Analysis and Modelling of MEMS Inertial

Measurement Unit
Alex G. Quinchia and
Carles Ferrer

Universitat

Aut`onoma de Barcelona
de Microelectr`onica de Barcelona
(CNM, CSIC)
Barcelona, Spain
Email: Alex.Garcia.Quinchia@uab.cat

Institut

Abstract Thanks to advances in the development of MicroElectromechanical Systems (MEMS), it has been possible to
fabricate small dimension and cheap accelerometers and gyros,
which are being used in many applications where the GPS/INS
integration is carried out, as for example to identify track
defects, navigation, geo-referencing, agriculture, etc. Although
these MEMS devices have a low-cost, they present different errors
which degrade the accuracy of the navigation systems in a short
period of time. Therefore, a suitable modelling of these errors is
necessary in order to improve the system performance. In this
work, Allan Variance and Power Spectral Density techniques
are used to identify the random processes that affect the inertial
sensor data. Once the random components are identified, they
are modelled using first-order Gauss-Markov and random walk
processes. Two models are assessed augmenting the states of the
Extended Kalman Filter (EKF) to 6 and 9. Subsequently, another
analysis and modelling of the inertial sensors which combines
Autoregressive Filters and Wavelet De-noising is implemented
and in this case the EKF of the loosely coupled GPS/INS
integration strategy is augmented with 6, 12 and 18 states. Finally,
the results show a comparison between these sensor error models
with real data under GPS outage conditions.

I. I NTRODUCTION
Inertial Navigation Systems (INS) can provide information
about position, velocity and attitude using measurements of
inertial sensors as gyros and accelerometers. The use of these
sensors can provide high resolution in the positioning and
in addition it is complementary to the GPS. Consequence
of the low-cost and small size of the inertial sensors, the
demand has increased and their usage has expanded in several
applications where GPS/INS are blended. Nonetheless, lowcost inertial sensors are characterized by high noise and large
uncertainties in their outputs, such as bias, scale factor and
non-orthogonality [1]. In other words, a low-cost INS presents
errors in position, velocity and attitude which grow rapidly.
Therefore, a suitable modelling of these inertial sensors is
necessary in order to improve the system performance.
The navigation instrument where the inertial sensors are
mounted is the Inertial Measurement Unit (IMU) that is
part of the INS. In a low-cost INS (MEMS grade) different
errors affect the measurement, which can be classified as
deterministic or stochastic errors.

Gianluca Falco
Emanuela Falletti and
Fabio Dovis
Politecnico di Torino and Istituto
Superiore Mario Boella
Turin, Italy
Email: falco@ismb.it

Deterministic errors are due to manufacturing and mounting

defects and can be calibrated out from the data; on the other
hand, the stochastic errors are the random errors that occur
due to random variations of bias or scale factor over time
[2]. There are several errors that affect the inertial sensors:
the misalignment errors are the result of non-orthogonalities
of the sensor axes and are usually treated as deterministic
error. The scale factor represents the sensibility of the sensor
and it is the result of manufacturing tolerances or aging; it
is usually divided between linear and non-linear part where
the linear part is obtained from calibration while the nonlinear is modelled with a stochastic process. In the case of
the bias, it is divided between bias turn-on and bias drift, the
bias turn-on is constant but it varies from turn-on to turnon and is considered as a deterministic error, the bias drift
presents a random behaviour and needs to be modelled with
a stochastic process. Regarding the random error, this is an
additional signal resulting from noise of the sensor itself or
other components that interfere with the signal provided by
the sensor.
In this sense, a correct noise models in the inertial sensors
leads to better performance for the GPS/INS integrated navigation system. So, this work is focused on the identification of
the stochastic error of the inertial sensors and it is organized
as follows. We begin with the Allan Variance (AV) analysis
and the Power Spectral Density (PSD) technique, used to
identify the characteristics of the underlying random processes
that affect the inertial sensors and followed by an analysis
of the random errors. Subsequently, a combination between
Wavelet De-noising and AR models is done using different
levels of decomposition and different orders in the AR models.
Finally, the models that were identified using AV and PSD are
compared with the models that were obtained with Wavelet
De-noising and AR models. These models are assessed using
real data collected in urban roadways.
II. I NERTIAL SENSORS
A. Power Spectral Density

STOCHASTIC ERROR

Power Spectral Density is an important descriptor of a

random process because it provides information of the signal

978-1-4673-2343-7/12/$31.00 2012 IEEE

that is not easy to extract from the time domain.

The PSD is related to the autocorrelation function with
Z
Sx (jw) = F [Rxx ( )] =
Rxx ( )ejwt d
(1)

The AV obtained from Eq. (2) is related to the two-sided

PSD by:
Z
sin4 (f T )
2 (T ) = 4
df Sx (f )
(3)
(f T )2
0

where, Sx (jw) is the Power Spectral Density of the process

x, F [] indicates Fourier transform, and Rxx ( ) is the autocorrelation of the process x.
The PSD contains information for analysing and characterizing the process x from a set of observed data. Basically, the
PSD is used to identify the IMU systematic errors from the
frequency components.
The parameters of the errors affecting the inertial sensors
may be estimated from the frequency components of the PSD,
which eventually can be used in the stochastic model of the
INS.
The PSD brings information to determine if either the
specific noise exist or is suspected to influence the inertial
sensor. Obviously, the number of random noises that might be
present in the PSD curve depends on the type of sensors. The
derivation of these noise terms are well detailed in [3], [4],
[5].

where Sx (f ) is the PSD of the random process x, written in

(1).
An interpretation of equation (3) is that the Allan variance
is proportional to the total noise power of the sensor output
when passed through a bandpass filter with transfer function
sin4 (f T )/(f T )2 . This filter depends on T , which suggests
that different types of random processes can be examined by
adjusting the correlation time (T ). Thus, the AV provides a
mean of identifying and quantifying various noise terms that
exist in the data [3].

B. Allan Variance
The Allan Variance is a time domain analysis technique
originally developed to study the frequency stability of oscillators, later on it was adapted to analyse random drift of
inertial sensors [6]. This technique has been used to determine
the characteristics of the random processes that give rise to
the measurement noise of the sensors. As such, AV helps
identifying the source of a given noise term in the observed
data [3].
The Allan Variance is estimated as follows:
2 (T ) =

NX
2n
1
(k+2n 2k+n + k )2
2T 2 (N 2n)

(2)

k=1

where T represents the correlation time or cluster time, i.e.,

the time associated with a group of n consecutive observed
data samples, N is the length of the data that will be analysed
and is the output velocity in case of the accelerometers and
output angle in case of the gyros, these measurements are
made at discrete times from the inertial sensors.
The basic idea to estimate the AV is to take a long sequence
of data (N ), in this case, the data that is collected from the
stationary IMU. Once the data is obtained and after removing
the turn-on bias, the output of the inertial sensors is integrated
in order to get , then having obtained , the AV estimate can
be obtained through equation (2).
In AV, the uncertainty in the data is assumed to be generated
by noise sources of specific character, as for instance rate
random walk, angle random walk, bias instability, etc. In
order to obtain the covariance of each noise source affecting
the sensor output, it is necessary to analyse the AV result
computed by (2). This is usually done by plotting a log-log
AV curve, from which the covariance values for each error
can be extracted.

C. Autoregressive Processes (AR)

To avoid the problem of inaccurate modelling of inertial
sensor random errors due to inaccurate autocorrelation function determination, another method can be applied, introduced
in [7]. There are different works where the AR models have
been evaluated and compared with other methods, some of
them are [8], [9], [10], [11].
First-order Gauss-Markov (GM) process has been very
useful for modeling random errors, however, better stochastic
modeling of the inertial sensor errors can be achieved by
modeling these errors as higher order AR models [8]. In
addition, the autocorrelation of the random error often seems
to follow a higher order GM process, which can be modeled
using an appropriate AR model.
An AR process is a time series produced by linear combination of past values, which can be described by the following
linear equation [9]
x(n) =

p
X

k=1

k x(n k) + 0 w(n)

(4)

where x(n) is the process output that is a combination of

past outputs, plus a white noise w(n) with standard deviation
0 , p is the order of the AR process and k are the model
parameters. It is assumed that these coefficients (0 ,k ) are
computed so that the linear system is stable, making the model
stationary [12]. It should be noted in Eq. (4) that if p = 1,
then the AR process approximates first-order GM process. On
the other hand, if p = 1 and 1 = 1 it becomes a Random
Walk process.
In this paper we focus on AR models up to 3rd order, since a
higher would increase the computational load and might result
in unstable solutions [10].
III. E XPERIMENTAL W ORK
A. Inertial measurement unit and data acquisition
In order to evaluate and compare these models, the IMU
3DM-GX3 MEMS grade of MicroStrain was used. It combines
a triaxial accelerometer, triaxial gyro, triaxial magnetometer
and a temperature sensor. All quantities are temperature compensated and are mathematically aligned to an orthogonal

B. AV analysis
Having collected the sensor data, the identification of the
inertial sensors noise was carried out using the Allan variance
and the power spectral density. Before, it was necessary to
eliminate the turn-on bias for each sensor as well as the
approximate measure of the gravity provided by the output of
the z-axis accelerometer. Subsequently, for AV analysis the
acceleration and the angular rate were integrated to obtain
the instantaneous velocity and angle. Then the log-log plot
of Allan variance standard deviation versus cluster times (T )
was obtained using Eq. (2). The results are plotted in Fig. 1
for the accelerometer and Fig. 3 for gyro data respectively.

which is very good. Thus, the velocity random walk or in

other words the noise term (N ) for z-axis accelerometer is
determined as

N = 0.047 0.00050 (m/s/ hr)

(5)
2

10

Acceleration random walk (K)

Allan standard deviation (m/s/hr)

coordinate system [13]. The IMU was configured with a

sampling frequency of 100Hz and the second stage filter
implemented in the microcontroller was adjusted with a filter
width of 15, which means an attenuation of 14.16% at 20Hz,
for more details of this digital filter see [14].
The test for static analysis was conducted along two days,
each day recording 14 hours of data; then, the more stable
7 hours without vibrations of each day were extracted for
analysis.

10

Bias instability (B)

0

10

Velocity random walk (N)

1

10

10 2
10

Fig. 2.

10

10

10
Cluster times (seg)

10

10

10

IMU 3DM-GX3-25 Allan Variance for z-axial accelerometer

10

Gyro X
Gyro Y
Gyro Z

Acc X
Acc Y
Acc Z

Allan standard deviation (deg/hr)

Allan standard deviation (m/s/hr)

10

10

10

10 2
10

10 2
10

10

10
10
Cluster times (seg)

10

Fig. 3.
Fig. 1.

10

10

10
Cluster times (seg)

10

10

10

10

IMU 3DM-GX3-25 Allan Variance for 3 gyro axes

IMU 3DM-GX3-25 Allan Variance for 3 accelerometer axes

Fig. 1 shows the AV estimated on the 3DM-GX3 accelerometers. It presents three types of errors: velocity random walk
(N), bias instability (B) and acceleration random walk (K).
These errors are identifiable drawing straight lines for each
error with its corresponding slope. The plot also shows that
the z-axis has a bias instability lower than the other two
accelerometers. To clarify, Fig. 2 depicts straight lines for each
noise of z-axis accelerometer: in this case the accelerometer
has N ,B and K with slopes 1/2, 0 and 1/2 respectively. It
can be seen that the dominant noise in short cluster times is the
velocity random walk, while the dominant error in long cluster
times is the acceleration random walk. From the straight line
with slope 1/2 fitted to the beginning of the N noise, it
can be read a value = 0.047 (m/s/hr) at a cluster times of
1hr. Since the velocity random walk (N ) is present in a cluster
times interval where the number of independent cluster is very
large, the estimation accuracy of the AV is approximately 1.1%

The Allan variance standard deviation versus cluster times

(T ) for gyro data is depicted in Fig. 3, where two types of
noises can be recognized: an angle random walk for short
cluster times and a bias instability for long cluster times.
For the x-axis gyro (blue curve), the bias instability is
present in the time range between 321.92 (s) and 654.01 (s).
The value of this error can be measured with a flat line at
29.57 (deg/hr). Dividing this standard deviation by the factor
0.664 as suggested in [3], the B coefficient can be achieved:
B = 44.533 5.14 (deg/h)

(6)

For more details of the procedure to analyse the AV curve

see [3], [5], [6], [15], [16].
Table I summarizes the error coefficients with their respective uncertainty for accelerometers and gyro data. The
correlation time (Tc ) of the bias instability (B) and the
standard deviation for each sensor of the IMU 3DM-GX3-25
is shown in Table II.

TABLE I
6

10

I DENTIFIED ERROR COEFFICIENTS FOR ACCELEROMETERS AND GYRO OF

THE

3DM-GX3 IMU WITH AV

4

Acc X
Acc Y
Acc Z

Gyro X
Gyro Y
Gyro Z

Velocity random
walk (N)

(m/s/ h)
0.045 0.00023
0.045 0.00022
0.047 0.00050

Bias instability
(B) (m/s/h)
5.1581 0.0370
4.5507 0.0506
1.8336 0.0524

Acceleration
random walk (K)
(m/s/h3/2 )
166.30 4.6398
24.95 2.8368
13.53 1.8685

Bias instability
(B) (deg/h)

Rate random walk

(K) (deg/h3/2 )

Angle random
walk
(N)
(deg/ h)
2.420 0.0974
1.988 0.0565
2.164 0.0599

PSD (m/s/h)2/Hz

10

10

Velocity random walk (N)

0

10

Bias instability (B)

2

10

Acceleration random walk (K)

44.533 5.14
38.810 2.51
31.717 2.29

10 5
10

Fig. 5.

TABLE II

10

10

10
10
Frequency (Hz)

10

10

10

Power Spectral Density Accelerometer Z IMU 3DM-GX-25

I DENTIFIED T c (B) AND STANDARD DEVIATION FOR 3DM-GX3 IMU

Correlation time
Tc (sec)

Acc X

Acc Y

Acc Z

Gyro X Gyro Y

Gyro Z

1.29

41.09

20.74

654.01 83.27

166.49

0.0068 0.0065 0.0063 0.0055 0.0045 0.0048

STD

C. PSD analysis
In order to check the validity of the noise coefficients
obtained with Allan variance, the PSD was implemented
using Welchs method. The results are shown in Fig. 4 for
accelerometer data.
6

10

Acc X
Acc Y
Acc Z

10

PSD (m/s/h)2/Hz

10

10

10

10

the z-axis accelerometer curve is plotted in Fig. 5 with straight

lines for each noise N,B and K with their respective slopes
0,1,2. The acceleration random walk is present in the
low frequency components between 1 104 Hz and 2.29
103 Hz. It is obtained fitting a straight line with a slope of
2 starting from 1 104Hz up to it meets the vertical line of
f = 1Hz. For details of the intercepts to determine the noise
parameters see [3]. The bias instability (B) is the dominant
noise between 2.29 103 Hz and 7.1 102 Hz with a slope
of 1, while the velocity random walk is present between
0.1248Hz and 20Hz. After 20Hz there is an attenuation
because of the digital moving average filter which is used to
remove high frequency spectral noise produced by the MEMS
sensors, this digital filter is implemented in the microcontroller
of the IMU 3DM-GX3-25.
Table III summarises the values of different errors that affect
the inertial sensors using PSD method. This verifies the results
that were obtained with AV analysis, where was also identified
velocity random walk (N), bias instability (B) and acceleration
random walk (K) for accelerometers data and angle random
walk (N) and bias instability (B) for gyro data. It can be seen
that most of the estimated values are within the incertitude
interval computed by AV (Table I).

10

TABLE III
1

10 5
10

Fig. 4.

10

10

10
10
Frequency (Hz)

10

10

10

I DENTIFIED ERROR COEFFICIENTS FOR ACCELEROMETERS AND GYRO OF

THE 3DM-GX3 IMU WITH PSD

Power Spectral Density Accelerometers IMU 3DM-GX3-25

Fig. 4 shows the one-side PSD for accelerometers data,

it can be seen again that there are three types of noise, the
acceleration random walk (K), the bias instability (B) and the
velocity random walk (N). This plot also confirms that z-axis
accelerometer has a bias instability smaller than the other two
accelerometers, and the velocity random walk is almost the
same for all the accelerometers (slope 0) which is coherent
with the results obtained with AV (slope 1/2).
The values for each noise parameter was extracted as in AV,
but in this case the interceptions are different. For instance,

Bias instability
(B) (m/s/h)

Acc X
Acc Y
Acc Z

Velocity random
walk (N)

(m/s/( h)
0.045
0.044
0.047

4.6447
4.6700
1.7733

Acceleration
random walk (K)
(m/s/h3/2 )
168.60
26.66
14.60

Bias instability
(B) (deg/h)

Rate random walk

(K) (deg/h3/2 )

Gyro X
Gyro Y
Gyro Z

Angle random
walk
(N)
(deg/ h)
2.297
1.937
2.058

43.438
39.614
30.705

Having identified the random errors using AV and PSD,

a)
1500

|Gy(f)| (Deg/h)

the parameters obtained with AV were used in the model

considering that they provide a percentage error. Thus the
3DM-GX3 accelerometers stochastic error ase was modeled
as:
ase = 1st GM (B) + RW (K)
(7)

1000

500

0
0

gse = 1st GM (B)

(8)

where B is modeled as first-order Gauss-Markov process

(1st GM). This noise can be modelled by a combination of
Markov noise states [4]. In this case a first-order GaussMarkov process was fitted to the flat part of the AV curve
taking into account B and its correspondent correlation time
(Tc )(see Table II). Regarding the noise term N, it presents high
frequency components that have correlation time much shorter
than the sample time, most of these sources can be eliminated
by design. In this work it was removed by filtering, so it is not
in the error model. There are different approaches to model
these noise terms, some of them are presented in [16], [17].
For the sake of comparison with the analysis developed so
far, next section presents an inertial sensor error model developed combining AR models and wavelet de-noising technique.
D. AR model and wavelet de-noising
The Discrete Wavelet Transform (DWT) is a widely used
technique in digital signal processing, one of its characteristics
is that allows to do a multiresolution analysis of a signal,
in other words, DWT decomposes a signal into different frequency components at different resolutions. Moreover wavelet
de-noising takes advantage of the sub-band decomposition
performed by the DWT and removes the noise by eliminating
the frequency components that are less relevant, this procedure
is well described in [7], [18], [19], [20]. There are different ways to remove the noise with this technique, one of
them is removing all the details. Another way is applying
a threshold on the DWT coefficients, where it is assumed
that the magnitude of the actual signal is greater than the
noise level, and the noise is white noise. This is known
as wavelet thresholding. There are various techniques for
choosing de-noising parameters but so far there is not any
universal threshold determination rule [18].
The idea of combining AR models and wavelet de-noising
is to reduce high frequency noise in order to obtain a residual
noise with slow variations, so its autocorrelation function will
decrease slowly. This slow variation in the residual noise is
more suitable to model when a low order AR process is used.
Having the inertial static data, the wavelet de-noising technique is applied, then after removing part of the noise the
AR model coefficients are estimated with Burgs method. This
procedure is executed for each sensor and for the three AR
models: 1st, 2nd and 3rd order.

10

15

20

25

30

35

40

45

50

20
25
30
Frequency (Hz)

35

40

45

50

b)
1500

|Gy(f)| (Deg/h)

where the noise term K is modeled as a random walk (RW),

while the bias instability is modeled as first-order GaussMarkov process (1st GM).
Regarding the 3DM-GX3 gyro stochastic error gse , the
model was defined as:

1000

500

0
0

10

15

Fig. 6. a) Spectrum of Gyro Y. b) Spectrum of Gyro Y after applying wavelet

de-noising with 4 levels of decomposition.

The number of decomposition levels for the DWT was

determined from the inertial kinematic data. Figure 6 shows
the spectrum of the gyro Y with a sampling frequency of
100 Hz under kinematic conditions before and after applying
wavelet de-noising with 4 levels of decomposition with soft
thresholding.
IV. R ESULTS
In this work we use loosely-coupled integration with feedback which corrects the INS error through a close-loop. This
approach combines the inertial solution with the GPS data
having the residual error of position and velocity as input to
the Extended Kalman Filter (EKF). The INS error dynamics
equations are built in the EKF, having initially 9 states for
position, velocity and attitude error. Details of this strategy of
integration can be found in [2], [21], [22].
The EKF was adapted for each error model in order to
evaluate the six stochastic error models that were obtained
from the previous analysis. Firstly, the two models extracted
from AV and PSD were implemented, so the vector error
states of the Extended Kalman Filter was augmented with 6
and 9 states. Subsequently, three autoregressive models were
assessed augmenting EKF with 6, 12 and 18 states. It should
be noted that if the order of the AR model increases by one
the variables in the state vector of the Extended Kalman Filter
(EKF) will increase by six since this model will be used
in each of the six inertial sensors. Table IV summarizes the
stochastic models for the 3DM-GX3 sensors and the number
of states that are required in the loosely-coupled GPS/INS
integration.
The EKF for the loosely-coupled integration has 15 states
for two models, one is the model obtained with AV\PSD
where the bias instability (B) of both accelerometers and gyro
are modelled with a first-order Gauss-Markov process. The
second model with 15 states is a first-order AR model.
In order to assess the performance of the inertial sensor
error models, a vehicle was equipped with the 3DM-G3X
MEMS grade IMU and the Sat-Surf platform with u-blox
LEA-5X receiver [23]. The data set was collected in urban

TABLE IV
N UMBER OF STATES IN THE LOOSELY COUPLED INTEGRATION
ARCHITECTURE FOR DIFFERENT ERROR MODELS

behaviour of these models under different scenarios having

GPS signal blockages.
V. C ONCLUSION

AV\PSD ase

15 States

18 States

GM (B)

GM (B)

21 States

27 States

2nd orderAR

3rd orderAR

+
RW (K)

AV\PSD gse

GM (B)

AR ase \gse

1st orderAR

GM (B)

roadways inside the city of Turin, Italy. After the data collection campaign, the loosely-coupled integration architecture
with the error models presented in this paper were evaluated
by intentionally introducing one GPS outage lasting 60 s.
Figure 7 shows the horizontal error for each of the error model
developed.
250
GM(B)
1st order AR
GM(B)+RW(K)
3rd order AR

Error (m)

200

150

100

50

In this work different stochastic error models for the measurement noise components of a MEMS-based IMU have
been derived from experimental data and compared. It can
be concluded that PSD method can help to identify error
parameters, however the accuracy of the spectral density
estimation has to be considered. The bunching of the high
frequency showed in the analysis could be remove applying
frequency averaging technique, this way a better estimation
of the noise terms can be done. Regarding Allan variance
method, it allows a characterization of the random errors
by computing a simple operation. Despite this, the accuracy
of the Allan variance analysis is not good in large cluster
times, although this can be improved overlapping sequences
of AV results. Finally, AR models combined with wavelet denoising have more modelling flexibility than first-order GaussMarkov, random walk and random constant processes and also
present better results. However, it is necessary to consider the
computational cost of the wavelet filtering and the number of
states of the filter in a real-time system. It is also important to
mention that depending on the application, the selection of the
decomposition level and the thresholding method have to be
carefully analysed, due to the fact that frequency components
which are associated with, e.g., vehicles dynamics, may be
eliminated after performing wavelet de-noising.
ACKNOWLEDGMENT

0
0

100

Fig. 7.

200

300
400
Time (s)

500

600

700

Horizontal error with simulated 60 sec GPS outage

It can be noticed that the error is small during the first part
of the trajectory. The minimum error during the outage (from
570 s to 630 s) corresponds to the 3rd order AR model which
is depicted in the red line.
Table V presents the computation position error for each
model during the outage.
TABLE V
H ORIZONTAL ERROR WITH SIMULATED (60 s) OUTAGE

Error AV/PSD (m)

15 States
204.43

Error AR (m)

194.26

18 States
183.24

21 States

27 States
163.80

Although the 3rd order AR model presents the minimum

error, it has to be mentioned that these results can be improved
computing the deterministic errors and implementing INS
alignment. In addition, in order to guarantee an optimal an
accurate INS/GPS integration it is necessary to analyse the
temperature effect on the inertial sensors since MEMS-based
IMUs behaviour depends on temperature. Regarding the 2nd
order AR model, it has made unstable the EKF, so it has been
removed from the results. It is also necessary to consider the

Support funded by Spanish Ministry of Economy and Competitiveness under DELPHIS project: TEC 2009-09712.
R EFERENCES
[1] E.-H. Shin, Estimation techniques for low-cost inertial navigation,
Ph.D. dissertation, University of Calgary, Alberta, Canada, may 2005.
[2] P. Aggarwal, Z. Syed, A. Noureldin, and N. El-Sheimy, MEMS-Based
Integrated Navigation, ser. GNSS technology and applications series.
Artech House, 2010.
[3] IEEE standard specification format guide and test procedure for singleaxis interferometric fiber optic gyros, IEEE Std 952-1997, p. i, 1998.
[4] IEEE standard specification format guide and test procedure for linear, single-axis, non-gyroscopic accelerometers, IEEE Std 1293-1998
(R2008), pp. 1 249, 21 2011.
[5] N. El-Sheimy, H. Hou, and X. Niu, Analysis and modeling of inertial
sensors using allan variance, Instrumentation and Measurement, IEEE
Transactions on, vol. 57, no. 1, pp. 140149, jan. 2008.
[6] C. C. Naranjo, Analysis and modeling of MEMS based inertial sensors,
Masters thesis, School of Electrical Engineering Kungliga Tekniska
Hgskolan, Sweden, 2008.
[7] S. Nassar, Improving the inertial navigation system (INS) error model
for INS and INS/DGPS applications, Ph.D. dissertation, University of
Calgary, Alberta, Canada, nov 2003.
[8] J. Georgy, A. Noureldin, M. Korenberg, and M. Bayoumi, Modeling
the stochastic drift of a MEMS-based gyroscope in gyro/odometer/GPS
integrated navigation, Intelligent Transportation Systems, IEEE Transactions on, vol. 11, no. 4, pp. 856 872, dec. 2010.
[9] A. Noureldin, T. Karamat, M. Eberts, and A. El-Shafie, Performance
enhancement of MEMS-based INS/GPS integration for low-cost navigation applications, Vehicular Technology, IEEE Transactions on, vol. 58,
no. 3, pp. 1077 1096, march 2009.
[10] N. El-Sheimy, S. Nassar, K.-P. Schwarz, and A. Noureldin, Modeling
inertial sensor errors using autoregressive (AR) models, NAVIGATION,
vol. 51, no. 5, pp. 259268, 2005.

[11] M. Park and Y. Gao, Error and performance analysis of MEMS-based

inertial sensors with a low-cost GPS receiver, Sensors, vol. 8, no. 4,
pp. 22402261, 2008.
[12] A. Analytic Sciences Corporation Technical Staff. Gelb, Applied optimal
estimation. Cambridge, Massachusetts: MIT Press, 1974.
[13] 3DM-GX3-25 Technical Product Overview data sheet, MicroStrain,
Williston VT, USA.
[14] 3DM-GX3-25 Data Communications Protocol, MicroStrain, Williston
VT, USA.
[15] X. Zhang, Y. Li, P. Mumford, and C. Rizos, Allan variance analysis
of error characteristics of MEMS inertial sensors for an FPGA- based
GPS/INS system, in Proc. Int. Symp. on GPS/GNSS, Yokohama, Japan,
November 2008, pp. 127133.
[16] S. Moafipoor, L. Bock, J. Fayman, G. Mader, and P. de Jonge, Development and assessment of a low dynamic vehicle navigation system,
in Proc. International Technical Meeting of The Institute of Navigation,
San Diego, CA, EEUU, Jan. 2011, pp. 895907.
[17] Y. Yuksel, N. El-Sheimy, and A. Noureldin, Error modeling and
characterization of environmental effects for low cost inertial mems
units, in Position Location and Navigation Symposium (PLANS), 2010
IEEE/ION, may 2010, pp. 598 612.
[18] C. W. Kang, C. H. Kang, and C. G. Park, Wavelet de-noising technique
for improvement of the low cost MEMS-GPS integrated system, in
International Symposium on GPS/GNSS, october 2010.
[19] N. El-Sheimy, S. Nassar, and A. Noureldin, Wavelet de-noising for
IMU alignment, Aerospace and Electronic Systems Magazine, IEEE,
vol. 19, no. 10, pp. 32 39, oct. 2004.
[20] X. Wu, L. Duan, and W. Chen, A Kalman Filter approach based on
random drift data of fiber optic gyro, in Industrial Electronics and
Applications (ICIEA), 2011 6th IEEE Conference on, june 2011, pp.
1933 1937.
[21] A. Quinchia and C. Ferrer, A low-cost GPS&INS integrated system
based on a fpga platform, in Localization and GNSS (ICL-GNSS), 2011
International Conference on, june 2011, pp. 152 157.
[22] G. Falco, G. A. Einicke, J. T. Malos, and F. Dovis, Performance analysis of constrained loosely coupled GPS/INS integration solutions, in
Proc. International Global Navigation Satellite Systems Society IGNSS
Symposium 2009, Qld, Australia, December 2009.
[23] Navsas. (2012, May) Sat-surfer training board & software suite for
GNSS training. [Online]. Available: http://www.navsas.eu/