Yasuhiro Kawakami, Yasuhito Eguchi, Tsuyoshi Nimiya, Haruo Suemitsu and Takami Matsuo

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

Int. J. Advanced Mechatronic Systems, Vol. 5, No.

2, 2013

113

Velocity and acceleration estimation by iterative


learning observer and performance validation with
MEMS-based inertial sensors
Yasuhiro Kawakami, Yasuhito Eguchi, Tsuyoshi Nimiya,
Haruo Suemitsu and Takami Matsuo*
Department of Architecture and Mechatronics,
Oita University,
700 Dannonaru, Oita, Japan,
Fax: +81-975547507
E-mail: shigaku yk@yahoo.co.jp
E-mail: v12e6004@oita-u.ac.jp
E-mail: nimiyatsuyoshi@gmail.com
E-mail: suemitu@oita-u.ac.jp
E-mail: matsuo@oita-u.ac.jp
*Corresponding author
Abstract: We have recently proposed the adaptive velocity estimator based on the adaptive
control theory. In this paper, another velocity estimator is designed using the iterative learning
observer proposed by Chen et al. Since the ILO requires the passivity condition, we cannot apply
the ILO to the acceleration estimator directly. The series connection of the velocity estimator is
used as an acceleration estimator. Moreover, we perform the experimental validation of proposed
estimators by using the measurement data with the Crossbow NAV440 that is a product of
MEMSIC, Inc. Using Crossbow NAV440, we can obtain the roll, pitch, and yaw angles and their
angular velocities. The performances of the proposed differential estimators are verified by using
the measurement data with NAV440. The proposed velocity and acceleration estimators are
compared with the approximate differentiator and the exact differentiator.
Keywords: adaptive observer; parameter update law; differentiator.
Reference to this paper should be made as follows: Kawakami, Y., Eguchi, Y., Nimiya, T.,
Suemitsu, H. and Matsuo, T. (2013) Velocity and acceleration estimation by iterative learning
observer and performance validation with MEMS-based inertial sensors, Int. J. Advanced
Mechatronic Systems, Vol. 5, No. 2, pp.113121.
Biographical notes: Yasuhiro Kawakami received his Master degree in Engineering from Oita
University, Oita, Japan in 2013. He joined KYOCERA Corporation, Inc. He has worked in
adaptive estimation theory and its applications.
Yasuhito Eguchi received his Bachelor degree in Engineering from Oita University, Oita, Japan
in 2012. He is currently a Master course student at Graduate School of Engineering, Oita
University, Japan. He has worked in adaptive control system theory and its applications.
Tsuyoshi Nimiya received his Master degree in Engineering from Oita University, Oita, Japan in
2012. He joined Kawasumi Laboratories, Inc. He has worked in control system theory and its
applications.
Haruo Suemitsu received his Bachelor degree in Engineering from Oita Institute of Technology,
Oita, Japan in 1978. He is currently an Assistant Professor in the Department of Architecture and
Mechatronics, Oita University, Japan. He has worked in control system theory and its
applications.
Takami Matsuo received his Doctoral degree in Engineering from Kyushu University, Fukuoka,
Japan in 1985. He is currently a Professor in the Department of Architecture and Mechatronics,
Oita University, Japan. He has worked in control system theory and its applications. He is a
member of IEEE and SIAM.

Copyright 2013 Inderscience Enterprises Ltd.

114

Y Kawakami et al.
This paper
of velocity
Conference
Agriculture

is a revised and expanded version of a paper entitled Performance comparison


estimators with MEMS-based inertial sensors presented at 2012 International
on Advanced Mechatronic Systems (ICAMechS 2012), Tokyo University of
and Technology, Japan, September 1821, 2012.

1 Introduction
The necessity to evaluate the time-derivative of signals
arises frequently in many areas of research. For
position/speed controller, good measurements of position,
velocity, and frequency are required. Nevertheless, in
practical implementation, only the position is measured
by position sensors such as an optical encoder. Hence,
the velocity should be estimated in some ways with the
position measurement. Moreover, in tracking or detecting
moving objects in an image sequence, the need of velocity
estimation from measured position data is still a difficult
task and a challenging problem. The simplest numerical
method for differentiating a signal is the backward
difference. However, the backward difference operator
has a noise-amplifying characteristic (Tille and Montanari,
2001).
Many approaches to estimate the derivative of noisy
signals have been proposed. The use of low-pass
filters causes additional phase delay that decreases the
performance of the closed loop. Assuming that the position
can be approximated with a low-degree polynomial, a
differentiator based on the Newton predictor was proposed
(Tille and Montanari, 2001). The extended Kalman filter
base estimation approach can be used under the assumption
that the velocity is expressed as one of the state variables
generated by passing white noise through a linear and
stable all-integrator model (Belanger et al., 1998). Ibrir
(2004) presented a time-varying linear observer to estimate
the first (n 1)th derivative of any bounded signal. The
time-derivative observer was formulated as a high-gain
observer where the observer gain was calculated through
a Lyapunov-like dynamical equation (Ibrir, 2004). Levant
(1998) also presented a differentiator that is exact on
signals with a given upper bound for Lipshitzs constant
of the derivatives via the sliding mode technique. The
exact differentiator does not need a mathematical model
of the signal and provides for exact finite-time-convergent
estimation for noise-free signals. However, in the presence
of considerable measurement noises, the implementation
of a simple smoothing element is recommended (Levant,
1998).
We have recently proposed the adaptive velocity,
acceleration, and frequency estimators based on the
adaptive control theory (Matsuo et al., 2008a, 2008b).
Especially, in Nomura et al. (2010) and Kitsuka et al.
(2010a, 2010b), we presented two types of adaptive
differentiators to estimate the first and second order
derivatives of an output signal. In the case of the output
signal with unknown dynamic structure, a velocity estimator
was proposed based on the gradient-type adaptive update
law. Then, a non-passifiable adaptive scheme was proposed
using the high-gain input observer (Fomichev, 1998) in

order to design an acceleration estimator. The stability


of the error system with non-passifiable adaptive update
law was guaranteed by an LMI condition. This adaptive
scheme is more simple than the backstepping method
(Michino and Mizumoto, 2011). We call these estimators
the non-model-based adaptive differentiator.
In this paper, we summarise the non-model-based
differentiator and show the robustness of the velocity
estimator against the high frequency signals. Since the
approximated transfer function of the velocity estimator
is a kind of a band-pass filter, the estimator allows us
to reduce the noises and high frequency vibrations. Since
the proposed velocity estimator includes the discontinuous
function such as the signum function, the continuous
velocity estimator is preferable when applying to real
noisy data. From this viewpoint, we design another
velocity estimator using the iterative learning observer
(ILO) proposed by Chen and Chowdhury (2007). However,
the acceleration estimator cannot be designed by using the
ILO, because the ILO requires a kind of the passivity
condition. Thus, we use a series connection of the velocity
estimator based on the ILO as an acceleration estimator.
Moreover, we perform the experimental validation of
proposed estimators with the Crossbow NAV440 that is
a registered trademark of MEMSIC, Inc. (MEMSICs
NAV440, 2013). The Crossbow NAV440 is an integrated
GPS and attitude and heading reference system (AHRS)
that utilises low drift MEMS-based inertial sensors with
GPS aiding to provide an unmatched price and performance
(MEMSICs NAV440, 2013). The core of the NAV440 is a
rugged 6-degrees of freedom (DOF) MEMS inertial sensor
cluster. The 6-DOF MEMS inertial sensor cluster includes
three axes of MEMS angular rate sensing and three axes
of MEMS linear acceleration sensing. We can obtain the
roll, pitch, and yaw angles and their angular velocities.
The performance of the proposed differential filters is
verified by using the measurement data with NAV440. In
this paper, we adopt the yaw angle as the measurement
data. Moreover, we compare the proposed method with
the approximate differentiator and the exact differentiator
proposed by Levant.

2 Adaptive velocity estimator


2.1 Estimation law of velocity
Let x(t) be a position signal at a time t. We define 1 (t)
as the derivative with respect to time of the measurement
signal x(t), i.e.,
dx(t)
= 1 (t).
dt

(1)

Velocity and acceleration estimation by iterative learning observer and performance validation
By estimating 1 (t) with the measurement signal x(t),
we obtain the estimate of the velocity signal. We apply
the adaptive observer with the parameter adjustment law
to estimate the parameter 1 (t) under the assumption the
velocity is slow time-varying. Since the relative degree
of the system (1) is one, we can adopt the parameter
adjustment law with = 0, i.e., the adaptive law is
equivalent to the gradient-type law.
The time-varying parameter, 1 (t), is written by the
following equation:
1 (t) = 10 + (t),

(2)

where 10 is an unknown constant, and (t) is assumed that


its upper bound is known, i.e.,
|(t)| 0 ,

(3)

where 0 is a known constant.


We have proposed an adaptive observer to estimate
the time-derivative whose upper bound is known (Nomura
et al., 2010; Kitsuka et al., 2010a, 2010b). The estimate,

x(t),
of the derivative of the signal x(t) is given by the
following adaptive observer and the update laws:
x
(t) = k(
x(t) x(t)) + 1 (t) (t)sgn(
x(t) x(t))

1 (t) = 1 (
x(t) x(t))
(t) = |
x(t) x(t)|
where 1 > 0, k > 0. The estimate of x(t)

is given by

x(t)
= 1 (t) =

1 (
x( ) x( ))d

(4)

We call this estimator the adaptive velocity estimator.

2.2 Robustness against additive noises


Define the estimation errors of the state x(t) and the
parameter 1 (t) as
= 1 (t) 1 (t).
e(t) = x
(t) x(t), (t)
It is assumed that the available signal y(t) is the state x(t)
with the additional noise n(t) as

115

To derive the approximated transfer function of the adaptive


velocity estimator, we ignore the terms (t)sgn(e n) and
(t), which is equivalent to the adaptive velocity estimator
for the constant parameter . The error dynamics is given
by

e = k(e n) + (t)
and its Laplace transform is
E(s) =

{(s)
+ kN (s)}
s+k

where
E(s) = L{e(t)}, (s)
= L{(t)},
and
N (s) = L{n(t)}. When the additional noise is added to
the output, the parameter update law is given by

= 1 (e(t) n(t)).
The Laplace transform of the parameter update law is
obtained by

s(s)
= 1 (E(s) N (s)).
We have the transfer function form N (s) to the estimation

error of the parameter (s)


as

(s)
=

1 s
N (s).
s2 + ks + 1

The obtained transfer function is a kind of bandpass filters.


Thus, the adaptive velocity estimator allows us to attenuate
the noises and high frequency vibrations.
Remark 2.1: Levant (1998) also presented a differentiator
that is exact on signals with a given upper bound for
Lipshitzs constant of the derivatives via the sliding
mode technique. The exact differentiator does not need a
mathematical model of the signal and provides for exact
finite-time-convergent estimation for noise-free signals.
The exact differentiator of the signal x(t) proposed by
Levant (1998) is given as
z(t)
= u(t)

u(t) = u1 (t) |z(t) x(t)|sgn(z(t) x(t))


u 1 = l sgn(z(t) x(t)).

(5)
(6)
(7)

y(t) = x(t) + n(t).

3 Velocity estimator based on ILO


The adaptive velocity estimator with the available signal
y(t) is given by
x
= k(
x y) + 1 (t) (t)sgn(
x y)

1 (t) = 1 (
x(t) y(t)), (t) = |
x(t) y(t)|.
We have the error dynamics as
(t)sgn(e n) (t).
e = k(e n) + (t)

3.1 Iterative learing observer


In this subsection, we review the ILO structure (Chen and
Chowdhury, 2007). Chen and Chowdhury (2007) proposed
the ILO to alleviate the constraint of identifying constant
and slowly varying parameters by adaptive observers. The
design principle is to employ an additional term for the
purpose of identifying time-varying parameters that may
be periodic or aperiodic signals while estimating system

116

Y Kawakami et al.

states. The capability of identifying time-varying parameters


relies on the additional term that is updated in real time
by previous information (Chen and Chowdhury, 2007).
The updating mechanism used in the ILO for parameter
estimation is called learning (Chen and Chowdhury, 2007).
Consider a linear system described by
x(t)

= Ax(t) + Bu(t) + F (t)


y(t) = Cx(t)

}
(8)

where x(t) Rn , y(t) Rp , u(t) Rm , Rq is the


unknown time-varying parameter that may be constant,
periodic, or aperiodic signal, and A Rnn , B Rnm ,
C Rpn , F Rnq are constant matrices. The following
assumptions are made:

Theorem 3.1: Consider the system (8) and the ILO (9). If
there exists a positive definite matrix P satisfying
(A LC)T P + P (A LC) = Q < 0
and the ILO parameters, K1 and K2 , can be selected such
that
0 < K1T K1 I
P F = (K2 C)T
> > 1,
then both the state estimation error and the parameter
estimation error are bounded.
Remark 3.1: In this paper, we point out that the system (8)
has to satisfy the strictly positive realness condition. The
sufficient condition to construct the ILO is the existence of
the matrix P > 0 such that

the pair {A, C} is detectable

the time-varying parameter, (t), may vary quickly

the derivative of (t) may not exist at some time


instants

(A LC)T P + P (A LC) = Q

(11)

the time-varying parameter,(t), is bounded.

P F = C T K2T .

(12)

The parameter update laws of the conventional adaptive


observers use the integral operators. In contrast, the ILO
employs a delay operator in composing learning rule. The
ILO is given by the following equations:

x
(t) = A
x(t) + Bu(t) + L(y(t) y(t)) + F v(t)
(9)
y(t) = C x
(t)

v(t) = K1 v(t ) + K2 (y(t) y(t))


where x
and y are the estimated state and output. The
matrix L is selected such that (A LC) is stable. The
signal v(t) is the estimate of the time-varying parameter,
(t). It is updated online by both its previous information,
v(t ), and output estimation errors, y(t) y(t). The
parameter, > 0, is the updating interval. A better choice
of the parameter, , is the sampling interval. K1 Rqq ,
K2 Rqp are the gain matrices.
The error equation is obtained by
x
(t) = (A LC)
x(t) + F ((t) v(t))
y(t) = C x
(t)

}
(10)

where x
(t) = x(t) x
(t), y(t) = y(t) y(t).
Comparing with integrator-based adaptive observers,
Chen asserted that it possesses the following advantages
(Chen and Chowdhury, 2007):
1

it can deal with time-varying parameters without


existence of derivatives at some time instants

the estimated parameters may vary quickly

it does not require a persistently exciting signal.

Chen and Chowdhury (2007) have proven the following


theorem.

This condition is equivalent to the strictly positive realness


of the system (A, F, C):
x = Ax + F u
y = Cx.
Remark 3.2: The parameter estimation law (9) is equivalent
to a discrete-time version of the passivity-based adaptive
update law when K1 = 1. When 0 < K1 < 1, the parameter
estimation law (9) includes a negative delayed feedback
term.

3.2 Simulation results


3.2.1 Velocity estimation
To construct the velocity estimator based on the ILO,
consider the following system:
dx(t)
= 1 (t)
dt

(13)

where x(t) is the measurement signal and 1 (t) is the


unknown time-varying parameter. We can select the system
parameters in (8) as follows:
A = 0, B = 0, C = 1, F = 1.
We give a simulation example. Selecting the initial
condition and the unknown time-varying parameter as
x(0) = 0, 1 (t) = sin 5t, the measurement signal is obtained
by x(t) = 51 (1 cos 5t). Using the measurement signal,
x(t), the ILO obtains the estimates x
and (t) that are
shown in Figures 1 and 2, respectively. The parameters
of the ILO are selected as L = 1, K1 = 1, K2 = 5, Q = 1,
= 0.01.

Velocity and acceleration estimation by iterative learning observer and performance validation
Figure 1

d1 (t)
= 2 (t).
dt

The state x (solid blue line) and its estimate x

(dotted red line)

0.5
0.45
0.4

x and x

0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
0

10

time

Figure 2

117

The time-varying parameter (solid blue line) and


its estimate v(t) (dotted red line)

1.5

(16)

For each equation, the velocity estimator can be designed.


From equation (15) and from equation (16), we can
estimate 1 (t) and 2 (t) using the velocity estimator
based on the first-order ILO, respectively. The estimate
of the acceleration is given by 2 (t). However, we
cannot prove the asymptotic stability of the whole system.
Figure 3 shows the estimates of the time-varying parameter
1 (t) = sin 5t, when the parameters of the velocity
estimator based on the ILO are selected as L = 0.2,
K1 = 1, K2 = 12.5, Q = 10, = 0.001. The estimation
error is too large. Figure 4 shows the estimates of the
time-varying parameter, when the parameters of the velocity
estimator based on the ILO are selected as L = 0.2,
K1 = 0.8, K2 = 12.5, Q = 10, = 0.001. The estimation
error becomes small.
Figure 3 The time-varying parameter (solid blue line) and
its estimate v(t) (dotted red line) by the acceleration
estimator based on the first-order ILO with K1 = 1
6

x 10

4
3
2

and

and

0.5

0.5

1
0
1
2
3

1.5
0

10

time

5
0

3.2.2 Acceleration estimation

10

time

In this section, we apply the ILO to acceleration estimation.


We define 2 (t) as the second order derivative with respect
to time of the measurement signal, i.e.,
d2 x(t)
= 2 (t).
dt2

Figure 4 The time-varying parameter (solid blue line) and


its estimate v(t) (dotted red line) by the series
connection of the velocity estimator based on the
first-order ILO with K1 = 0.8

(14)
1.5

We can select the system parameters in (8) as follows:


[

[ ]
01
0
A=
,B =
00
0
[ ]
[ ]
0
C = 10 ,F =
1

and

0.5

However, the ILO for the above second-order system cannot


guarantee the stability, because CF = 0.
As a last resort, we use a series connection of the
velocity estimator based on the first-order ILO as an
acceleration estimator. Equation (14) can be rewritten by

0
0.5
1
1.5
0

6
time

dx(t)
= 1 (t)
dt

(15)

10

118

Y Kawakami et al.

Figure 5

(a) The yaw angle, (b) the yaw angular velocity, and (c) the acceleration estimate
100

yaw[deg]

50

50

100

10

15

20

25

30

35

20

25

30

35

20

25

30

35

time[s]

(a)

yaw rate [deg/s]

200
100
0
100
200
300

10

15
time[s]

(b)
2000

yaw acc [deg/s ]

1500
1000
500
0
500
1000
1500

10

15
time[s]

(c)

4 Inertial measurement unit


The Crossbow NAV440 made by MEMSIC, Inc. is
an integrated GPS and AHRS that utilises low drift
MEMS-based inertial sensors with GPS aiding to provide
an unmatched price and performance (MEMSICs NAV440,
2013). The core of the NAV440 is a rugged 6-DOF
MEMS inertial sensor cluster. The NAV440 provides full
inertial data (angles, rates, accels) and GPS position, along
with inertially derived velocity that provides significant
improvement in stability and latency compared with
stand-alone GPS velocity measurements. The 6-DOF
MEMS inertial sensor cluster includes three axes of MEMS
angular rate sensing and three axes of MEMS linear
acceleration sensing. The performance is given as follows:

Position/velocity

Position accuracy 1 (m CEP): < 2.5

X,Y velocity accuracy (m/s rms): 0.4

Z velocity accuracy (m/s rms): 0.5

1PPS accuracy (ns): 50.

Attitude (angle)

Range: roll, pitch (deg): 180, 90

Accuracy (deg): < 0.2

Resolution (deg): 0.02.

Velocity and acceleration estimation by iterative learning observer and performance validation

Range: roll, pitch, yaw (deg/sec): 200

Bias stability in-run (deg/hr): < 10

Bias stability over temp (deg/sec): < 0.02

Resolution (deg/sec): < 0.02

Bandwidth (Hz): 25.

Acceleration

Input range: X/Y/Z (g): 4

Bias stability in-run (mg): < 1

Bias stability over temp (mg): 4

Resolution (mg): < 0.5

Bandwidth (Hz): 25.

Figure 6 The comparisons between the measurement data of


yaw angular velocity and its estimates,
(a) yaw angular velocity (blue solid line) and the
estimate of the adaptive velocity estimator
(red dotted line) (b) yaw angular velocity
(blue solid line) and the estimate of the velocity
estimator based on the ILO (red dotted line)
400

angular rate[deg/s]

Angular rate

200
0
200
400
0

10

15

20

25

30

35

20

25

30

35

time[s]

5 Experimental results

rate[deg/s]
angular angular
rate[deg/s]

119

(a)
400
200
0
200
400
0

10

15

time[s]

(b)

Figure 7 The comparisons between the measurement data of


yaw angular velocity and its estimates,
(a) yaw angular velocity (blue solid line) and the
estimate of the approximated differentiator
(red dotted line) (b) yaw angular velocity
(blue solid line) and the estimate of the exact
differentiator (red dotted line)
400

angular rate[deg/s]

We can obtain the roll, pitch, and yaw angles and their
angular velocities (rates) by NAV440 measurement. Since
the angular acceleration cannot be obtained, we use the
output of the adaptive velocity estimator using the angular
velocity instead of the angular acceleration measurement.
In this paper, we adopt the yaw angle and the
yaw angular velocity as the measurement data. Figure 5
shows the angle, the angular velocities of yaw obtained
by NAV440 measurement, and the acceleration estimate
obtained by the adaptive velocity estimator with the angular
velocity.
To verify the performance of the proposed differential
filter, we add the Gaussian noise N (0, 102 ) or N (0, 0.5)
to the measurement signal of the yaw angle.

200
0
200
400

5.1 Velocity estimation

10

15

20

25

30

35

20

25

30

35

time[s]

(a)
400

angular rate[deg/s]

Figure 6 shows the comparison between the adaptive


velocity estimator and the velocity estimator based on the
ILO in the presence of the additive noise N (0, 102 ).
The parameters of the adaptive velocity estimator (4) are
selected as k = 500, 1 = 20, 000. The parameters of the
velocity estimator based on the ILO (9) are = 104 ,
L = 1, K1 = 0.7, Q = 4, K2 = 20. Figure 7 shows the
comparison between the approximate differentiator and the
exact differentiator in the presence of the additive noise
s
and
N (0, 102 ). The approximate differentiator is 0.01s+1
the parameters of the
exact
differentiato
(Levant,
1998)

are selected as = 700, l = 770. From these figures,


the proposed velocity estimators can estimate the yaw
rate and its estimation performance is more robust against
noises than the approximate differentiator and the exact
differentiator.

200
0
200
400

10

15

time[s]

(b)

Table 1 shows the mean square errors of the angular


velocity estimates between 0 (s) and 30 (s) in the
presence of the additive noise N (0, 2 ). The values are
the means of three times running with different initial
seeds. In Table 1, AVE, AD, ED, and ILO means the
adaptive velocity estimator, the approximate differentiator,
the exact differentiator, and the velocity estimator based
on ILO, respectively. The parameters of the adaptive

Y Kawakami et al.

120

velocity estimator, the approximate differentiator and the


exact differentiator are same as in Figures 6 and 7.
Moreover, we compare the performance for different K1 s.
The parameters of the ILOs (ILO2, ILO2, and ILO3) are
selected as in Table 2. The estimation performance depends
on the selection of the parameters K1 and K2 . Before the
estimation, the parameter calibration is necessary according
to the noise level.
Table 1 The mean square errors of velocity estimators
2

AVE

ILO1

ILO2

ILO3

AD

ED

0.01
0.5

11.1
12.4

18.1
31.6

11.0
13.9

10.9
15.3

18.7
71.7

14.7
26.7

estimator (4) are selected as k = 500, 1 = 20, 000. The


parameters of the velocity estimator based on the
first-order ILO (9) are = 104 , L = 3, K1 = 0.9, Q = 20,
K2 = 3.333. Figure 9 shows the comparison between the
series connection of the approximate differentiator and that
of the exact differentiator in the presence of the additive
s
noise N (0, 102 ). The approximate differentiator is 0.1s+1
and the parameters ofthe exact differentiator (Levant, 1998)
are selected as = 700, l = 770. From these figures,
the proposed velocity estimators can estimate the yaw
rate and its estimation performance is more robust against
noises than the approximate differentiator and the exact
differentiator.
Figure 9

Table 2 The ILO parameters

Figure 8

ILO3

1
4

0.9
5

0.7
1

104

104

104

0.2
16

10
7

20
4

The comparisons between the measurement data of


yaw angular acceleration and its estimates,
(a) yaw angular acceleration (blue solid line) and the
estimate of the adaptive velocity estimator
(red dotted line) (b) yaw angular acceleration
(blue solid line) and the estimate of the velocity
estimator based on the ILO (red dotted line)

acceleration [rad/s2]

2000

1000
0
1000
2000
0

10

15

20

25

30

35

15
20
time[s]

25

30

35

time[s]

(a)
2000
1000
0
1000
2000
0

1000

10

(b)

0
1000
2000
0

10

15

20

25

30

35

20

25

30

35

time[s]
]

(a)
2000

acceleration [rad/s2]

2000

acceleration [rad/s2]

ILO2

K1
L

K2
Q

ILO1

acceleration [rad/s2]

Parameter

The comparisons between the measurement data of


yaw angular acceleration and its estimates,
(a) yaw angular acceleration (blue solid line) and the
estimate of the approximated differentiator
(red dotted line) (b) yaw angular acceleration
(blue solid line) and the estimate of the exact
differentiator (red dotted line)

1000
0
1000
2000
0

10

15

time[s]

(b)

5.2 Acceleration estimation


We use the series connections of the velocity estimators as
an acceleration estimator. Figure 8 shows the comparison
between the series connection of the adaptive velocity
estimator and that of the velocity estimator based on
the first-order ILO in the presence of the additive noise
N (0, 102 ). The parameters of the adaptive velocity

Table 3 shows the mean square errors of the angular


acceleration estimates between 0 (s) and 30 (s) in the
presence of the additive noise N (0, 2 ). The values are
the means of three times running with different initial
seeds. In Table 1, AVE, AD, ED, and ILO means the
adaptive velocity estimator, the approximate differentiator,
the exact differentiator, and the velocity estimator based on
ILO, respectively. The parameters of the adaptive velocity
estimator, the approximate differentiator and the exact
differentiator are same as in Figures 8 and 9. Moreover, we
compare the performance for different K1 s. The parameters
of the ILOs (ILO2,ILO2, and ILO3) are selected as in
Table 4. The estimation performance also depends on the
selection of the parameters K1 and K2 .

Velocity and acceleration estimation by iterative learning observer and performance validation
Table 3 The mean square errors of acceleration estimators
2

AVE

ILO1

ILO2

ILO3

AD

ED

0.01
0.5

232.3
261.0

500.8
500.9

212.4
246.3

222.6
275.7

289.5
297.8

378.6
392.1

Table 4 The ILO parameters


Parameter

K1
L

K2
Q

ILO1

ILO2

ILO3

1
5

0.9
3

0.7
1

104

104

104

0.01
0.1

3.333
20

8.333
5

6 Conclusions
We proposed a velocity estimator based on the ILO. The
experimental results using the inertial measurement unit
show that the proposed differentiators are robust against
additive noises.

Acknowledgements
This work is partially supported by the Grant-in-Aid for
Scientific Research (21656108), Japan Society for the
Promotion of Science.

References
Belanger, P.R., Dobrovolny, P., Helmy, A. and Zhang, X.
(1998) Estimation of angular velocity and acceleration from
shaft-encoder measurements, Int. J. Robot Res., Vol. 17,
No. 11, pp.12251233.
Chen, W. and Chowdhury, F.N. (2007) Simultaneous
identification of time-varying parameters and estimation
of system states using iterative learning observers,
International Journal of Systems Science, Vol. 38, No. 1,
pp.3945.

121

Fomichev, V. (1998) Some algorithms for inversion of


linear dynamic systems, Computational Mathematics and
Modeling, Vol. 9, No. 3, pp.260267.
Ibrir, S. (2004) Linear time-derivative trackers, Automatica,
Vol. 40, No. 3, pp.397405.
Kitsuka, Y., Nomura, T., Suemitsu, H. and Matsuo, T. (2010a)
Adaptive estimation of velocity and acceleration with high
gain input observer, Proc. of Int Conf. on Modelling,
Identification and Control, pp.436441.
Kitsuka, Y., Nimiya, T., Suemitsu, H. and Matsuo, T. (2010b)
Non-model-based velocity and acceleration estimators for a
suspension system with parallel connection of a hydraulic
actuator, Proc. of 2010 IEEE Multi-Conference on Systems
and Control, pp.549554.
Levant, A. (1998) Robust exact differentiation via sliding mode
technique, Automatica, Vol. 34, No. 3, pp.379384.
Matsuo, T., Wada, S. and Suemitsu, H. (2008a) Model-based
and non-model-based velocity estimators for mobile robots,
International Journal of Innovating Computing, Information
and Control, Vol. 4, No. 12, pp.31233133.
Matsuo, T., Adachi, K. and Suemitsu, H. (2008b) Frequency
estimation with an lmi-based adaptive update law,
International Journal of Advanced Mechatronic Systems,
Vol. 1, No. 2, pp.100107.
MEMSICs NAV440 (2013) [online]
http://www.memsic.com/products/inertial-systems/
inertial-systems.html (accessed 11 February 2013).
Michino, R. and Mizumoto, I. (2011) High gain adaptive output
feedback control of non-linear systems with a class of
control uncertainties, International Journal of Advanced
Mechatronic Systems, Vol. 3, No. 4, pp.312323.
Nomura, T., Kitsuka, Y. and Matsuo, T. (2010) Nonmodel-based
estimation for velocity and acceleration by adaptive
identification method, IEEJ Trans. on Electrical and
Electronic Engineering, Vol. 5, No. 3, pp.372374.
Sato, K. and Tsuruta, K. (2010) Adaptive friction compensation
for linear slider with adaptive differentiator, Preprints of the
5th IFAC Symposium on Mechatronic Systems, pp.467472.
Tille, A. and Montanari, M. (2001) A low-noise estimator
of angular speed and acceleration from shaft-encoder
measurements, ATKAAF, Vol. 42, Nos. 34, pp.169176.

You might also like