See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/260014657

Unscented Kalman Filters and Particle Filter Methods for Nonlinear State
Estimation

Article · December 2014

DOI: 10.1016/j.protcy.2013.12.457

CITATIONS READS

28 355

3 authors:

Katalin György Andras Kelemen

Sapientia Hungarian University of Transylvania Sapientia Hungarian University of Transylvania
11 PUBLICATIONS   43 CITATIONS    21 PUBLICATIONS   61 CITATIONS   

SEE PROFILE SEE PROFILE

László Dávid
Sapientia Hungarian University of Transylvania
31 PUBLICATIONS   159 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

High power electron beam welding and processing equipment, View project

All content following this page was uploaded by László Dávid on 30 June 2017.

The user has requested enhancement of the downloaded file.

 Available online at www.sciencedirect.com

ScienceDirect
Procedia Technology 12 (2014) 65 – 74

The 7th International Conference Interdisciplinarity in Engineering (INTER-ENG 2013)

Unscented Kalman filters and Particle Filter methods for nonlinear

state estimation
Katalin Györgya*, András Kelemenb, László Dávidc
a,b,c
Department of Electrical Engineering, Faculty of Technical and Human Sciences,
Sapientia University, Tîrgu Mureş, 540485, România

Abstract

For nonlinear state space models to resolve the state estimation problem is difficult or these problems usually do not admit
analytic solution. The Extended Kalman Filter (EKF) algorithm is the widely used method for solving nonlinear state estimation
applications. This method applies the standard linear Kalman filter algorithm with linearization of the nonlinear system. This
algorithm requires that the process and observation noises are Gaussian distributed. The Unscented Kalman Filter (UKF) is a
derivative-free alternative method, and it is using one statistical linearization technique. The Particle Filter (PF) methods are
recursive implementations of Monte-Carlo based statistical signal processing. The PF algorithm does not require either of the
noises to be Gaussian and the posterior probabilities are represented by a set of randomly chosen weighted samples.

© 2013 The Authors. Published by Elsevier B.V.

© 2013 The
Selection andAuthors. Published
peer-review by Elsevier Ltd.
under responsibility Open access of
of Department under CC BY-NC-ND
Electrical license.Engineering, Faculty of Engineering,
and Computer
Selection
“Petru and peer-review
Maior” University ofunder
Tîrguresponsibility
Mureș. of the Petru Maior University of Tirgu Mures.

Keywords: nonlinear state estimation; Unscented Kalman Filter; Particle Filter algorithm.

1. Introduction

Many various problems in science and especially many various control algorithms require determination of the
states for studied or controlled system. In more realistic cases only the outputs of the plant (rather than the state
vector) can be measured. In this case the state estimation process often plays an important role in the process control

* Corresponding author. Tel.: +40-758-662404.

E-mail address: [email protected], [email protected], [email protected]

2212-0173 © 2013 The Authors. Published by Elsevier Ltd. Open access under CC BY-NC-ND license.
Selection and peer-review under responsibility of the Petru Maior University of Tirgu Mures.
doi:10.1016/j.protcy.2013.12.457
66 Katalin György et al. / Procedia Technology 12 (2014) 65 – 74

implementation and in the monitoring applications. In case of the industrial processes there are many disturbing
factors which influence the process control such as model and measurement uncertainties. In this paper we use the
Gaussian probability distribution function to represent these uncertainties. The simple Kalman filter (KF) can be
used only in case of linear dynamic systems and this algorithm propagates the mean and covariance of the
probability distribution function of the model state in an optimal way. Almost all practical systems have some
nonlinearity. If the studied system can be described with a nonlinear dynamic model than the most commonly used
algorithm for the state estimation is the Extended Kalman Filter (EKF). In this case the state distribution is
propagated analytically through a linear approximation of the system around the operating point at each time instant.
This linear approximation may introduce errors in the estimated states, in other words with this method the results
may not be appropriate for some systems. The Unscented Kalman Filter (UKF) is a derivative free method, and it
resolves this problem by using a deterministic sampling approach. The Particle Filters (PF) method is a recursive
implementation of the Monte Carlo based statistical signal processing. The aim of this work is to compare these state
estimation methods for different nonlinear state space models.
Section 2 describes the principles of the nonlinear state estimation problem. The section 3 presents the general
form of the of the EKF algorithm. Section 4 first introduces briefly the unscented transformation and after that the
UKF algorithm. The section 5 presents the particle filter and the section 6 studies different examples to compare the
performances of the EKF, UKF and PF algorithms. One of these examples is the nonlinear reactive magnetron
sputtering process. Discussion of the results and the conclusions are presented in Section 6.

2. The nonlinear state estimation problem

To define the state estimation problem first we consider the general form of the discrete nonlinear process model,
with m input signals, p output signals and the order of system is n. The discretization was made with sampling time
Ts and we use the following notation for signal sequence xk x(Ts ˜ k ) . In this reason the nonlinear discrete model is

xk F ( x k 1 , u k 1 )  w k 1
(1)
y G( x k , u k )  v k
k

where xk ƒn is the state vector, ukƒm is the system input vector and ykƒp is the noisy output vector of the
system. The functions F: ƒnoƒn and G:ƒpoƒn are nonlinear and they need to be continuous. The wk is one n
dimensional process noise sequence and vk is p dimensional observation (measurement) noise sequence. Both noises
are Gaussian (normal distribution), independent random processes with zero means and known time invariant
covariance matrices. If the E{} is the expected value operator we can write:

w ~ N (0,V w ) Q cov{w} E{w ˜ wT }

(2)
v ~ N (0, V v ) R cov{v} E{v ˜ v T }

The objective of the estimation problem is to recursively estimate xk from the output measurements yk.. In
accordance with Bayes theory this mean, that recursively calculate the estimation of xk at time k given the dates
y1,…,yk up to time k. This is required calculation of the probability distribution function pdf ( xk | y1k: ) . We suppose
that the initial pdf function of the state vector ( pdf ( x0 | y0 ) ) is known and the pdf ( xk | y1k: ) is obtained recursively
in two sections: prediction step and update (correction) step.

3. The EKF algorithm for nonlinear state estimation

The Kalman Filter (KF) propagates the mean and covariance of the probability distribution function of the model
state in an optimal way with minimization of the mean square error. The KF dynamics results from the consecutive
cycles of prediction and filtering. The KF algorithm for state estimation can be applied just to the linear discrete
time system model, when the model is nonlinear then we have to use the EKF algorithm. The EKF algorithm we
 Katalin György et al. / Procedia Technology 12 (2014) 65 – 74 67

apply to the discrete nonlinear system model (1), where wk and vk are Gaussian distribution noises with known
parameters (2). The extended algorithm is almost similar with KF algorithm, but in this case there was introduced
one local linearization step of the model equations.
The steps of the EKF state estimation algorithm is presented below [1,5,12]:
0. Initialization step at k=0.
- Initial estimated state vector ~
x o E{x 0 }
~
- Initial covariance matrix: P E{( x  ~
0, x x ) ˜ (x  x )T }
0 0 0 0
1. Local linearization step ( k t 1 ): linearizing the nonlinear model functions F() and G() we can calculate the
following matrices:

ª wF1 wF1 º ª wG1 wG1 º

« wx  
wxn » « wx wxn »
« 1 » « 1 »
)k «    » Ck «    » (3)
« wFn 
wFn » « wG p 
wG p »
« wx1 wx n » « wx wx n »¼
¬ ¼x ~
x k 1 ¬ 1 x ~
x k 1

The linearization method utilizes just the first term in the Taylor expansion of the nonlinear functions
2. Prediction step: Calculation of the predicted state mean and covariance (time update)

~ ~ ~
x k F (~
x k 1 , u k 1 ) Px,k ) k ˜ Px,k ˜ )Tk  Q (4)

3. Calculation of the filter gain vector:

Kk
~
 ~
Px,k ˜ CkT ˜ Ck ˜ Px,k ˜ CkT  R
1
 (5)

4. Correction step: The estimates are updated with latest observation (measurement update)

~ ~ ~ ~
xk x k  K k ˜ ( yk  G( ~
x k , u k )) Px,k ( I  K k ˜ Ck ) ˜ Px,k (6)

The EKF gives an approximation of the optimal estimate. The non-linearities of the system’s dynamics are
approximated by a linearized version of the non-linear system model around the last estimated state [5]. This
algorithm can be divergent if the consecutive linearizations are not a good approximation of the non-linear model.

4. The Unscented Transformation and UKF algorithm for nonlinear state estimation

Consider the non-linear system described by the equation (1). The state distribution is also represented by
Gaussian random variables, but this method is using a minimal set of carefully chosen sample points. These points,
called sigma points, completely capture the true mean and covariance of the states and are propagated through the
nonlinearity. For calculating the statistics of a random variable which undergoes a nonlinear transformation we can
use the unscented transformation (UT) [2, 4].
Consider a random variable x (dimension n), which is propagating through a nonlinear function y f (x) . The
mean of x is x E{x} and the covariance of x is Px E{( x  x ) ˜ ( x  x ) T } . To calculate the statistics of y, we form
a matrix  of 2 ˜ n  1 sigma vectors according to the following:
68 Katalin György et al. / Procedia Technology 12 (2014) 65 – 74

0 x
i x J ˜P x i i 1,...,n (7)
i x J ˜ P  x i i n  1,...,2 ˜ n

where J n  O is a scaling parameter and O D 2 ˜ (n  k f )  n . The constant D determines the spread of the sigma
points around the mean value x and usually D[10-4,1]. The constant kf is a secondary scaling parameter and it is
usually set to 0 for state estimation. P x i is the i-th column of the matrix square root. These i vectors are
propagated through the nonlinear function y i f (i ) i 0,...,2 ˜ n .The mean and covariance for y are
approximated using a weighted sample mean and covariance of the posterior sigma points:

2˜n 2˜n
y | ¦Wi( m) ˜ yi Py | ¦Wi(c) ˜ ( yi  y ) ˜ ( yi  y )T (8)
i 0 i 0

The weights are given by

W0( m) O /(n  O ) W0(c ) O /(n  O )  (1  D 2  E )

(9)
Wi( m) Wi(c ) 1 /(2 ˜ (n  O )) i 1,...,2 ˜ n

where E is used to incorporate prior knowledge of the distribution of x (for Gaussian distribution the optimal value is
2). The standard UKF state estimation algorithm, with additive (zero mean) noise, is presented below [3,5]:
0. Initialization step at k=0:
- initial estimated state vector: ~
x o E{x 0 } ;
~
- initial covariance matrix: P E{( x  ~
x ) ˜ ( x  x )T } .
x,0 0 0 0 0
1. Sigma points’ calculation for k>=1

ª~ ~ ~
k 1 x ~
x  J ˜ Px,k 1 xk 1  J ˜ Px,k 1 º
~ (10)
«¬ k 1 k 1 »¼

2. Propagation of the sigma points:

- transform the sigma points through the state-update function: *k F (k 1 , u k 1 ) ;
- calculate the apriori state estimate and apriori covariance, where the weights Wi(m) and Wi(c ) are defined in
accordance with relations (9):

2˜n ~ 2˜n
~
xk ¦Wi
( m)
˜ (*k ) i Px,k ¦Wi
(c)
˜ ((*k ) i  ~
xk ) ˜ ((*k ) i  ~
xk )T  Q (11)
i 0 i 0

.3. Update of the output vectors:

- transform the sigma points through the measurement-update function 8k* G(k 1 , u k ) ;
- calculate the mean and covariance of the measurement vector:

2˜n ~ 2˜n
~
yk ¦Wi
( m)
˜ (Yk* ) i Py,k ¦Wi
(c )
˜ ((Yk* ) i  ~
y k ) ˜ ((Yk* ) i  ~
y k )T  R (12)
i 0 i 0

4. Calculate the cross covariance matrix:

 Katalin György et al. / Procedia Technology 12 (2014) 65 – 74 69

~ 2˜n
Px, y ,k ¦Wi
(c )
˜ (( X k* ) i  ~
xk ) ˜ ((Yk* ) i  ~
yk )T (13)
i 0

5. Calculation of the Kalman filter gain vector

~ ~
Kk Px, y,k ˜ ( Py,k ) 1 (14)

6. Calculate the estimated state and the covariance in accordance with the standard Kalman filter algorithm:

~ ~ ~ ~ ~
xk x k  K k ˜ ( yk  ~
yk ) Px,k Px,k  K k ˜ Py,k ˜ K kT (15)

The UKF principle is simple and easy to implement because it does not require the calculation of Jacobians at
each time step [5]. The most computationally intensive operation in the UKF corresponds to calculating the new set
of sigma points at each time update.

5. The Particle Filter algorithm

The Particle filter algorithms is a recursive implementation of the Monte Carlo based statistical signal processing
method. The method approximates the Bayesian posterior probability density function (pdf) with a set of randomly
chosen, weighted samples. Each sample of the state vector is referred to as a particle. A sufficiently large number of
particles guarantee almost sure convergence to the true probability distribution function [6,7,8]. The discrete
weighted approximation to the posterior pdf we can define

Ns
p( xk | y1:k ) | ¦ wki ˜ G ( xk  xki ) (16)
i 1

where G() is the Dirac delta measure, the Ns is the number of samples, the {xki | i 1,...,N s } the set of random
samples at time k, the {wki | i 1,...,N s } are normalized weights

Ns
i
¦ wk 1 (17)
i 1

The Sequential Importance Sampling (SIS) is the basic framework for particle filters: The main idea is to
represent the required posterior density by a set of random samples with associated weights and to compute
estimates based on these samples and weights [6]

p( y k | xki ) ˜ p( xki | xki 1 )

wki v wki 1 ˜ (18)
q( xki | x1i:k 1 , y k )

where q(xk|x1:k-1,yk) is the importance probability density function and v is the proportionality symbol. A main
problem with the SIS algorithm is the degeneracy phenomena, where after a few iterations; just one particle will
have non negligible weight. This implies that a large computational effort is devoted to updating particles whose
contribution to the approximation of p(x k|y1:k) is almost zero [6]. One measure of the degeneracy is the estimated
value of the effective sample size:

Ns
N eff ( ¦ ( wki ) 2 ) 1 (19)
i 1
70 Katalin György et al. / Procedia Technology 12 (2014) 65 – 74

Serious degeneracy is indicated when the Neff is small and Neff≤N. The degeneracy phenomena can be reduced
with a good choice of the prior density function and if we use different resampling method. The importance
resampling algorithm consists in choosing the prior density p(xk|xk-1) as importance density q(xk|x1:k-1,yk), in
accordance with that we can write:

wki v wki 1 ˜ p( yk | xki ) (20)

The basic idea of resampling is to eliminate particles what have small weights and to concentrate on particles
with large weights. In the literature are several method to solve this problem: multinomial resampling, residual
sampling, systematic resampling, etc.
The particle filter algorithm for state estimation is presented below [7]:
1. Initialization (k=0): Set the initial state vector {x0i | i 1,...,N s } and weights w0i 1 / N s , i 1,...,N s (assuming
with that the all particles are equally probable at the start of the algorithm).
2. Measurement update: Update the weights wki wki 1 ˜ p( yk | xki ) , whit normal probability density function
(normpdf) and normalize the weight wki wki / ¦ wki .
3. Resampling: apply the chosen resampling method (if the degeneracy phenomen is serious) to the set of
particles {xki | i 1,...,N s } and their weights {wki | i 1,...,N s } to obtain a new set of particles and set of weights.
4. Compute the estimate :

Ns
~
xk j
¦ wk ˜ xk
j
(21)
j 1

5. Set k = k+1, and iterate the algorithm to step 2 at end of the simulation time interval.

6. Simulation studies

In this section we present some example and simulation results to compare the different characteristics of the
three nonlinear state estimation method EKF, UKF and PF algorithms. The simulations of the estimation algorithms
were made for three different types of systems. First we test this estimation algorithm for one theoretical nonlinear
system, where we introduce different nonlinearities in observation model and at the end we test the presented
algorithms to one complicate nonlinear system models (reactive sputtering system model). The numerical algorithms
were implemented in Matlab environment. In all examples the simulations have been made in the following
conditions:
- the process noise and measurement noise are applied to the system, both noises are Gaussian with zero mean
and with known standard deviation ( w ~ N (0,V w ), v ~ N (0,V v ) );
- the initial values (the initial state vector and the initial covariance matrix), the process noise and measurement
noise covariance matrices are chosen to be the same for all algorithms (EKF and UKF):

~
Px,0 V x2 ˜ I n Q V w2 ˜ I n R V v2 ˜ I p (22)

- for PF algorithm the initial state vector is a Gaussian distribution vector with ~
x 0 ~ N (P x ,V x ) ;
- the measurements update sampling time of the Kalman filters coincides with the system sampling time (Ts);
- for the UKF algorithm the scaling parameters are set to the following values: D=1, E=2, kf=0;
- for PF algorithm the number of particle will be notated by N s and the resampling method is the systematic
resampling method.
To compare the performances of the three approaches we can calculate the cumulative square error for every case
(separately for each state). We can use the following relationship:
 Katalin György et al. / Procedia Technology 12 (2014) 65 – 74 71

1§N ·
Ei ¨ ¦ ( xi , j  ~
xi , j ) 2 ¸¸ i 1,..,n (23)
N ¨© j 1 ¹

where n is the number of states and N is the number of samples.

6.1. Nonlinear theoretic system:

For comparison of the three approaches consider the following nonlinear systems:

x1,k 0.5 ˜ x1,k 1  x2,k 1 ˜ u k 1  w1,k 1

(24)
x 2, k 0.05 ˜ x1,k 1 ˜ x2,k 1  u k 1  w2,k 1

The estimation algorithms were tested first with one linear output function, in this case is.

yk 2 ˜ x1,k  x2,k  vk (25)

and after that with one nonlinear output function:

yk  x1,k ˜ x2,k  vk (26)

For both case the initial state and the tuning parameters are: V x 10 ,V v 4 ,V w 0.1 , Px=0 and Ns=300,
and the input is a variable step signal. The real measured output with noise and without noise and the estimated
outputs using the EKF, UKF and PF algorithms, when the output is calculate with linear function (25) are shown in
Fig.1 a. The cumulative square presented in Fig.1.b. The real states (non-measurable values with noises and without
noises) and the estimated states using the EKF, UKF and PF algorithms when the output is calculate with nonlinear
function (26) are shown in Fig.2, the cumulative square error Fig.3. We can observe the essential error value in case
of EKF algorithm.

Variation of y Cumulative estimation errors for y

60 4

50 3.5

40
3

30
Error amplitude of y

2.5
Amplitude

20 Measured y with noise

y without noise 2
10 Estimated y (EKF)
Estimated y (UKF) 1.5
0 Estimated y (PF)

1
-10

-20 0.5

-30 0
0 10 20 30 40 50 60 70 80 90 100 1 2 3
No.of samples 1-EKF 2-UKF 3-PF

a) b)

Fig.1. The measured outputs with noises and without noises and the estimated outputs using the EKF, UKF and PF algorithms (a), the cumulative
square error (b) when the output function is linear
72 Katalin György et al. / Procedia Technology 12 (2014) 65 – 74

Variation of x1 Variation of x2
30 10

x1 with noise
"Ideal x1" without noise x1 with noise
Estimated x1 (EKF) 8 "Ideal x2" without noise
25
Estimated x1 (UKF) Estimated x2 (EKF)
Estimated x1 (PF) Estimated x2 (UKF)
6 Estimated x2 (PF)
20

4
15

2
Amplitude

Amplitude
10

5
-2

0
-4

-5
-6

-10 -8
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
No.of samples No.of samples

a) b)

Fig.2 Estimated states x1 (a) and x2 (b) using EKF, UKF and PF algorithms for the when the output function is nonlinear

Cumulative estimation errors for x1 Cumulative estimation errors for x2

50 2.5

45

40 2

35
Error amplitude of x1

Error amplitude of x2

30 1.5

25

20 1

15

10 0.5

0 0
1 2 3 1 2 3
1-EKF 2-UKF 3-PF 1-EKF 2-UKF 3-PF

a) b)

Fig.3. The cumulative square error for x1 (b) and for x2 (b) when the output function is nonlinear

6.2. Reactive sputtering process

In this section we try to use the presented estimation algorithms for one complex analytic nonlinear system
model, what characterized the reactive sputtering process. A very sensitive aspect of the reactive sputtering process
is the dynamic equilibrium of the reactive gas inside the chamber and of the metal atoms which form the compound
with the reactive gas atoms on the surface of the substrate [9,10,11]. The dynamic model of the reactive magnetron
sputtering process is defined by the system of equations [12]:

dp N
k1 ˜ (qin  q p  (D tM ˜ FN ˜ (1  T t ) ˜ At  D cM ˜ FN ˜ (1  T c ) ˜ Ac )
dt
dT t 1
(2 ˜ D tM ˜ FN ˜ (1  T t )  J ˜K N ˜ T t ) (27)
dt N Ti
dT c 1 §A · §A ·
( J ˜K N ˜ T t ˜ ¨¨ t ¸ ˜ (1  T c )  2 ˜ D cM ˜ FN ˜ (1  T c )  J ˜K M ˜ (1  T t ) ˜ ¨ t
¸ ¨A
¸ ˜T c )
¸
dt N Ti © Ac ¹ © c ¹

In this mathematical model the following notation and numerical parameters values has been used:
pN- the partial pressure of reactive gas (nitrogen) in the sputtering chamber;
Tt- the fractional surface of the target covered by compound molecules;
 Katalin György et al. / Procedia Technology 12 (2014) 65 – 74 73

Tc- the fraction of the condensation area covered by compound molecules;

FN- the flux of reactive gas molecules (N2) on the target or on the substrate (FN=k1 pN);
qin- the input reactive gas flow;
qp- the gas flow evacuated by the vacuum pump;
At, Ac- the target area (0.84 1e-2 m2) and the substrate (condensation) area (0.22 m2);
mN, mTi- mass of the reactive gas molecule (28 a.u) and of the metal (47.9 a.u.) (1.a.u.=1.66 1e-27 kg);
ηM, ηM sputtering yield of the elemental (Ti) metallic (1.5) and of the compound (titanium nitride) material (0.3);
DtM, DcM- sticking coefficients for the nitrogen molecule (to the titanium target or to the covered part);
NTi- the superficial density of the titanium atoms on the surface of the metallic target (140 1e-12 m-2 );
J- the particle density of argon ions on the surface of the target, which can be calculated using the relationship
J=Id/(At e), where e is the charge of electron (1.6 1e-19 C) and Id is the intensity of the discharge current;
k1- proportional coefficient, calculated in function of temperature and chamber volume.
R – ideal gas constant (8314 J/molK);
NA – Avogadro’s number (6.022 1e23 mol-1);
T – temperature (300K);
V – volume of the sputtering chamber (75 1e-3 m3).
This mathematical model in state space representation has three state variables (x1=pN, x2=Tt and x3=Tc) two
input signals (u1=qin and u2=Id) and for the output signal we can choose between the fractional surface of the target
Tt, the fractional surface of the condensation area Tc covered by compound molecules, but we can also chose the
pulverization rate (Rp) or the speed of deposition (aD) what can be calculate with following nonlinear relationships.

mTi ˜ R p ˜ At
Rp J ˜ (K N ˜ T t  K M ˜ (1  T t )) aD ˜[ (28)
UTi ˜ Ac

where UTi is the metal density (4.51 1e3 kg/m3 ) and [ is one accommodation coefficient (0.8).
The reactive sputtering process has been simulated employing a Runge-Kutta step control algorithm. The
sampling time was typically set to 0.001 sec, and the initial states vector:[4e-3 0.09 0.37]. For this case the
numerical simulation has following parameters:
V x [1e  4 1e  2 1e  2]; V w [1e  5 1e  2 1e  3]; V v 1e  2; P x x0 , N s 100
The first input signal (qin) is one step variable signal and the second input (Id) is constant. First the output is set y=Tc
and at the second case the output is y=aD .Simulation results for the first case are presented in Fig.4.
Variation of y -3
0.42 x 10 Cumulative estimation errors for y
2.5
Measured y with noise
Estimated y (EKF)
0.41
Estimated y (UKF)
Estimated y (PF) 2
0.4
Error amplitude of y

0.39 1.5
c
teta

0.38
1

0.37

0.5
0.36

0.35 0
0 50 100 150 200 250 1 2 3
No.of samples 1-EKF 2-UKF 3-PF

a) b)

Fig.4. The measured and estimated outputs with EKF, UKF and PKF algorithms (a) and the cumulative errors (b) when y=x3

When the output is the aD, that is mean that the output function is also nonlinear, the EKF results is worst but
results in case of PF and UKF are more better (Fig.5).
74 Katalin György et al. / Procedia Technology 12 (2014) 65 – 74

Variation of y
14 Cumulative estimation errors for y
30

12 Measured y with noise

Estimated y (EKF)
25
Estimated y (UKF)
10
Estimated y (PF)
20
8

Error amplitude of y
aD[Ao/sec]

6 15

4
10
2

5
0

-2 0
0 50 100 150 200 250 1 2 3
No.of samples 1-EKF 2-UKF 3-PF

a) b)

Fig.5. The measured and estimated outputs with EKF, UKF and PKF algorithms (a) and the cumulative errors (b) when y=aD

7. Conclusion

This paper shows that the UKF and the PF algorithms for state estimation is an interesting alternative to the EKF
because they have some improved performances. The PF and UKF methods are simpler to implement compared to
EKF. The computational load of UKF is comparable to the EKF approach, where the Jacobians are computed
numerically in each step of the algorithm. The performance of the PF depends of the number of particles. If this
number is great then the calculation time is also significant in correlation with the others methods.

References

[1] Riberio MI. Kalman and Extended Kalman Filters, Concept, Derivation and Properties, Institute for Systems and Robotics, Instituto
Superior Tecnico Lisboa ,2004. Available: http://users.isr.ist.utl.pt/~mir/pub/kalman.pdf.
[2] Terejanu GA. Unscented Kalman Filter Tutorial, Department of Computer Science and Engineering University at Buffalo, Buffalo, NY.
Available: http://users.ices.utexas.edu/~terejanu /files/tutorialUKF.pdf.
[3] Van der Merwe R, Wan EA. The square-root unscented Kalman filter for state and parameter estimation, Oregon Graduate Institute of
Science and Technology, Oregon, 2000. Available: http://www.researchgate.net/publicati on/2895916_The_Square-Root_Unscented_Kal
man_Filter_for_State_and_Parameter-Estimtion.
[4] Wan EA, van der Merwe R. The Unscented Kalman filter. Available: http://stomach.v2.nl/Docs/TechPubs/Tracking_and_AR
/wan01unscented.pdf.
[5] Kandepu R, Foss B, Imsland L. Applying the unscented Kalman filter for nonlinear state estimation, Science Direct, Journal of Process
control, 2008 [Online]. Available: http://www.itk.ntnu.no/ansatte/Foss_Bjarne/pubs/journals/journal-47.pdf.
[6] Sanjeev Arulampalam M, Maskell S, Gordon N, Clapp T. A Tutorial on Particle filters for Oline Nonlinear/Non-Gaussian Bayesian
Tracking, IEEE Transactions on signal Processing, 2002; 50(2): 174-188.
[7] Alfonso Manya. Particle Filter and Extended Kalman Filter for Nonlinear Estimation: A comparative Study (2008), Available:
http://users.isr.ist.utl.pt/~pjcro/cadeiras/dsfps0708/SEMS/PF_EKF_TermPaper.pdf.
[8] Doucet A, Johansen A. A tutorial on Particle Filtering and Smoothing, 2008. Available: http://www.cs.ubc.ca/~arnaud/
doucet_johansen_tutorialPF.pdf.
[9] Berg S, Blom HO, Larsson T, Nender C. Modeling of reactive sputtering of compound materials, J.Vac. Sci. Technol., 1987; A5 (2): 202-
207.
[10] Berg S, Nyberg T. Fundamental understanding and modeling of reactive sputtering processes, Science direct, Thin Solid films, 2005; 476
(2): 215-230.
[11] György K, Kelemen A, Papp S. Modeling and stability analysis of the nonlinear reactive sputtering process, Scientific Bulletin of the „Petru
Maior” University of Tg. Mureş, 2011; 8 (2): 33-37.
[12] Papp S, György K, Kelemen A, Jakab-Farkas L. Applying the Extended and Unscented Kalman Filters for Nonlinear State Estimation,
Inter-Eng 2012 Conference Proceedings, Tg Mureş, 2012: 233-239.

View publication stats