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

net/publication/3155367

Observability of error States in GPS/INS integration

Article in IEEE Transactions on Vehicular Technology · April 2005

DOI: 10.1109/TVT.2004.841540 · Source: IEEE Xplore

CITATIONS READS
202 2,256

5 authors, including:

Sinpyo Hong Ho Hwan Chun

Pusan National University Pusan National University
29 PUBLICATIONS 650 CITATIONS 215 PUBLICATIONS 3,866 CITATIONS

SEE PROFILE SEE PROFILE

Jason Speyer
University of California, Los Angeles
348 PUBLICATIONS 8,782 CITATIONS

SEE PROFILE

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

Master Thesis View project

All content following this page was uploaded by Sinpyo Hong on 07 May 2015.

The user has requested enhancement of the downloaded file.

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 54, NO. 2, MARCH 2005 731

Observability of Error States in GPS/INS Integration

Sinpyo Hong, Man Hyung Lee, Senior Member, IEEE, Ho-Hwan Chun, Sun-Hong Kwon, and
Jason L. Speyer, Fellow, IEEE

Abstract—Observability properties of errors in an integrated between the GPS antenna and IMU is often quite difficult. The
navigation system are studied with a control-theoretic approach error in the estimated value of the lever arm, the relative posi-
in this paper. A navigation system with a low-grade inertial mea- tion of the GPS antenna with respect to IMU, can be of great
surement unit and an accurate single-antenna Global Positioning
System (GPS) measurement system is considered for observability concern in accurate navigation systems [1]. The lever arm error
analysis. Uncertainties in attitude, gyro bias, and GPS antenna in large vehicles can be much greater than the centimeter-level
lever arm were shown to determine unobservable errors in the error in CDGPS measurements. In addition, the lever arm error
position, velocity, and accelerometer bias. It was proved that all can increase errors in the estimates of position, attitude, and in-
the errors can be made observable by maneuvering. Acceleration ertial sensor biases [2], [3].
changes improve the estimates of attitude and gyro bias. Changes
in angular velocity enhance the lever arm estimate. However, both For a vehicle that moves with slow changes in attitude and
the motions of translation and constant angular velocity have no acceleration, the attitude of the vehicle is unobservable with a
influence on the estimation of the lever arm. A covariance simula- single-antenna GPS measurement system [4]. Moreover, as will
tion with an extended Kalman filter was performed to confirm the
be shown later in this paper, the component of gyro bias in the
observability analysis.
direction of specific force is not observable if the gyro has a large
Index Terms—Global Positioning System, GPS/INS, inertial
error. Thus, yaw error can increase significantly fast with time.
measurement unit (IMU), inertial navigation, lever arm, observ-
ability. These unobservable modes increase errors in the estimates of
position, velocity, and accelerometer bias. Especially, position
error can be considerable with large attitude error if the lever
I. INTRODUCTION arm is long.

T HIS paper considers the integration of an accurate single-

antenna Global Positioning System (GPS) measurement
system with a low-grade inertial measurement unit (IMU). The
One approach to dealing with the above problems is designing
an estimator for the uncertain values such as the lever arm, at-
titude, and biases of IMU. This approach can be particularly
positioning accuracy of the GPS measurement is assumed to be useful when there is no other instrument to measure the uncer-
centimeter level. This type of accuracy can be obtained from tain values. It is well known that motions of a vehicle can im-
carrier phase differential GPS (CDGPS). The errors in IMU are prove observability of the states of inertial navigation systems
assumed to be so large that gyro error is much greater than the (INS) and inertial sensor errors. The effect of maneuvering on
rotational speed of earth and accelerometer error is much greater the observability during in-flight alignment (IFA) was investi-
than the rotational speed of earth multiplied by the velocity es- gated in [5]–[9]. A control-theoretic approach to the observ-
timation error. Low-cost inertial sensors made with the current ability study on IFA was first introduced in [9] using a piece-
microelectromechanical systems (MEMS) technologies usually wise constant system modeling [10]. The research on the ob-
have these noise levels. servability enhancements was mainly concerned with the effect
There are a couple of important problems in this type of inte- of the translatory motions such as changes in acceleration. In
gration. Lever arm uncertainty can be a serious problem in ac- this paper, a control-theoretic framework was introduced for the
curate navigation systems. While the GPS antenna is mounted observability analysis of errors in the integrated systems of GPS
on the outside surface of a vehicle, IMU is usually placed in- and INS. Changes in attitude as well as acceleration were con-
side the vehicle. Thus, the direct measurement of the distance sidered in the observability analysis.
The errors in the observability analysis were described in the
Manuscript received March 3, 2004; revised July 20, 2004. This work was earth-centered earth-fixed (ECEF) frame and the body frame.
supported by the Korea Science and Engineering Foundation through the Ad- Errors in position, velocity, IMU attitude, biases in gyro and ac-
vanced Ship Engineering Research Center at Pusan Nation University. The re-
view of this paper was coordinated by Dr. R. Klukas. celerometer, and lever arm were considered in the observability
S. Hong is with the Advanced Ship Engineering Research Center, Pusan Na- analysis. The terms associated with position and velocity errors
tional University, Busan 609-735, Korea (e-mail: [email protected]). in the velocity error propagation equation were neglected be-
M. H. Lee is with the School of Mechanical Engineering, Pusan National
University, Busan 609-735, Korea. cause they were relatively very small compared with other error
H.-H. Chun is with the Department of Naval Architecture and Ocean Engi- terms in the integration of low-grade IMU and accurate GPS
neering, Pusan National University, Busan 609-735, Korea. measurements. This simplification made the observability anal-
S. H. Kwon is with the Department of Naval Architecture and Ocean Engi-
neering, Pusan National University, Busan 609-735, Korea. He is also presently ysis uncomplicated. Among the inertial sensor errors such as
a visiting scholar at the Department of Civil Engineering, Texas A&M Univer- biases, scale-factor errors, and alignment errors, biases are the
sity, College Station, TX 77843–3136 USA. most unpredictable and dominant in low-grade sensors. Since
J. L. Speyer is with the Mechanical and Aerospace Engineering Department,
University of California, Los Angeles, CA 90095-1597 USA. the period of testing for the error estimation is relatively short
Digital Object Identifier 10.1109/TVT.2004.841540 compared with the time-constant of bias drifts, the biases in in-
0018-9545/$20.00 © 2005 IEEE
732 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 54, NO. 2, MARCH 2005

ertial sensors ae considered to be constant in this paper. Time outside surface of the vehicle. Rapid changes in acceleration and
derivatives of GPS measurements were used for observability angular rate are desirable for the observability enhancement by
test. This approach changes the observability test on the 18 error maneuvering. Small vehicles are usually capable of the motion
states to that on at most nine error states; errors in attitude, gyro changes required for time-varying observability with their own
bias, and lever arm. Compared with the analysis based upon the power. Large vehicles may need external forces such as air wind,
null space test of observability matrices [3], [4], this approach water waves, or gravity for the rapid motion changes.
makes the observability analysis simpler. Many notations in this paper follow those in [14]. For a vector
In this paper it is shown that the time-invariant error dynamics , is the vector decomposed in a coordinate frame .
model for the integrated navigation system has seven unobserv- denotes the rotation matrix from a frame to a frame .
able modes when the position of a vehicle is measured with a denotes the column vector of the angular velocity of a frame
single antenna GPS measurement system. Errors in the lever relative to a frame , decomposed in a frame . denotes the
arm, attitude, and component of gyro bias in the direction of screw-symmetric cross-product matrix of . , , , and
gravity are not observable if the vehicle moves with constant denote the estimate, estimation error, time derivative, and
attitude and acceleration. Both the errors in attitude and lever transpose of a matrix , respectively. denotes the absolute
arm contribute to the position error. It was shown in previous re- value of a vector . and denote the cross-product
search that if the lever arm error is neglected, the time-invariant and dot (scalar) product of vectors and , respectively.
error dynamics model with two horizontal channels has three
denotes the th time derivative of a matrix . denotes an
unobservable modes [11], [12]. Later, in [4] and [9], the error
identity matrix. “0” denotes a zero matrix with an appro-
model for the three channels was also shown to have three un- priate dimension. , , , , and used for coordinate frames
observable modes. The addition of the unobservable gyro bias denote the earth-centered inertial (ECI) frame, ECEF frame,
in this paper for the three channels results from the omission of
earth-fixed tangential frame (east, north, up), body-fixed naviga-
the terms associated with the position and velocity errors in the
tion frame (north, east, down), and body frame (forward, right,
velocity error propagation equation. down), respectively.
It is shown in this paper that the unobservable errors in
the time-invariant system can be made observable through II. NAVIGATION ERROR MODEL
maneuvering of the vehicle. Based on the observability analysis
of time-varying systems, all of the above seven unobservable A navigation error propagation model is introduced in this
modes are shown to be made observable if both the directions of section. The errors in the estimates of position, velocity, attitude,
angular velocity and acceleration change. Acceleration changes biases in the inertial sensors, and lever arm from their true values
are shown to improve the estimation of attitude and gyro bias. are considered in the error propagation equations. Velocity error
The components of errors in attitude and gyro bias that are propagation equation is made simple by neglecting error terms
orthogonal to the direction of the acceleration change are made associated with errors in position and velocity. These error terms
observable. This result is consistent with that obtained from the are relatively small in the integration of low-grade inertial sen-
piecewise constant system modeling [9]. Changes in angular sors with accurate GPS measurements. The simplified velocity
velocity decrease the lever arm error. The components of the error propagation equation makes the observability analysis in
lever arm error that are orthogonal to the axis of rotation are Section III uncomplicated. Reference frames for error states are
made observable. The motion with a constant angular velocity chosen in such a way that the observability analysis in the next
does not have any effect on the estimation of the lever arm in section is more convenient. Inertial sensor biases and errors in
the GPS/INS systems in which low-grade inertial sensors are attitude and lever arm are represented in the body frame. Posi-
employed. Covariance simulation results were given to support tion and velocity vectors are represented in the ECEF frame.
the observability analysis results. The navigation equations in the ECEF frame are [14], [15]
One of the main contributions of this paper is a control-theo-
(1)
retic approach to the observability analysis on the time-varying
error dynamics model of GPS/INS. Even though it is recently (2)
known that rotational motions improve the estimation of the (3)
lever arm, rigorous research on the effect of rotational motions
on observability was rarely published [3], [13]. This paper ex- where , , and are the position, velocity, and gravity,
tends the application of control-theoretic approaches to more respectively, and is the specific force. Let the errors in the
general time-varying error models. With this approach, the ef- mechanization of the navigation equations be modeled as
fect of motions of rotation as well as translation on the estima- (4)
tion of errors in navigation systems can be studied. The second
(5)
contribution is that the relationship between the motions of the
vehicle and the observability of errors in the low-grade INS (6)
aided by a single antenna GPS measurement system is explicitly (7)
given. (8)
One useful application of the observability property is the
measurement of the relative distance between two objects, one where is the position error, is the velocity error,
of which is inside of a vehicle and the other of which is on the is the attitude error, is the cross-product matrix of ,
HONG et al.: OBSERVABILITY OF ERROR STATES IN GPS/INS INTEGRATION 733

is the accelerometer measurement, is the accelerometer bias,

is the accelerometer noise, is the gyro measurement,
is the gyro bias, and is the gyro noise. Attitude error in INS
error analysis has usually been represented in the navigation
frame such that where is the reference
frame of INS [14]–[16]. The relation between and is
for small . Relatively simple observability analysis will
be obtained with in the next section. The linearized error
propagation equations are [14], [15]

(9)
(10)
(11)
(12)
Fig. 1. GPS/INS measurement system.
(13)

where ; , , and are the cross-product servability properties are investigated by examining the prop-
matrices of , , and , respectively. The error propaga- erties of time derivatives of the GPS position measurement. In-
tions for and are related with each other such that vestigation of the time derivatives of the measurement changes
. the observability test on the 18 error states to that on at most
The state augmentation technique [11], [17], [18] for errors nine error states; errors in attitude, gyro bias, and lever arm.
in the inertial sensors is used in this paper. Since biases are the These error states are represented in the same reference frame,
major error sources in the low-grade inertial sensors [19], other the body frame. Comparing with the attitude error represented in
errors such as scale factor error and misalignment errors are ne- the navigation reference frame, the attitude error given the same
glected in this paper. As in [9] and [12], biases are assumed to reference frame of the inertial sensor biases and lever arm error
be constant because the test time for the error estimation is rel- makes time derivatives of measurement estimation error sim-
atively short compared with the time constant of the bias drift. pler. These nine error states together with the simplified velocity
The maximum singular value of is in the order of 10 . The error propagation model make the observability test straightfor-
magnitudes of and are in the orders of 10 and 10, re- ward. In addition, test of the time derivatives of the measurement
spectively. The magnitudes of , , and are in the orders makes the physical interpretation of the observability analysis
of 0.1, 0.2, and 0.01, respectively, in carrier-phase differential easy.
GPS. The magnitudes of and are assumed to be in the or-
ders of 0.1 and 0.001. Thus, in this paper, the gravity gradient A. Observability Definitions
and rotational motion of earth can be considered less important.
Instead of (10) and (11), the following equations are used in the Before the main part of this section is given, the definitions of
following sections to simplify the observability analysis: observability of linear systems used in this paper are introduced.
Consider the linear system
(14)
(15)

A single antenna GPS measurement system is given in Fig. 1

where is the lever arm between the GPS antenna and IMU.
where and are, respectively, the and ma-
Lever arm uncertainty is modeled such that
trices whose entries are continuous functions of time defined
(16) over ( , ).
Definition 3.1: The dynamic equation is observable at if
(17) there exists a finite such that for any state at time ,
where is the lever arm error in the body frame. Then, the GPS the knowledge of the output over the time interval
measurement estimation error can be written as [3], [13] suffices to determine the state .
Define a sequence of observability matrices
(18) by the equation

where is the cross-product matrix of the lever arm and

is the error in the GPS measurement.

III. OBSERVABILITY PROPERTIES

In this section, observability analyses are made for both the Suppose and in the system are analytic functions
time-invariant and time-varying error dynamics models. Ob- of . Then, the time-varying system is observable
734 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 54, NO. 2, MARCH 2005

at time if there exists a finite time such that the rank Then, the INS mechanization error equations and measurement
of the matrix estimation error can be written as

(24)
.. (19) (25)
.
where and include the first-order approximation
errors and sensor noises. For simplicity of expression, the
is [20]. Suppose and in the system are constant. symbol of dependency of matrices and vectors on time is
Then, the time-invariant linear system is observable if and only omitted in the rest of this section. For example, and
if the rank of the matrix are replaced with and , respectively. Suppose
is an unob-
servable mode of the system . Then

.. (26)
.
(27)
is . If the linear time-invariant system is observable, then it (28)
is observable at every initial time, and the determination of the
(29)
initial state can be achieved in any nonzero time interval [20].
The above observability tests are the same as finding a state (30)
vector such that ..
.

..
. (31)

where

for a time . If there is no nonzero state that satisfies (32)

the above conditions, then the system is observable at . If the with
system is time-invariant, the system’s observability does not de-
pend upon time. In the rest of this section, observability prop- (33)
erties of errors in the integration of GPS and INS are provided.
(34)
The proofs of them are given in Appendix A.
(35)
B. Observability Condition
In this section a general observability condition for the in- Since is a full-rank matrix, , in
tegrated system of GPS/INS is introduced. By the application (28)–(31) implies that
of observability condition to the time derivatives of GPS mea-
surement estimation error, an observability test on the 18 error (36)
states is transformed to that on at most nine error states. For the
with
analysis on the time-varying system, it is assumed in this paper
that and are 1 times continuously differentiable (37)
where is the number of error states. In this paper .
Let Let

(20) (38)
Since and are constant, and , we have
(21)
(39)
(40)
(22) (41)

where . Define

(23) (42)
HONG et al.: OBSERVABILITY OF ERROR STATES IN GPS/INS INTEGRATION 735

for with .
Thus (44) can be satisfied with arbitrary nonzero elements of ,
.. .. .. (43)
. . . , and the component of in the direction of . Therefore,
the system has seven unobservable modes. Let be the matrix
whose columns represent the unobservable modes such that
Then, the condition , is the same as
(47)
(44)

The remaining conditions , is always satisfied Then the results of the observability analysis is summarized with
with the following relations: the following property.
Property 3.1: The time-invariant system has seven
(45)
unobservable modes. The set of seven columns of is a basis
of the unobservable subspace.
(46) These unobservable modes induce errors in other estimates
such that
and (37). Thus, the observability of the system depends
on the rank of . If it has a full column rank, then the system (48)
is observable. Since the null space of in the nine-dimensional (49)
real vector space determines the unobservable subspace of the
(50)
system , a basis of the null space of will be given in
the following observability analysis. Remark 3.5: An important observability property of the inte-
Remark 3.1: in (32) is equal to grated navigation system of a low-grade IMU with a single-an-
. This is the tenna GPS measurement system is that the component of gyro
same as . This shows bias in the direction of the specific force is unobservable if
that is the sum of errors in the estimates of specific force a vehicle moves with constant attitude and acceleration. Con-
and tangential and centrifugal accelerations of antenna 1. These sider the case in which a vehicle moves on a straight horizontal
errors are induced by the attitude estimation error . in line with a constant velocity. Then, the yaw estimation error in-
(32) is the same as . Thus it is the creases as time passes. The error increase rate is proportional to
error in the estimate of the centrifugal acceleration of antenna the vertical component of the gyro bias estimation error. How-
1. This error is induced by the uncertainty in the gyro bias . ever, in the navigation system with an accurate IMU that can de-
is the same as . Thus it is tect the rotational motion of earth, the increase rate of the yaw
the sum of errors in the estimates of tangential acceleration and estimation error can be negligible because gyro bias is observ-
centrifugal acceleration. The estimation errors are caused by the able [4], [9].
lever arm error .
Remark 3.2: Equations (37) and (44)–(46) state that the un- D. Systems With Time-Varying Acceleration
observable modes are determined by the nine error states ,
This section investigates the effect of acceleration changes
, and that make all the time derivatives of the estimation
on the observability of GPS/INS systems. It is assumed that the
error of the measurement acceleration zero vectors. If there are
vehicle moves with a constant attitude. For the convenience of
no such nonzero states, then the system is observable.
expression, a group of matrices are introduced in the following.
Remark 3.3: Note that the reference frame of is the same
Each column in the matrices represents an unobservable mode
as that of and . It can be easily seen that if the attitude error
that satisfies (44) for various motion conditions given in Prop-
was represented with another reference frame, time derivatives
erties 3.2–3.6. Let
of would be more complicated.
Remark 3.4: From (33)–(35) and (39)–(41), it can be seen
that is determined by , , , and their time derivatives.
Thus, the observability properties depend on the specific force
and angular velocity of the vehicle. If the speed of the vehicle
is not very fast, the following relations hold:
. Thus, the observability properties of navigation errors are
expected to be independent of the reference frame of the INS
mechanization equations such as ECEF, local geographic frame,
(51)
or (local fixed) tangential frame.

C. Time-Invariant Systems where ; and are defined in the fol-

In this section observability property is presented for the time- lowing Properties 3.2 and 3.4, respectively. consists of
invariant system in which acceleration and attitude are constant. seven modes. Any three-dimensional with the constraint
Suppose and in (22) and (23) are constant. Then can be expressed as a linear combination of the
736 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 54, NO. 2, MARCH 2005

first three columns. The component of in the direction of Remark 3.6: Property 3.4 states that a change in acceleration
can be expressed with the fourth column. Any three-dimen- direction makes the components of attitude error that are per-
sional vector can be expressed as a linear combination of pendicular to the acceleration change observable except for a
last three columns. consists of five modes: It represents very rare case. This result is in agreement with the observability
three-dimensional and the components of and in the analysis made by the piecewise constant modeling in [9].
direction of . consists of five modes. The component of
in the direction of can be expressed with the first column. E. General Time-Varying Systems
The second column represents the component of in the Next, observability properties are presented for the case in
direction of with the constraint that is in the direction of which both the acceleration and attitude of a vehicle change. In
. The ratio of to is the same as that of to . The last the following, roll, pitch, and yaw are denoted by , , and ,
three columns represent three-dimensional . consists of respectively. As in the previous subsections a group of matrices
five modes. The first column represents the component of in are introduced in the following for Properties 3.7–3.12:
the direction of with the constraint that is in the direction
of . The ratio of to is the same of that
of to . The second column represents the
component of in the direction of . The last three columns
represent three-dimensional . consists of four modes:
It represents the component of in the direction of and (52)
three-dimensional . Observability properties for time-varying
accelerations are given below.
Property 3.2: Suppose that is constant and and where , ,
all the time derivatives of it have the same direction such , and is the position vector from the vehicle to the center of
rotation, represented in the body frame. represents a compo-
that , where nent of in the direction of . consists of six modes: Any
are real numbers. If three-dimensional with the constraint can be
, then the system represented with a linear combination of the first three columns.
has seven unobservable modes. The set of seven column The last three columns represent the three-dimensional .
vectors in is a basis of the unobservable subspace. Other- represents the component of in the direction of . rep-
wise, the system has five unobservable modes. The set resents the component of in the direction of . consists
of five column vectors in is a basis of the unobservable of two modes. The first column represents the component of
subspace. in the direction of . The second column represents the com-
Property 3.3: Suppose that is constant, that and ponent of in the direction of with the constraint that
have different directions, and that is constant. Then, the is in the direction of . The ratio of to is
system has five unobservable modes. The set of the five the same as that of to . Then the observability
column vectors in is a basis of the unobservable subspace. properties for the general time-varying system are given below.
Property 3.4: Suppose that is constant, that Property 3.7: If is constant, then the system has
and have different directions, and that all of the three unobservable modes . The set of the last three column
time derivatives of have the same direction such that vectors in is a basis of the unobservable subspace.
, Property 3.8: If and all orders of the time derivatives
where are real numbers. If of it have the same direction, then the system has an
unobservable mode. is a basis of the unobservable subspace.
, then the system has five unobservable Property 3.9: If and are constant, then the system
modes. The set of the five column vectors in is a basis of has six unobservable modes. The set of the six column
the unobservable subspace. Otherwise, the system has vectors in is a basis of the unobservable subspace.
four unobservable modes. The set of the four column vectors in Property 3.10: If is constant and both and are linear
is a basis of the unobservable subspace. functions of time, then the system has an unobservable
Property 3.5: Suppose that is constant and , , and mode. is a basis of the unobservable subspace.
are linearly independent. Then the system has three Property 3.11: If and both and are linear functions
unobservable modes . The set of the last three column vectors of time, then the system has an unobservable mode.
in is a basis of the unobservable subspace. is a basis of the unobservable subspace.
From Properties 3.2–3.5, the following property can be ob- Property 3.12: If a vehicle rotates on the horizontal plane
tained. with a constant radius and a constant angular acceleration
Property 3.6: Suppose that is constant and is time- , then the system has two unobservable modes. The
set of the two column vectors in is a basis of the unobserv-
varying. If any three vectors in are lin- able subspace.
early independent, then the system has three unobserv- The observability of the system depends upon the rank
able modes . Otherwise, has at least one additional un- of . However, is very complicated for general motions, and
observable mode. corresponding analytic observability conditions are not easy to
HONG et al.: OBSERVABILITY OF ERROR STATES IN GPS/INS INTEGRATION 737

TABLE I TABLE II
ACCELERATIONS AND ATTITUDES IN THE NUMERICAL OBSERVABILITY TESTS DESCRIPTIONS OF NUMERICAL TESTS

TABLE III
, , , DIMENSION OF THE NULL SPACE OF IN THE NUMERICAL TESTS
, ,
,
, .

obtain. In the rest of this section, results of a numerical observ-

ability test are given for the simple motions that can represent As can be seen in Table III, the lower bound of the numbers
many types of typical vehicle motions. Let of unobservable modes in each of the tests is the same as the
upper bound of the numbers of them for the same test. Thus, the
null space dimensions in Table III are actually the numbers of
unobservable modes of the tests. The numerical test on the ob-
.. .. .. (53) servability matrix shows that the lever arm error can be found
. . .
with rotational motion, not with translatory motion. Changes in
both the magnitude and direction of angular velocity make the
Note that in (43) is the same as . If has a full column system observable. The same numerical test results on the ob-
rank for a certain motion with , then also has a full servability were obtained for the following lever arms: [10 1
column rank. Thus the system is observable with that motion. 1] , [1 10 1] , and [1 1 10] .
Since , , and for are very complicated, is In the observability analysis given in this section, process and
chosen for the numerical test of observability in the following. measurement noises are neglected. Obviously, the presence of
To test the rank of , singular values of it were investigated. these noises can be harmful to the performance of state estima-
If the smallest singular value of is positive, then the matrix tors. In many cases, especially in the inertial navigation systems,
has a full column rank. The lever arm in the numerical tests was Kalman filtering techniques are often employed to evaluate the
. Table I shows accelerations and attitudes of the vehicle estimator behaviors in the presence of sensor noises. Error co-
in the numerical tests. The descriptions of the above tests are variance matrix in the Kalman filter can be used to inspect the
given in Table II. performance degradation of estimator due to the noises. Even
Unobservable modes in the above numerical tests were ob- though the effect of the noises is not considered in the observ-
tained between seconds 1 and 1 at 0.1 s intervals. The number ability analysis in this section, the analysis can be helpful in
of unobservable modes for each test was found to be constant understanding the limits of estimator performance. If a state is
for the time interval in the tests. The test results are given in unobservable, the state cannot be estimated even in the most fa-
Table III. vorable situation in which no noise is present in the system. If
Since the above observability test is based on the null space the unobservable mode is unstable, estimation error can grow
test on that is a part of the full observability matrix , the without limit.
dimensions of null spaces in Table III are actually upper bounds Low-cost MEMS inertial sensors usually have a lot of broad-
of the number of unobservable modes. The table shows that band noises. Thus, for the state estimator design, proper mod-
upper bounds of the numbers of unobservable modes in Tests eling of sensor noises is required. Allan variance technique [21]
7 and 8 are zeros. Since the number of unobservable modes is can be a useful means to investigate the time-domain character-
nonnegative, it can be inferred that all the errors are observable istics of random signals. Allan variance tests showed that noises
in Tests 7 and 8. On the other hand, the properties given in this in the low-cost MEMS inertial sensors can usually be modeled
section provide lower bounds of the numbers of unobservable as the sum of random constant bias and white noise for the time
modes for Tests 1 through 6. Properties 3.3, 3.5, 3.9, 3.10, 3.11, period of one or two minutes [22], [23]. For longer time periods,
and 3.12 state that the lower bounds of the numbers of unob- the noises can be modeled as exponentially correlated processes
servable modes in Tests 1–6 are 5, 3, 6, 1, 1, and 2, respectively. with time constant longer than 100 s. When the time required
738 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 54, NO. 2, MARCH 2005

for the maneuvering is less than or equal to 100 s, noise mod-

eling of random constant bias and white noise can be justified
for low-cost MEMS inertial sensors.

IV. SIMULATION RESULTS

The observability analysis given in the previous section does

not provide sufficient information on the performance of error
estimators. It only decides whether a measurement system is ob-
servable or not. For the analysis of the degree of observability
[24], covariance analysis is employed in this section. It can also
be considered as an efficient means of producing statistical in-
formation on the behavior of state estimators [25]. Assuming all
the errors in the process and sensor to be Gaussian white, numer-
ical simulation results on the standard deviation (STD) of errors
are provided in the following. In the simulation INS mechaniza-
tion, differential equations were integrated with low-grade IMU
data at 100 Hz. The solutions to the mechanization differential
equations were corrected with an extended Kalman filter at 1 Hz
using the measurement from the GPS antenna.
The statistics of sensor noises in the numerical simulation Fig. 2. Vehicle trajectory.
are as follows: all the noises in measurements from GPS and
IMU are assumed to be Gaussian white. The STD of the GPS
measurement noise in the tangential frame was [0.06 0.06 0.08]
in meters. Bias and the STD of the accelerometer noise were
[ 0.04 0.05 0.1] and 0.01 in m/s , respectively. Bias and the
STD of the gyro noise were [ 0.09 0.05 0.1] and 0.1 in ,
respectively. The lever arm of GPS antenna was [1.0 1.0 1.0]
in meters in the body frame.
The motion of the vehicle in the simulation is given in
Figs. 2–7. The vehicle was motionless for five seconds after
the start of the numerical simulation. It accelerated and moved
Fig. 3. Vehicle velocity.
northward with a constant velocity for five seconds, respec-
tively. The vehicle changed its roll angle for five seconds. Then,
it moved without changing its velocity and attitude for five
seconds. The vehicle moved up and down, changing its pitch
angle for five seconds. Once again, it moved without changing
its velocity and attitude for five seconds. Finally, it moved
eastward and back westward, changing its yaw angle for ten
seconds. Thus, the vehicle experienced a motion of a full six
degrees of freedom in the simulation.
Figs. 8 –12 show errors in the estimates of INS states. The
STDs of roll and pitch estimation errors in Fig. 8 decreased
significantly at the beginning. This is because the horizontal Fig. 4. Vehicle acceleration.
components of static unobservable attitude error can be approx-
imated to the accelerometer bias divided by gravity, which is In Fig. 10, the horizontal components of gyro bias error were
about from 0.3 to 0.4 degrees [13]. The figure shows that the continuously reduced from the beginning. The figure confirms
STD of yaw error decreased when the horizontal components that these errors are observable without maneuvering. The figure
of the specific force changed. The STD of yaw error during the also shows that the vertical component of gyro bias error re-
first 5 s given in Fig. 8 can be misleading. Even though the STD mained almost constant until the twenty-fifth second. It was re-
of yaw estimation error remained constant while the vehicle was duced by large amounts when the specific force and attitude
motionless in the beginning, the yaw estimation error in Fig. 9 changed simultaneously between seconds 25 through 30 and 35
was shown to increase in the beginning. This is the case when through 45. The figure shows that the change in the horizontal
the signs of the initial yaw error and gyro bias were the same. If acceleration is not effective in the estimation of the vertical com-
the sign of gyro bias had been different from that of the initial ponent of gyro bias.
yaw error, the absolute value of yaw error would have decreased In Fig. 12, it can be seen that both right-hand and down-
in the beginning. ward components of errors in the lever arm estimate began to
HONG et al.: OBSERVABILITY OF ERROR STATES IN GPS/INS INTEGRATION 739

Fig. 5. Vehicle specific force. Fig. 9. Error in the estimate of yaw angle.

Fig. 6. Vehicle attitude. Fig. 10. Standard deviation of gyro bias estimation error.

Fig. 7. Vehicle attitude rate. Fig. 11. Standard deviation of accelerometer bias estimation error.

Fig. 8. Standard deviation of attitude estimation error.

Fig. 12. Standard deviation of lever arm estimation error.

be reduced significantly after the vehicle began to experience a

change in roll rate at the fifteenth second. The figure also show V. CONCLUSIONS
that forward and downward components of the lever arm error In this paper the observability of errors in the integration of
were reduced noticeably when pitch angle started to change a low-grade IMU with an accurate single antenna GPS mea-
at the twenty-fifth second. Thus, Property 3.8 was numerically surement system was studied. A control-theoretic approach was
demonstrated. adopted for the observability analysis of a time-varying error
The simulation results showed that acceleration changes were dynamics model. The effects of the motions of both translation
effective to improve attitude estimation. They also showed that and rotation on the error estimation were investigated. Errors
changes in angular velocity reduced lever arm estimation error. in the estimates of position, velocity, attitude, biases of inertial
However, changes in both acceleration and attitude were neces- sensors, and GPS antenna lever arm from their true values were
sary for the estimation of the vertical component of gyro bias. considered in the observability analysis.
740 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 54, NO. 2, MARCH 2005

Velocity error propagation equation was made simple by ne- Thus, if

glecting error terms associated with errors in position and ve- , then (54)–(56) are satisfied with the
locity since these error terms are relatively small. The reference following three cases. First, for any value of
frames of errors were chosen in such a way that the observability with . Second, and is in the direction
analysis was less complicated. Inertial sensor biases and errors of . Third, and is arbitrary. Otherwise,
in attitude and lever arm were represented in the body frame. Po- the above equations are satisfied with the following three cases.
sition and velocity vectors are represented in the ECEF frame. First, and is in the direction of . The
Observability properties were investigated by examining the second and third cases are the same as the above second and
properties of the time derivatives of the GPS position measure- third cases. This completes the proof of Property 3.2.
ment. Analysis with the time derivatives of the measurement
changed the observability test on the 18 error states to that on B. Proof of Property 3.3
at most nine error states: errors in attitude, gyro bias, and lever From (44), it follows that
arm. These error states were represented in the same reference
frame, the body frame. These nine error states together with the (57)
simplified velocity error propagation model made the observ- (58)
ability test straightforward.
Observability analysis results showed that the time-invariant Obviously, the following three sets of , , and satisfy
error dynamics model has seven unobservable modes: attitude, (57) and (58); First, with in the direction of
lever arm, and the component of gyro bias in the direction of the . Second, with and where
specific force. Due to the unobservable gyro bias, yaw error can is an arbitrary real number. Third, an arbitrary value of with
increase quite fast if the change in the acceleration of the vehicle . This completes the proof of Property 3.3.
is small. It was also shown that all the unobservable errors can be
C. Proof of Property 3.4
made observable throughthe maneuvering.Acceleration changes
improve the estimates of attitude and gyro bias. The changes in an- From (36) with , it follows that
gular rate enhance the estimate of the lever arm between the GPS
(59)
antenna and IMU. However, the motion with a constant angular
velocity has no influence on the lever arm estimation. Decompose and such that
A covariance simulation on the error state was given to verify
the results of observability analysis. The numerical simulation (60)
results showed that acceleration changes were effective to im- (61)
prove attitude estimation. They also showed that changes in an-
gular velocity reduced lever arm estimation error. Changes in where , , , , , and are real numbers. Then,
both acceleration and attitude were shown to be effective for from (59), it follows that
the estimation of the vertical component of gyro bias.
Even though the numerical simulation on the error covariance
can be considered as a useful means for the statistical analysis of (62)
the degree of observability, it has limitations on the simulation
of the behavior of state estimators for real measurement systems. Since , , and are linearly independent, it follows
Precise mathematical descriptions of vehicle trajectory, low-fre- that
quency components of errors in the inertial sensors, and multipath
error in GPS measurements are not easy to obtain. Field tests may (63)
be necessary as a separate work to obtain experimental evaluation (64)
of the effect of maneuvering on the error estimation. (65)

APPENDIX A However, (63) and (64) imply that

A. Proof of Property 3.2 (66)

Note that . Thus , ,

and for . These relations also hold Since is not parallel with , . Thus,
in the proofs of Properties 3.3–3.6. From (44), it follows that and . From (36) with
, it follows that
(54) (67)
(55) (68)
.. ..
. .
(56) (69)
HONG et al.: OBSERVABILITY OF ERROR STATES IN GPS/INS INTEGRATION 741

Thus, if for . This

, then (67)–(69) are completes the proof Property 3.9.
satisfied with . Otherwise, .
Obviously, (44) holds with any nonzero . This completes H. Proof of Property 3.10
the proof of Property 3.4. Let , , , where , , and
are constant initial angles, is a constant pitch rate, is a
D. Proof of Property 3.5 constant yaw rate, and is time. Then, it follows that
Decompose and into the forms of (60) and (61),
respectively, and decompose into the form

(70)

where , , and are real numbers. The proof of

Property 3.4 showed that (36) with is satisfied with
and . For , (36)
implies that (75)

where , , , .
(71) Then, the proof follows from the relation

Since , , and are linearly independent, and is

not zero, it follows that
(76)
(72)
with ,
(73) , and .
(74) This completes proof of Property 3.10.

Equations (72) and (73) imply that . With this I. Proof of Property 3.11
result, (74) implies that . Thus, . Indeed Let , , where , , and
(44) is satisfied with any nonzero . This completes the proof are constant initial angles, is a constant roll rate, is a
of Property 3.5. constant yaw rate, and is time. Then, it follows that

E. Proof of Property 3.7

The property can be easily derived from the relations
and , for
, because is constant. This completes the
proof of Property 3.7.

F. Proof of Property 3.8

(77)
Since and all orders of the time derivatives of it have

the same direction, , . Thus, Thus, we have (78) shown at the bottom of the page. If ,
for all . This completes the then
proof of Property 3.8.

G. Proof of Property 3.9 (79)

Property 3.9 can be proved with the following relations:
, Otherwise, is a full rank matrix if . This completes
, , and proof of Property 3.11.

(78)
742 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 54, NO. 2, MARCH 2005

J. Proof of Property 3.12 [15] M. Wei and K. P. Schwarz, “A strapdawn inertial algorithm using an
earth-fixed cartesian frame,” Navigation, vol. 37, no. 2, pp. 153–167,
Let , , and where , , and 1990.
are initial angles, is a constant yaw angular acceleration, [16] D. Goshen-Meskin and I. Y. Bar-Itzhack, “Unified approach to inertial
navigation system error modeling,” J. Guidance Contr. Dyn., vol. 15, no.
and is time. Then, it follows that 3, pp. 648–653, 1992.
[17] A. Gelb, Applied Optimal Estimation. Cambridge, MA: MIT Press,
1974.
[18] J. A. Farrell and M. Barth, The Global Positioning System & Inertial
Navigation. New York: McGraw-Hill, 1999.
[19] J. A. Farrell, T. D. Givargis, and M. J. Barth, “Real-time differential
carrier phase GPS-aided INS,” IEEE Trans. Contr. Syst. Technol., vol. 8,
(80) no. 4, pp. 709–721, 2000.
[20] C. T. Chen, Linear System Theory and Design. New York: Holt, Rine-
hart and Winston, 1984.
[21] D. W. Allan, “Statistics of atomic frequency standards,” Proc. IEEE, vol.
54, no. 2, pp. 221–230, Feb. 1966.
[22] D. Gebre-Egziabher, R. C. Hayward, and J. D. Powell, “A low-cost
GPS/inertial attitude heading reference system (AHRS) for general avi-
ation applications,” in Proc. 1998 IEEE Position, Location Navigation
Symp., Palm Springs, CA, 1998, pp. 518–525.
[23] H. Hou and N. El-Sheimy, “Inertial sensors errors modeling using Allan
variance,” in Proc. ION GPS/GNSS 2003, Portland, OR, pp. 2860–2867.
(81) [24] F. M. Ham and R. G. Brown, “Observability, eigenvalues, and Kalman
filtering,” IEEE Trans. Aerosp. Electron. Syst., vol. AES-19, no. 2, pp.
269–273, 1983.
Let and , where is a real [25] P. S. Maybeck, Stochastic Models, Estimation, and Control. New
number. The observability test condition implies York: Academic Press, 1979, vol. I.
that .
This implies that and
Sinpyo Hong received the B.A. degree from Pusan
where is a real National University, Busan, Korea, in 1982, the M.S.
number. Thus, the system has two unobservable modes degree from Korea Advanced Institute of Science and
in (52). This completes the proof of Property 3.12. Technology, Seoul, Korea, in 1985, and the Ph.D. de-
gree from the University of California, Los Angeles
(UCLA), in 1993, all in mechanical engineering.
From 1985 to 1986, he was a Design Engineer with
REFERENCES Korea Heavy Industries and Construction Co. From
[1] T. Bell, “Error analysis of attitude measurement in robotic ground ve- 1986 to 1990, he was a Research Engineer for the de-
hicle position determination,” Navigation, vol. 47, no. 4, pp. 289–296, velopment of a robot with Samsung Advanced Insti-
Winter 2000–2001. tute of Technology, Giheung, Korea. From 1993 to
[2] X. He and L. Jianye, “Analysis of lever arm effects in GPS/IMU inte- 1997, he was involved in the development of a navigation system for a unmanned
gration system,” Trans. Nanjing Univ. Aeronaut. Astronaut., vol. 19, no. aerial vehicle at UCLA. Since 2000 he has been working on the development of
1, pp. 59–64, Jun. 2002. an unmanned land vehicle at Pusan National University, where he currently is a
[3] S. Hong, Y. S. Chang, S. K. Ha, and M. H. Lee, “Estimation of alignment Research Professor with the Advanced Ship Engineering Research Center. His
errors in GPS/INS integration,” in Proc. Inst. Navigation GPS 2002, current research interests include design and analysis of integrated navigation
Portland, OR, pp. 527–534. systems and control systems.
[4] S. Hong, M. H. Lee, J. A. Rios, and J. L. Speyer, “Observability analysis Dr. Hong is a member of the Institute of Navigation.
of INS with a GPS multi-antenna system,” Korean Soc. Mech. Eng. Int.
J., vol. 16, no. 11, pp. 1367–1378, 2002.
[5] A. A. Sutherland Jr., “The Kalman filter in transfer alignment of inertial Man Hyung Lee (S’79–M’83–SM’01) was born in
guigance systems,” J. Spacecraft Rockets, vol. 5, pp. 1175–1180, 1968. Korea in 1946. He received the B.S. and M.S. degrees
[6] J. Baziw and C. T. Leondes, “In-flight alignment and calibration of iner- in electrical engineering from Pusan National Uni-
tial measurement units-part I: general formulation,” IEEE Trans. Aerosp.
versity, Pusan, Korea, in 1969 and 1971, respectively,
Electron. Syst., vol. AES-8, pp. 440–449, 1972.
and the Ph.D. degree in electrical and computer en-
[7] I. Y. Bar-Itzhack and B. Porat, “Azimuth observability enhancement
gineering from Oregon State University, Corvallis, in
during inertial navigation system in-flight alignment,” J. Guidance
Contr., vol. 3, pp. 337–344, 1981. 1983.
[8] B. Porat and I. Y. Bar-Itzhack, “Effect of acceleration switching during From 1971 to 1974, he was an Instructor in the De-
INS in-flight alignment,” J. Guidance Contr., vol. 4, pp. 385–389, 1981. partment of Electronics Engineering, Korea Military
[9] D. Goshen-Meskin and I. Y. Bar-Itzhack, “Observability analysis of Academy, Seoul. He was an Assistant Professor in the
piece-wise constant systems-part II: application to inertial navigation Department of Mechanical Engineering, Pusan Na-
in-flight alignment,” IEEE Trans. Aerosp. Electron. Syst., vol. 28, no. 4, tional University, from 1974 to 1978. From 1978 to 1983, he was a Teaching
pp. 1068–1075, 1992. Assistant, Research Assistant, and Postdoctoral Fellow at Oregon State Univer-
[10] , “Observability analysis of piece-wise constant systems-part sity. Since 1983, he has been a Professor in the College of Engineering, Pusan
I: theory,” IEEE Trans. Aerosp. Electron. Syst., vol. 28, no. 4, pp. National University, where he was a Pohang Iron and Steel Co., Pohang, Korea
1056–1067, 1992. (POSCO) Chair Professor in the School of Mechanical Engineering from 1997
[11] I. Y. Bar-Itzhack and N. Berman, “Control theoretic approach to inertial to 2003 and was Dean of the College of Engineering from March 2002 to Feb-
navigation systems,” J. Guidance Contr., vol. 11, no. 3, pp. 237–245, ruary 2004. His research interests are estimation, identification, stochastic pro-
1988. cesses, bilinear systems, mechatronics, micromachine automation, and robotics.
[12] Y. F. Jiang and Y. P. Lin, “Error estimation of INS ground alignment He is the author of more than 750 technical papers and was Program Chair of
through observability analysis,” IEEE Trans. Aerosp. Electron. Syst., vol. the 1998 4th ICASE Annual Conference.
28, no. 1, pp. 92–97, 1992. Dr. Lee is a member of the American Society of Mechanical Engineers,
[13] S. Hong, M. H. Lee, S. H. Kwon, and H. H. Chun, “A car test for the Society for Industrial and Applied Mathematics, and Society of Photo-Optical
estimation of GPS/INS alignment errors,” IEEE Trans. Intell. Transp. Instrumentation Engineers. He was General Cochairman of the 2001 IEEE
Syst., vol. 5, no. 3, pp. 208–218, 2004. International Symposium on Industrial Electronics (ISIE) and is a General
[14] K. R. Britting, Inertial Navigation System Analysis. New York: Wiley- Cochairman of The 30th Annual Conference of the IEEE Industrial Electronics
Interscience, 1971. Society (IECON ’04).
HONG et al.: OBSERVABILITY OF ERROR STATES IN GPS/INS INTEGRATION 743

Ho-Hwan Chun received the B.Sc. and M.Sc. de- Jason L. Speyer (M’71–SM’82–F’85) received the
grees in ship and ocean engineering from Pusan Na- S.B. degree in aeronautics and astronautics from the
tional University, Korea, in 1983 and 1985, respec- Massachusetts Institute of Technology, Cambridge,
tively. He received the Ph.D. degree from the Depart- in 1960 and the Ph.D. degree in applied mathematics
ment of Naval and Ocean Engineering, Glasgow Uni- from Harvard University, Cambridge, MA, in 1968.
versity, U.K., in 1988. His industrial experience includes research at
He was Yard Research Fellow at Glasgow Univer- Boeing, Raytheon, Analytical Mechanics Associ-
sity from 1988 to1990 and a Principal Researcher ated, and the Charles Draper Laboratory. He was the
with the Hyundai Maritime Research Institute, Harry H. Power Professor in Aerospace Engineering
Ulsan, Korea, from 1991 to 1993. He has been an at the University of Texas, Austin, and is currently
Associate Professor in the Department of Naval a Professor in the Mechanical and Aerospace Engi-
Architecture and Ocean Engineering, Pusan National University, since 1994. In neering Department, University of California, Los Angeles. He spent a research
2002, he became Director of the Advanced Ship Engineering Research Center leave as a Lady Davis Visiting Professor at The Technion—Israel Institute of
(ASERC), designated by the Ministry of Science and Technology, Korea. Technology, Haifa, in 1983. He was the 1990 Jerome C. Hunsaker Visiting
His main research area is hydrodynamics such as hull form design, hull and Professor of Aeronautics and Astronautics at the Massachusetts Institute of
structure interactions with waves, WIG and drag reduction, etc. Technology.
Dr. Speyer is a Fellow of the American Institute of Aeronautics and Astro-
nautics. He has twice been an elected member of the Board of Governors of the
IEEE Control Systems Society. He has been an Associate Editor of the IEEE
Sun-Hong Kwon received the B.S. degree in naval TRANSACTIONS ON AUTOMATIC CONTROL and Chairman of the Technical Com-
architecture from Pusan National University, Busan, mittee on Aerospace Control. From October 1987 to October 1991 and from
Korea, in 1978, the M.E. degree in ocean engineering October 1997 to October 2001, he was a member of the USAF Scientific Advi-
from Stevens Institute of Technology, Hoboken, NJ, sory Board. He received the Mechanics and Control of Flight Award and Dryden
in 1983, and the Ph.D degree in aerospace and ocean Lectureship in Research from the American Institute of Aeronautics and Astro-
engineering from Virginia Polytechnic Institute and nautics in 1985 and 1995, respectively. He received an Air Force Exceptional
State University, Blacksburg, in 1986. Civilian Decoration in 1991 and 2001 and the IEEE Third Millennium Medal
He joined Pusan National University in 1986 in 2000.
as a Faculty Member in the Department of Naval
Architecture and Ocean Engineering. His current
research interests are the application of wavelet
analysis to ocean engineering and development of sea wave monitoring system
by using radar images.

View publication stats