IMU Calibration on a Low Grade Turntable:

Embedded Estimation of the Instrument

Displacement from the Axis of Rotation
Alexander Kozlov, Igor Sazonov and Nina Vavilova
Navigation and Control Lab
Moscow State University
Moscow, Russia +7 495 939-5933
Email: a.kozlov.msu@gmail.com

Abstract— The paper concerns one simple method of cali- - it has one single axis of rotation (nearly horizontal);
bration of an assembled inertial measurement units (IMU) of - no rate measuring (not used if present);
different grades on a low grade single axis turntable. The - no rate stabilization required;
main feature of the method is that it has the weakest possible
requirements to the testbench. The method was presented at - no predefined angular positions or angle measuring;
ICINS 2010, ICINS 2013, and is now being used in industry for - no precise instrument alignment;
3–4 years. It appears to work well in practice. One of special - no programmable actuator control present (but more
points of this method is the situation, when the IMU is displaced convenient in practice);
from the axis of rotation. In some cases this fact can be neglected, - the axis of rotation is firmly fixed with respect to the
but in some cases not. It was shown earlier that parameters of this
displacement need not to be measured prior to the experiment, ground, and no vibrating within the desired accuracy of
but can be estimated automatically during the data processing, the IMU
similar to the rest of IMU parameters. So the device can be Please, see Fig. 1 for the schematic diagram of the calibra-
placed arbitrarily onto the stand, and no care of displacement
should be taken while conducting the experiment. This work tion experiment. It consists of three cycles of rotation round
concentrates primarily on aspects of observability and estimation a nearly horizontal axis, for about 10–20 minutes each. Note
accuracy of IMU displacement (which influence the accuracy of that there are no special requirements to these rotations except
the calibration itself), and shows more variety of experimental for a non-zero angular rate. Every cycle is preceded by 1–3
results than in previous publications. minute static initial alignment.
I. I NTRODUCTION
Conventional calibration techniques naturally depend on B. Sensor error model
the accuracy of the test bench. They require knowing either For inertial sensor errors we use linear model in small terms,
some precise positions, or angular rate, or both. But in assuming that some pre-calibration step was performed. This
fact, in a quite a lot of kinds of situations the information pre-calibration step can be based on the same experimental
given by inertial sensors is sufficient to calibrate themselves data (no additional operations required in a test bench), and it
without any additional information from other instruments. may be performed in different ways, but this is not the subject
Under the assumption that sensor errors are small enough, of this work. Thus, errors are assumed to be small enough
it appears to be possible to reduce errors to the second that second order terms can be neglected. The model includes
order of magnitude using the output from IMU only. In most accelerometer and gyro small constant biases, errors of scaling
cases the remaining sensor errors become negligible, or at factors, and small angles of sensitive axes misalignment. In ad-
least tolerable. Considering this, we can create a completely dition, the model can incorporate dynamic drift (g-sensitivity)
testbench-independent algorithm of calibration. This entirely coefficients and other parameters.
new approach to calibration was proposed by N.A. Parusnikov
The true components of unit force vector fz are being
from MSU [1]. More detailed description of the method, some
measured by accelerometers, where the subscript z stands for
aspects of its implementation for the case when several IMUs
instrumental reference frame. Similarly, the true components
are being calibrated simultaneously, and calibration results, are
of absolute angular rate vector ωz are being measured by
given below.
gyros. We use fz′ and ωz′ for corresponding sensor outputs,
II. I DEA AND THE BACKGROUND OF THE METHOD so that for 6 DoF IMU
A. The turntable  ′ 
fz1
 ′ 
ωz1
In this work, saying the turntable can be of low grade we fz′ =  fz2
′ 
, ωz′ =  ωz2′ .
′ ′
mean the following: fz3 ωz3
 Fig. 1. Schematic diagram of the calibration experiment

The conventional error model described above is now written C. The main concept
as follows. The idea of the method is to construct a linear state-space
system, containing all desired components in its state vector.
fz′ − fz ≡ ∆fz = ∆fz0 + Γfz + δfzs ,
This system then can be supplemented with measurements,
ωz′ − ωz ≡ −νz = −νz0 − Θωz − Dfz /g − δνzs , making it fully observable in experiments described above.
These measurements require no additional information, except
where minus signs in the second equation is just a convention, for precisely known local gravity force vector and Earth’s
and error components are grouped in matrices as outlined angular rate vector, both in geodetic frame.
below. For constant biases In order to derive the linear state-space system we need
to introduce some inertial navigation first. We begin with the
 0
  0 
∆fz1 νz1
∆fz0 =  ∆fz2 0 
, νz0 =  νz20 
. solution of Poisson attitude equation [3] for the transition ma-
∆fz30 0
νz3 trix L between instrumental (”z”) and geodetic (”x”) reference
frames [4]:
For errors of scaling factors (Γii , Θii ) and sensitive axes
misalignments in radians (Γij , Θij , i 6= j) L̇ = ω̂z L − Lûx , L(0) = L0 ,
 
    0 ωz3 −ωz2
Γ11 0 0 Θ11 Θ12 Θ13 ω̂z =  −ωz3 0 ωz1  ,
Γ =  Γ21 Γ22 0 , Θ =  Θ21 Θ22 Θ23 . ωz2 −ωz1 0
Γ31 Γ32 Γ33 Θ31 Θ32 Θ33 
0 ux3 −ux2


ûx =  −ux3 0 ux1  .

Here upper triangular part of Γ is forced to be zero by
ux2 −ux1 0
definition of the instrumental reference frame. This reference
has the first axis exactly matched with the sensitive axis of the Here ux stands for the Earth’s angular rate as projected onto
first accelerometer, and the second axis strictly in the plane of the local geodetic frame. Since ωz is not known, one uses
the first and the second accelerometers sensitive axes. Even sensor output ωz′ instead of ωz for computations. As a result,
though this frame is a kind of ”virtual” from some point of not L, but some different matrix L′ is obtained. This matrix
view, it is this reference that often used in inertial navigation. differs from the ideal value by small rotation term
The matrix D stands for g-sensitivity coefficients (if present),  
0 βz3 −βz2
and δfzs , δνzs — for stochastic errors.
L′ ≈ (E + β̂z )L, β̂z =  −βz3 0 βz1  .
We do not discuss here the temperature variations of sensor
βz2 −βz1 0
output since nothing new to that is going to be introduced
in this work. So any of the common approaches can be used The rotation term βz represents three small angles of rotation
to compensate for it. These approaches are usually include round each instrumental axis. For βz we then have the equation
experiments at several fixed temperatures and then linear (ignoring higher order terms)
interpolation or polynomial approximation of the results. The
β̇z ≈ ω̂z′ βz + νz .
possibility of conducting calibration experiments with non-
constant temperature is the subject of another ongoing study. This equation relates attitude and gyro errors to each other.
After that, the goal of calibration is to obtain estimates for Let’s now move to accelerometers. The only force acting on
∆fz0 , νz0 , Γ, Θ and D (with total number of components of them is the reaction force of the turntable. Every sensor should
21 or 30). These estimates will allow to compensate sensor ideally measure the projection of this force onto the sensitive
output for all error components except stochastic ones. axis. Assume then, that proof masses of accelerometers are
exactly on the axis of rotation. In this case measured force is
certainly the opposite to the gravity force. Real sensors have
errors introduced above, so we can write
   
0 0 g 0
Z ≡ (L′ )Tfz′ −0≈−g 0 0(L′ )Tβz +(L′ )T∆fz ,
g 0 0 0
Here we have some values on the left, which can be computed Fig. 2. An example of IMUs displaced from the axis of rotation
at any time, and which are linearly related to βz and ∆fz . The
information on the left will now be treated as measurements
in our model. B. Taking the displacement into consideration
Assuming the state vector X of the system
In some cases parameters of the displacement of accelerom-
T eter proof masses are known (e.g. they can be measured
X= βzT νz0T(∆fz0 )T Γ11 . . . Γ33 Θ11 . . .Θ33 D11 . . . D33 ,

directly along axes of the instrumental frame). Then ac-
and adding the system dynamic for parameters being calibrated celerometer outputs are corrected for the above tangential and
centripetal terms, and the model works as before.
β̇z ≈ ω̂z′ βz + νz , ˙ 0 = Γ̇ = Θ̇ = Ḋ = 0,
ν̇z0 = ∆f z But more interesting case is when components of s are not
we obtain linear state-space system with measurements. So known [2]. First of all, we accept that the displacement s1
that along the axis of rotation does not make any significant effect
on IMU. So we treat s1 as 0 and rename s = [0s2 s3 ]T . After
Ẋ = AX + q, Z = HX + r,
that, we add s2 and s3 to the state vector, so that
where the matrices A and H are known at any time (up to T
X= βzT νz0T(∆fz0 )TΓ11 . . .Γ33 Θ11 . . .Θ33 D11 . . .D33 s2 s3 .

the first order terms), and q and r are stochastic disturbances
with some specific characteristics. It appears then, that in the Then we modify the error model as follows
experiment outlined above (see Fig. 1) this system becomes
observable. Thus, we can use any optimal estimator (e.g. ˆ + δfzs .
∆fz = ∆fz0 + Γfz + ω̂ 2 s − ω̇s
Kalman filter) to obtain estimates for all desired parameters.
This approach, with slight variations and estimator tuning What is interesting here is that substituting the new model in
for different types of IMU, proved its efficiency in several our system gives us the new one, which is also very easy to
implementations in industry. make observable. It means that we can arrange the rotation of
the turntable in a way that components s2 and s3 could be
III. I NSTRUMENT DISPLACEMENT FROM THE AXIS OF estimated well. This was shown by processing data from real
ROTATION calibration experiments. Regarding the angular acceleration ω̇,
A. The effect of the displacement on IMU it can be computed by numerical differentiation for high-grade
IMUs. For low-grade IMUs angular accelerations should be
Models above were deduced under the assumption that avoided in experiment, and then ignored in equations.
accelerometer proof masses are located on the axis of rotation.
In general, it is not exactly the case. Due to inertia there appear C. Data processing and observability issues
an additional reaction which forces the IMU to stay on the
Although the state-space model under consideration is lin-
turntable and not to move rectilinearly. It is usually divided
ear, it is impossible to examine its observability analytically.
in two terms: tangential and centripetal force. Denoting the
That is because the system has an arbitrary function in its
displacement by a vector s in instrumental reference frame
structure, namely ωz (t). So the only way is to claim some
we get
ˆ general ideas based on system structure, and then to try making
ω̂ 2 s − ω̇s. experiments with different rotation rate profiles, and to study
From the equation above we can see that in many cases this the progress of parameter estimation in these experiments.
contribution can be ignored. For example, if ||s|| is less than First, it is fairly obvious that it is bad for observability to have
5 cm, ||ω|| does not exceed 3 degrees (1/20 radians) per the constant ω̂ 2 s term in equations, since it would mix up with
second, and ||ω̇|| is almost negligible, the magnitude of the ∆fz0 terms, which is also constant. Thus, the common thing
above error is less than 0.2 mm/s2 . is to have different angular rate magnitudes in an experiment.
But there are also lots of situations when significant dis- Second, since the rotation rate is a coefficient of displacement,
placement of accelerometer proof masses from the axis of the accuracy of estimation severely depends on the magnitude
rotation is unavoidable. The examples are calibrating several of this rate. Higher rates give higher accuracy of estimation.
IMUs at the same time on one turntable (see Fig. 2), IMUs Three inertial units are in consideration here.
√
with large spatial separation between accelerometers, particu- - Ring laser gyros (typ. performance of 0.01–0.03 ◦ / hr)
lar types of turntables, and high spin experiments. and quartz flexure suspension accelerometers (0.1 mg)
 200 50

Angular rate,
Angular rate,

degrees/sec
degrees/sec

Angular rate,
20

degrees/sec
100
0 0 0

−20 −100
−200 −50
400 600 800 1000 1200 1400 200 400 600 800 200 400 600 800
0.1 0.2
s2 estimate, m

0.3

s2 estimate, m

s2 estimate, m
3−σ boundaries 3−σ boundaries
0 3−σ boundaries 3⋅σ = 5 cm
3⋅σ = 1 mm estimate
0.2 0.1
estimate 3⋅σ = 2 mm estimate
−0.1 0.1
0
−0.2 0
400 600 800 1000 1200 1400 200 400 600 800 200 400 600 800
time, seconds time, seconds time, seconds

(a) high-grade IMU at low rates (b) medium-grade IMU at high rates (c) low-grade IMU at moderate rates

Fig. 3. Estimation progress for different types of IMU at different rotation profiles

TABLE I
assembled in navigation/tactical grade system, referenced
E STIMATION RESULTS SUMMARY
below as high-grade IMU. √
- Fiber optic gyros (0.5–1 ◦ / hr) and MEMS accelerom-
eters (1–2 mg) assembled in tactical grade system, refer- IMU grade High Medium Low
enced below as medium-grade IMU. √
- Hemispherical resonant gyros (3–5 ◦ / hr) and MEMS Gyro type RLG FOG HRG
accelerometers (1–2 mg) assembled in tactical/industrial Accelermeters quartz-flexure MEMS MEMS
grade AHRS system, referenced below as low-grade Rate magnitude 30 dps 200 dps 50 dps
Rate piecewise sine piecewise
IMU.
change 2 magnitudes 10 min period 1 magnitude
Sets of data were collected during calibration experiments Displacement 20 cm 30 cm 5 cm
similar to the described above. Each experiment had different Maximum 5 330 3.5
rotation profile and different rate magnitude. All datasets were centripetal term cm/s2 cm/s2 cm/s2
Estimate 3-σ 2 mm 1 mm 50 mm
processed by a software, which incorporates Kalman filtering
implemented in C/C++ program on PC. This software was
designed in Moscow State University Navigation and Control
Lab for calibration data processing. On Fig. 3 there are exam- is no actual need to measure the displacement prior to
ples of rotation profiles accompanied by estimation progress the experiment.
with 3-σ estimation error covariances given by estimator. - It is recommended for this situation, that the rotation rate
Results are incorporated in Table I. Actually, 3-σ values are should have tens of degrees per second magnitude and
given here just for common understanding of what orders of this magnitude should be changing. It should have at least
accuracy are in question here. In fact, real estimation accuracy 2 different nonzero values (e.g. 10 and 30 dps).
corresponds to the accuracy of the whole model, including - Angular accelerations should be avoided, since there is
all other parameters. The point is that residual accelerometer no instruments in IMU to measure it correctly.
errors caused by errors of estimation are always almost within
the calibration accuracy as if there was no any displacement. R EFERENCES
[1] N. Vavilova, A. Golovan, N. Parusnikov and I. Sazonov, Calibration of
a Strapdown Inertial Navigation System Using a Low-Accuracy Single
IV. C ONCLUSION Axis Turntable in Proc. ICINS 2010, pp. 71-72.
[2] A. Kozlov, I. Sazonov, N. Vavilova and N. Parusnikov,Calibration of an
Summarizing the experience of dealing with a considerable inertial measurement unit on a low-grade turntable with consideration of
number of different IMUs, we can state the following. spatial offsets of accelerometer proof masses in Proc. ICINS 2013, pp.
126-129.
- The above method of IMU calibration on a low grade [3] P.G. Savage, Strapdown inertial navigation integration algorithm design.
turntable works well even when the IMU is significantly Part 1: attitude algorithms in Journal of Guidance, Control, and Dynam-
displaced from the axis of rotation. This case requires to ics, vol. 21, No. 1, 1998, pp. 19-28.
[4] A.A. Golovan, O.V. Demidov, N.B. Vavilova, On GPS/GLONASS/INS
extend the mathematical model of the system. tight integration for gimbal and strapdown systems of different accuracy
- Geometrical parameters of the displacement are easily in Automatic Control in Aerospace, vol. 18, part 1, 2010, pp. 505-509.
estimated then, up to centimeter or millimeter level
depending on IMU type and rotation profile. So there