarding
rotor resistance estimation in the healthy condition, EKF provided a +50.232% bias, UKF
provided a +46.232% bias, DEKF provided a −0.005% bias, and DUKF provided a +0.072%
biased estimate. The range of variation is 50.237.
In Figure 16c, in the case of speed estimation under asymmetry conditions, EKF
provided a −9.5% bias, UKF provided a −8.3% bias, DEKF provided a −0.95% bias, and
DUKF provided a −2% biased estimate. The range of variation is 8.55. Finally, in Figure 16d,
For the rotor resistance estimation in the asymmetry condition, EKF provided a +2.3% bias,
UKF provided a +1.6% bias, DEKF provided a +0.25% bias, and DUKF provided a +0.45%
 Machines 2023, 11, 910 16 of 21

Machines 2023, 11, 910 17 of 22

biased estimate. The range of variation is 2.05. However, in both conditions, dual filters are
robust and provide low biased estimates.

Healthy Rotor

(a) (b)
Asymmetry Rotor

(c) (d)

Figure 16.16.
Figure Filter estimates
Filter with
estimates a mismatch
with of 10%
a mismatch decrease
of 10% in in
decrease Lm.Lm .

From
The Figure
results of16a,thespeed estimation
parameter sensitivityin the healthy
analysis condition, for
are presented EKF bothprovided
healthy anda
−115.12%
asymmetrybias,conditions,
UKF provided a −105.12%
considering bias, DEKF
a parameter provided
mismatch ofa10%
+0.06% bias, in
increase and DUKF
resistances
provided
and 10%a decrease
+0.07% biased estimate.asThe
in inductances, range
shown in of variation
Figure 17. is 115.27. From Figure 16b,
regardingFrom Figure
rotor 17a, speed
resistance estimation
estimation in theinhealthy
the healthy condition,
condition, EKFEKF provided
provided a −2.69%
a +50.232%
bias,
bias, UKF
UKF provided
provided a −2.25%bias,
a +46.232% bias,DEKF
DEKFprovided −0.1% bias,
provideda a−0.005% bias, and
and DUKF
DUKFprovided
provided a
+0.03% of
a +0.072% biased
biased estimate.
estimate. The
The range
range ofof variation
variation isis50.237.
2.72. From Figure 17b, regarding rotor
resistance estimation in the healthy condition, EKF
In Figure 16c, in the case of speed estimation under asymmetry provided a +0.838% bias, UKF provided
conditions, EKF
a +0.938%
provided bias, DEKF
a −9.5% bias, UKFprovided a +0.005%
provided a −8.3% bias, andDEKF
bias, DUKFprovided
providedaa−0.95%−0.072% of biased
bias, and
estimate.
DUKF The range
provided a −2%ofbiased
variation is 1.01.The range of variation is 8.55. Finally, in Figure
estimate.
16d, For In the
Figure
rotor17c, in the case
resistance of speed
estimation in estimation
the asymmetry undercondition,
asymmetry EKFconditions,
provided EKF a
provided
+2.3% a +1.65%
bias, UKF bias, aUKF
provided +1.6%provided
bias, DEKF a −0.15%
providedbias,a DEKF
+0.25%provided
bias, andaDUKF −0.55% bias,
pro-
and aDUKF
vided +0.45% provided a −1.65%The
biased estimate. of biased
range of estimate.
variation Theis range of variation
2.05. However, is 3.3.condi-
in both Finally,
in Figure 17d, for rotor resistance estimation
tions, dual filters are robust and provide low biased estimates.in the asymmetry condition, EKF provided
a +1.15% bias,ofUKF
The results the provided
parameterasensitivity
+2.05% bias, DEKFare
analysis provided a +0.25%
presented for both bias, and DUKF
healthy and
provided a +0.45% of biased estimate. The range of variation
asymmetry conditions, considering a parameter mismatch of 10% increase in resistances is 1.8. However, in both
andconditions,
10% decreasedualin filters are robust
inductances, asand
shownprovide low biased
in Figure 17. estimates.
Inference 2:
The inferences from parameter sensitivity analysis are threefold:
1. From Table 6, an observed low bias in the filter estimates occurs when the parameter
Rs is mismatched. A medium bias is present in the filter estimate when the parameter
Healthy Rotor

Lm is mismatched, and the estimates exhibit a high bias when Ls and Lr are mismatched.
Therefore, the stator resistance has low sensitivity, the mutual inductance has medium
sensitivity, and stator and rotor inductances are highly sensitive to the state estimation.
2. Even in the presence of parameter mismatches, dual filters remain robust in estimating
speed and rotor resistance.

(a) (b)
 Machines 2023, 11, 910 17 of 21
Machines 2023, 11, 910 18 of 23

3. Since dual filters offer a slightly biased estimate, they are proposed as effective soft
sensors for reducing false alarms.

Healthy Rotor

(a) (b)
Asymmetry Rotor

(c) (d)

Figure 17.17.
Figure Filter estimates
Filter with with
estimates a mismatch of 10%ofincrease
a mismatch in resistance
10% increase and 10%and
in resistance decrease in in-
10% decrease
ductances.
in inductances.

From
Table Figure
6. Range 17a, speed
Variation of allestimation
Filters. in the healthy condition, EKF provided a −2.69%
bias, UKF provided a −2.25% bias, DEKF provided a −0.1% bias, and DUKF provided a
Healthy Rotor Asymmetry Rotor
+0.03%
Casesof biased estimate. The range of variation is 2.72. From Figure 17b, regarding rotor
resistance estimation in the healthy Speedcondition, EKF Rr provided aSpeed
+0.838% bias, UKF Rr pro-
vided10%a increase
+0.938%inbias,
Rs DEKF provided 1.14 a +0.005%0.72 bias, and DUKF2.51 provided a −0.072%1.55 of
biased estimate. The
10% decrease in Ls and range of variation is 1.01.
9.95 37.354 44.5 40.6
LrIn Figure 17c, in the case of speed estimation under asymmetry conditions, EKF
provided a +1.65% bias, UKF provided
10% decrease in Lm 115.27
a −0.15%50.237
bias, DEKF provided
8.55
a −0.55% bias,
2.05
and
DUKF provided a −1.65% of biased estimate. The range of variation is 3.3. Finally, in
10% increase in Rs and
Figure 17d, for rotor resistance estimation in the asymmetry condition, EKF provided a
10% decrease in Ls , Lr , 2.72 1.01 3.3 1.8
+1.15%
Lm bias, UKF provided a +2.05% bias, DEKF provided a +0.25% bias, and DUKF
provided a +0.45% of biased estimate. The range of variation is 1.8. However, in both
conditions, dual filters are robust and provide low biased estimates.
3.5. Kurtosis
Inference 2:
TheThe designedfrom
inferences filterparameter
estimators, namely analysis
sensitivity EKF, UKF, areDEKF, and DUKF, are validated
threefold:
using the confidence level test of kurtosis. In this
1. From Table 6, an observed low bias in the filter estimates occurs context, an interval
when of ± 3sigma is
thex̂parameter
applied
Rs is mismatched. A medium bias is present in the filter estimate when the within
to all speed estimates to determine whether the estimates converge parame-the
99.73% confidence interval. Each speed estimation is obtained between x̂ + 3sigma and
x̂ −ter Lm is mismatched,
3sigma, which validates andtheir
the convergence.
estimates exhibit a high bias when Ls and Lr are mis-
matched. Therefore, the stator
The attainment of the fourth-order resistance
momenthas (kurtosis)
low sensitivity,
by the the mutual
designed inductance
filter estimators
has medium sensitivity, and stator and rotor inductances are highly
for healthy and asymmetrical rotor conditions is shown in Figures 18 and 19, respectively.sensitive to the
state estimation.
Zoomed figures for each filter estimation, depicting both the transient and steady states,
2. areEven
shown in in
the presence
Figures of parameter
18 and 19. mismatches, dual filters remain robust in esti-
mating speed and rotor resistance.
3. Since dual filters offer a slightly biased estimate, they are proposed as effective soft
sensors for reducing false alarms.
 3.5. Kurtosis
The designed filter estimators, namely EKF, UKF, DEKF, and DUKF, are validated
using the confidence level test of kurtosis. In this context, an interval of 𝑥 ± 3𝑠𝑖𝑔𝑚𝑎 is
applied to all speed estimates to determine whether the estimates converge within the
Machines 2023, 11, 910 99.73% confidence interval. Each speed estimation is obtained between 𝑥 + 3𝑠𝑖𝑔𝑚𝑎 and
18 of 21
𝑥 − 3𝑠𝑖𝑔𝑚𝑎, which validates their convergence.
The attainment of the fourth-order moment (kurtosis) by the designed filter esti-
matorsFigures
for healthy
18 andand asymmetrical
19 indicate that therotor
speedconditions is shown
state estimates in Figures
obtained by the 18 and
filter 19, re-
under
spectively. Zoomed
both healthy figures for conditions
and asymmetrical each filterconverge
estimation, depicting
within both the
the confidence transient
interval. Thus,and
the convergence
steady test validates
states, are shown the accuracy
in Figures 18 and 19.of the designed Kalman filters.

Speed Estimation Zoomed View at Transient state Zoomed View at Steady state
EKF

(a) (b) (a)

UKF

(d) (e) (f)

DEKF

Machines 2023, 10, x FOR PEER REVIEW 20 of 22

(g) (h) (i)

DUKF

(j) (k) (l)

Figure 18.
Figure Evolution of
18. Evolution of speed
speedestimation
estimationofof(a–c)
(a–c)EKF,
EKF,(d–f) UKF,
(d–f) (g–i)
UKF, DEKF,
(g–i) (j–l)(j–l)
DEKF, DUKF withwith
DUKF filterfilter
convergence healthyrotor.
convergence in healthy rotor.

Speed Estimation Zoomed View at Transient state Zoomed View at Steady state
EKF
 (j) (k) (l)

Machines 2023, 11, 910 19 offilter

Figure 18. Evolution of speed estimation of (a–c) EKF, (d–f) UKF, (g–i) DEKF, (j–l) DUKF with 21
convergence in healthy rotor.

Speed Estimation Zoomed View at Transient state Zoomed View at Steady state
EKF

(a) (b) (c)

UKF

(d) (e) (f)

DEKF

(g) (h) (i)

DUKF

(j) (k) (l)

Figure 19. Evolution of speed estimation of (a–c) EKF, (d–f) UKF, (g–i) DEKF, (j–l) DUKF with filter
convergence in asymmetry rotor.

4. Conclusions
This paper presents the detection of rotor asymmetry faults in wound rotor induction
motors (WRIM) using EKF, UKF, DEKF, and DUKF filters. Both healthy and asymmetric
rotor fault real-time data were gathered from an experimental setup of WRIM. Motor
parameters were calculated through standard tests. Since asymmetry faults are identified
as changes in rotor resistance, which, in addition, influences speed, both Rr and speed are
estimated. The results of state estimation are discussed under various operating conditions,
including healthy and asymmetric cases. An extensive analysis of healthy and rotor
Machines 2023, 11, 910 20 of 21

asymmetric fault detection, the robustness of different variants of Kalman filters, filter
sensitivity to parameter variations, and filter convergence were conducted.
From the results, it was observed that for detecting healthy and asymmetric condi-
tions, DUKF provided better performance for both speed and rotor resistance estimation.
Additionally, it can be concluded from parameter sensitivity analysis, dual filters were
robust in both healthy and asymmetric cases, and the results revealed that Rs has low
sensitivity, Lm has medium sensitivity, and stator and rotor inductances (Ls , Lr ) are highly
sensitive to all mismatch conditions. This analysis also led to the conclusion that, under all
the aforementioned sensitivity conditions, dual filters provide low-biased estimates and
can be proposed as effective variants of the Kalman filter to prevent false alarms.

Author Contributions: Conceptualization, F.J.B.; Methodology, F.J.B.; Software, F.J.B.; Validation,

F.J.B.; Formal Analysis, F.J.B.; Investigation, F.J.B.; Resources, F.J.B.; Data Curation, F.J.B.; Writing—
Original Draft Preparation, F.J.B.; Writing—Review and Editing, F.J.B. and K.S.; Visualization, F.J.B.;
Supervision, K.S.; Project Administration, K.S. All authors have read and agreed to the published
version of the manuscript.
Funding: This research received no external funding.
Data Availability Statement: The data presented in this study are available on request from the
corresponding author. The data is not publicly available due to working on another paper.
Conflicts of Interest: The authors declare no conflict of interest.

Nomenclature

WRIM Wound Rotor Induction Motor

SCIM Squirrel Cage Induction Motor
DFIG Doubly Fed Induction Generator
EKF Extended Kalman Filter
UKF Unscented Kalman Filter
DEKF Dual Extended Kalman Filter
DUKF Dual Unscented Kalman Filter
RAF Rotor Asymmetry Fault
SAF Stator Asymmetry Fault
FDD Fault Detection and Diagnosis
MCSA Motor Current Signature Analysis
IDFT Iterative localized Discrete Fourier Transform
CWT Continuous Wavelet Transform
SCSVM Stator Current Space Vector Magnitude
IMSC Instantaneous Magnitude of the Stator Current
Rs Stator Resistance
Rr Rotor Resistance
Ls Stator Inductance
Lr Rotor Inductance
Lm Mutual Inductance

References
1. Bebars, A.D.; Eladl, A.A.; Abdulsalam, G.M.; Badran, E.A. Internal electrical fault detection techniques in DFIG-based wind
turbines: A review. Prot. Control Mod. Power Syst. 2022, 7, 18. [CrossRef]
2. Ameid, T.; Menacer, A.; Talhaoui, H.; Harzelli, I. Rotor resistance estimation using Extended Kalman filter and spectral analysis
for rotor bar fault diagnosis of sensorless vector control induction motor. Measurement 2017, 111, 243–259. [CrossRef]
3. Ferrucho-Alvarez, R.E.; Martinez-Herrera, A.L.; Cabal-Yepez, E.; Rodriguez-Donate, C.; Lopez-Ramirez, M.; Mata-Chavez, R.I.
Broken rotor bar detection in induction motors through contrast estimation. Sensors 2021, 21, 7446. [CrossRef] [PubMed]
4. Farzaneh, K.; Poshtan, J.; Poshtan, M. Detection of broken rotor bars in induction motors using nonlinear Kalman filters. ISA
Trans. 2010, 49, 189–195.
5. Yetgin, A.G. Effects of induction motor end ring faults on motor performance. Experimental results. Eng. Fail. Anal. 2019, 96,
374–383. [CrossRef]
Machines 2023, 11, 910 21 of 21

6. Kia, S.H. Detection of Stator and Rotor Asymmetries Faults in Wound Rotor Induction Machines: Modeling, Test and Real-Time
Implementation. In Emerging Electric Machines-Advances, Perspectives and Applications; IntechOpen: London, UK, 2021.
7. Vedreño-Santos; Francisco; Riera-Guasp, M.; Henao, H.; Pineda-Sánchez, M.; Puche-Panadero, R. Diagnosis of rotor and stator
asymmetries in wound-rotor induction machines under nonstationary operation through the instantaneous frequency. IEEE
Trans. Ind. Electron. 2013, 61, 4947–4959. [CrossRef]
8. Zamudio-Ramirez, I.; Antonino-Daviu, J.A.; Osornio-Rios, R.A.; de Jesus Romero-Troncoso, R.; Razik, H. Detection of winding
asymmetries in wound-rotor induction motors via transient analysis of the external magnetic field. IEEE Trans. Ind. Electron. 2019,
67, 5050–5059. [CrossRef]
9. Bolognani, S.; Peretti, L.; Zigliotto, M. Parameter Sensitivity Analysis of an Improved Open-Loop Speed Estimate for Induction
Motor Drives. IEEE Trans. Power Electron. 2008, 23, 2127–2135. [CrossRef]
10. IEEE Standard 493–1997; IEEE Recommended Practice for the Design of Reliable Industrial and Commercial Power Systems (Gold
Book). IEEE: Piscataway, NJ, USA, 1998.
11. Toliyat, H.A.; Kliman, G.B. (Eds.) Handbook of Electric Motors; CRC Press: Boca Raton, FL, USA, 2018; Volume 120.
12. Climente-Alarcon, V.; Riera-Guasp, M.; Antonino-Daviu, J.; Roger-Folch, J.; Vedreno-Santos, F. Diagnosis of rotor asymmetries in
wound rotor induction generators operating under varying load conditions via the Wigner-Ville Distribution. In Proceedings of
the International Symposium on Power Electronics, Electrical Drives, Automation and Motion, Sorrento, Italy, 19–22 June 2012;
IEEE: Piscataway, NJ, USA, 2012; pp. 1378–1383.
13. Kia, S.H. Monitoring of wound rotor induction machines by means of discrete wavelet transform. Electr. Power Compon. Syst.
2018, 46, 2021–2035. [CrossRef]
14. Khalaf, I.R.; Watson, S.J.; Djurović, S.; Crabtree, C.J. An effective approach for rotor electrical asymmetry detection in wind
turbine DFIGs. IEEE Trans. Ind. Electron. 2018, 65, 8872–8881.
15. Marzebali, M.H.; Abolghasemi, V.; Ferdowsi, S.; Bazghandi, R. Manipulation of stator current signature for rotor asymmetries
fault diagnosis of wound rotor induction machine. IET Sci. Meas. Technol. 2022, 16, 523–532. [CrossRef]
16. Zhale, H.; Rahideh, A. Rotor electrical fault detection of wind turbine induction generators using an unscented Kalman filter.
Iran. J. Sci. Technol. Trans. Electr. Eng. 2020, 44, 979–988.
17. Said, M.S.N.; Benbouzid, M.; Benchaib, A. Detection of broken bars in induction motors using an extended Kalman filter for rotor
resistance sensorless estimation. IEEE Trans. Energy Convers. 2000, 15, 66–70. [CrossRef]
18. Kumar, S.; Prakash, J.; Kanagasabapathy, P. A critical evaluation and experimental verification of Extended Kalman Filter,
Unscented Kalman Filter and Neural State Filter for state estimation of three phase induction motor. Appl. Soft Comput. 2011, 11,
3199–3208. [CrossRef]
19. Murat, B.; Bogosyan, S.; Gokasan, M. Speed-sensorless estimation for induction motors using extended Kalman filters. IEEE
Trans. Ind. Electron. 2007, 54, 272–280.
20. Krishnan, R. Electrical Motor Drive; Prantice Hall: Upper Saddle River, NJ, USA, 2001.
21. Van Der Merwe, R.; Wan, E.A. The square-root unscented Kalman filter for state and parameter-estimation. In Proceedings of the
2001 IEEE International Conference on Acoustics, Speech, and Signal Processing (Cat. No. 01CH37221), Salt Lake City, UT, USA,
7–11 May 2001; IEEE: Piscataway, NJ, USA, 2001; Volume 6, pp. 3461–3464.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.