Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

Measurement Science and Technology

PAPER

A novel MEMS-RIMU self-calibration method based on gravity vector


observation
To cite this article: Jianhua Cheng et al 2021 Meas. Sci. Technol. 32 055108

View the article online for updates and enhancements.

This content was downloaded from IP address 200.130.19.206 on 01/10/2021 at 19:21


Measurement Science and Technology

Meas. Sci. Technol. 32 (2021) 055108 (12pp) https://doi.org/10.1088/1361-6501/abd798

A novel MEMS-RIMU self-calibration


method based on gravity vector
observation
Jianhua Cheng, Ping Liu, Zhenyu Wei and Guangdi Luo
College of Intelligent System Science and Engineering, Harbin Engineering University, Harbin
Heilongjiang, People’s Republic of China

E-mail: liuping@hrbeu.edu.cn

Received 7 August 2020, revised 6 December 2020


Accepted for publication 30 December 2020
Published 26 March 2021

Abstract
The microelectromechanical system (MEMS) redundant inertial measurement unit (RIMU) is a
low-cost navigation system, which can improve reliability effectively. In existing MEMS-RIMU
calibration schemes, extra equipment with higher accuracy is required. However, on some
occasions, a rotation platform or other reference equipment cannot be provided, and calibration
of the MEMS-RIMU is limited. To solve the problem, a novel self-calibration method for the
MEMS-RIMU based on gravity vector observation is proposed, which can estimate the bias,
scale errors and alignment errors of accelerometers and gyroscopes without extra equipment or
rotation platforms. In this paper, a novel calibration model for accelerometers and gyroscopes is
built by analyzing the relationship between gravity vector observation and the measurement
value of the MEMS-RIMU. There is no attitude information in the calibration model, so the
RIMU can be calibrated without attitude references via extra equipment. Then, a derivative
unscented Kalman filter is proposed to estimate the bias, scale errors and alignment errors of
accelerometers and gyroscopes, which can reduce the computation load. Finally, the simulation
and experimental results demonstrate the efficiency and accuracy of the proposed calibration
scheme.
Keywords: redundant inertial measurement unit (RIMU), calibration, unscented Kalman filter,
microelectromechanical system (MEMS)

(Some figures may appear in colour only in the online journal)

1. Introduction However, the accuracy of MEMS-RIMU will decline due to


bias, scale errors and alignment errors. Studies have shown
The inertial measurement unit (IMU) is a core part of the iner- that calibration of the MEMS-RIMU can compensate for the
tial navigation system (INS) that provides angular velocity and errors effectively [5].
specific force [1]. To improve the reliability of the IMU, a RIMU calibration is a method used to estimate the bias,
redundant inertial measurement unit (RIMU) was proposed scale errors and alignment errors by comparing the sensors’
and is widely used in INS [2]. RIMUs commonly include more information with known reference information [6]. A labor-
than three gyroscopes and three accelerometers. When one atory calibration scheme in [7] was shown to calibrate each
of the gyroscopes (or accelerometers) is broken, the RIMU sensor in the RIMU accurately. However, a high-precision
can work normally [3]. As technology has developed, micro- rotation platform is required for calibration. Therefore, this
vehicles, robots and other intelligent devices have become method cannot be used in the outer field calibration. Con-
more popular. For this purpose, microelectromechanical sys- sidering instantaneity and the calculation load, we can estim-
tem (MEMS) inertial sensors with advantages of low cost, low ate the parameters which are equivalent to the triaxial angu-
power and small size play an important role in RIMUs [4]. lar velocity and specific force instead of all the sensors in the

1361-6501/21/055108+12$33.00 1 © 2021 IOP Publishing Ltd Printed in the UK


Meas. Sci. Technol. 32 (2021) 055108 J Cheng et al

RIMU [8]. Furthermore, in the process of navigation, the angu- Furthermore, the proposed approach has the advantage of
lar velocity and specific force from the RIMU cannot be used minor calculation and higher precision.
directly. The data fusion algorithm is used to transform the The rest of this paper is organized as follows. In section 2,
multi-dimensional data which are obtained from the RIMU the redundant sensor fusion based on the weighted LS
into three-dimensional data [9]. Therefore, calibration of the algorithm for the RIMU is given. In section 3, the calibra-
equivalent triaxial bias, scale errors and alignment errors is tion models of accelerometers and gyroscopes are presented.
more significant for the out-field calibration. Section 4 establishes the dynamic model and the observation
Generally, the calibration process is divided into two parts: model, and introduces the derivative UKF algorithm. Simula-
building calibration models and estimating parameters [10]. tion and experimental results are shown in section 5. Finally,
A calibration model can be built by analyzing the relations conclusions are given in section 6.
between RIMU output and reference information. Paramet-
ers such as bias, scale errors and alignment errors, are usually
2. Redundant sensor fusion algorithm
estimated by the least squares (LS) algorithm, Kalman filter
(KF) or other estimation algorithms.
The redundant sensor fusion algorithm is designed for trans-
To achieve outer field calibration, star sensors are used to
forming the RIMU output to triaxial navigation information in
calibrate the RIMU [11]. However, as an expensive sensor, a
the body frame. The relation between the oblique sensor and
low-cost navigation system is usually not equipped with a star
body frame is shown in figure 1. According to the relation, the
sensor, such as a MEMS navigation system. A magnetometer
output of the oblique sensor can be expressed as:
is a kind of low-cost sensor that can be used for calibrating the
RIMU [12, 13]. But magnetic interference can be a problem mi = sinαi cosβi xx + sinαi sinβi xy + cosαi xz . (1)
during the calibration process. Compared with the star sensor
and magnetometer, GPS has the advantages of low cost and Assuming that there are n gyroscopes (or n acceleromet-
good stability [8]. A problem with this scheme is that when ers) in the RIMU, the observation model of the sensors can be
the GPS signal is lost or disturbed, the performance of the expressed as:
calibration will decline seriously [14]. Besides, gravity vec-
tors and the Earth’s rotation angular velocity are used in self- m = Hx + v, (2)
calibration in the RIMU as references [15–17]. However, some
low precision inertial sensors cannot be sensitive to the Earth’s  
rotation, such as the MEMS-RIMU [18]. Self-calibration of sinα1 cosβ1 sinα1 sinβ1 cosα1
the MEMS-RIMU is a key problem, especially when there is  .. .. .. 
H= . . . , (3)
no rotation platform or other navigation equipment [19].
sinαn cosβn sinαn sinβn cosαn
To estimate the parameters, the LS algorithm is widely
used [20]. However, the shortcomings of the LS algorithm
where m = [m1 , m2 , . . . , mn ]T , mi denotes the output of ith
are low precision and poor adaptability of the non-linear
sensor. Here,v = [v1 , v2 , . . . , vn ]T , and vi denotes the observa-
model. To solve this problem, numerous non-linear estimation
tion noise of the ith sensor. H represents the configuration
algorithms are presented, such as the extended Kalman filter
matrix of the RIMU. Here, x = [xx , xy , xz ]T denotes the triaxial
(EKF), unscented Kalman filter (UKF), neural networks, and
angular velocity (or specific force) in the body frame [27]. We
so on [21–24]. Among the estimation algorithms, the UKF is
define the covariance matrix of the observation noise as:
a non-linear estimation algorithm with highperformance [25].
Due to the fact that the dynamic model of calibration is lin-
R = diag([δ1 , δ2 , . . . , δn ]). (4)
ear and the observation model is non-linear, the unscented
transformation (UT) for the dynamic model will increase the The error of sensor fusion can be written as:
amount of calculation. In [26], a derivative unscented Kalman
filter (DUKF) is proposed to reduce the amount of computa- e = m − Hx̂. (5)
tion for integrated navigation. We can also apply the algorithm
to estimate parameters in the RIMU calibration area.
According to the analysis above, a novel self-calibration To estimate x, the weighted LS algorithm is used. The prin-
algorithm for the MEMS-RIMU based on gravity vector ciple of the weighted LS algorithm is to minimize the sum of
observation is proposed. Firstly, the redundant sensor fusion squares [28]. We define
algorithm is proposed to obtain the accelerometer and rota-
tion velocity of the system in the body frame. Secondly, an X
n

accelerometer calibration model is built by analyzing rela- F(x̂) = e2i /δi2


tions between gravity vectors and accelerometer output. Then, i =1
T −1
according to the relations between the accelerometer and gyro- =e R e
scope output, a gyroscope calibration model can be estab- = (m − Hx̂)T R−1 (m − Hx̂)
lished. Finally, a dynamic model and observation model
are given for estimating parameters by the DUKF. During = mT R−1 m − x̂T HT R−1 m
the calibration process, we do not need extra references. − mT R−1 Hx̂ + x̂T HT R−1 Hx̂. (6)

2
Meas. Sci. Technol. 32 (2021) 055108 J Cheng et al

Earth yb
Rotation zb
yn zn

xb
xn
o
g
o

Figure 1. The relation between the oblique sensor and body frame. Figure 2. The frames used for calibration.

Minimizing the sum of squares in (11), we can obtain: at = Ca (I + Sa )am − bta − ϵm , (11)

−mT R−1 H + x̂T HT R−1 H = 0, (7) where am denotes the accelerometer’s output after sensor
fusion in section 2, at is the true gravity vector and ϵm is a
x̂ = (HT R−1 H)−1 HT R−1 m. (8) zero mean white noise vector of the accelerometer. Here, bta
denotes the bias matrix of the accelerometer, Ca represents the
Compared with the LS algorithm, the observation noise is scale error matrix, Sa is the scale error matrix, and bta , Ca and
considered in the weighted LS algorithm. According to (8), the Sa can be expressed as follows, respectively:
gyroscope fusion and accelerometer fusion of the RIMU can
 T
be expressed as follows: bta = bax bay baz , (12)

ω m = (HT R−1 H)−1 HT R−1 ω, (9)


 
1 0 0
Ca =  θxy 1 0 , (13)
am = (HT R−1 H)−1 HT R−1 a. (10) θxz θyz 1
To reduce the computation load and improve the real-time  
performance of the calibration scheme, we do not need to cal- Sax 0 0
ibrate every sensor in the RIMU. We can design an appropriate Sa =  0 Say 0 . (14)
calibration algorithm to calibrate the data after sensor fusion, 0 0 Saz
such as ω m and am .
According to (13) and (14),
 
3. System modeling based on gravity vector Sax 0 0
observation Ca (I + Sa ) = I +  (1 + Say )θxy Say 0 .
(1 + Saz )θxz (1 + Saz )θyz Saz
In this section, the calibration models of accelerometers and (15)
gyroscopes are introduced by analyzing the relationships
between gravity vector and inertial sensor data after sensor To simplify the calculation process, the matrix of scale
fusion. The introduced frames are shown in figure 2. errors and alignment errors can be defined as:

• Navigation frame (n-frame). The navigation frame is a local  


Sax 0 0
geographic frame [29]. The frame origin is set at the loca- Ka =  kxy Say 0 , (16)
tion of the vehicle. Its axes are aligned with the directions kxz kyz Saz
of the north, east and local vertical.
• Body frame (b-frame). The body frame is fixed to the where kxy = (1 + Say )θxy , kxz = (1 + Saz )θxz , kyz = (1 +
vehicle [30]. The origin o is set at the mass center of the Saz )θyz .
vehicle. The axes xb , yb and zb are aligned with the roll, Then, the accelerometer model in the b-frame can be
pitch and yaw axes of the vehicle, respectively. written as:

abt = (I + Ka )abm − bta − ϵm . (17)


3.1. Accelerometer calibration model

For the calibration of accelerometers, a 9-parameter According to the accelerometer model , the calibration
accelerometer model is widely used [31]. The 9-parameter model for the accelerometer can be built. From (11), it is
accelerometer model can be written as: known that at can be obtained from the gravity field model

3
Meas. Sci. Technol. 32 (2021) 055108 J Cheng et al

of the Earth. We define the local gravity vector in the n-frame where ωxb , ωyb and ωzb denote the true angular velocity of the
as ant , and it can be expressed as: vehicle. Then, the relation between the accelerometer and
gyroscope measurements can be written as:
ant = Cnb abt , (18)
(I + Ka )ȧbm = −[ω×]bn Cbn ant + Cbn ȧnt + ϵ̇m . (27)
where Cnbis the transition matrix form b-frame to n-frame. The
norm of ant satisfies: Substituting (17) into (27):

2 2 Cbn ȧnt = (I + Ka )ȧbm + [ω×]bn Cbn ant − ϵ̇m


∥ant ∥ = Cnb abt = (Cnb abt )T (Cnb abt ) = abt .
2
(19)
= (I + Ka )ȧbm + [ω×]bn [(I + Ka )abm − bta ]
Substituting (17) into (19), the norm of ant can be expressed as: − [ω×]bn ϵm − ϵ̇m . (28)

2
To simplify the calculation process, the following parameters
∥ant ∥ = [(I + Ka )abm − bta − ϵm ]T [(I + Ka )abm − bta − ϵm ] are defined:
) (I + Ka )T − (bta )T − ϵTm ][(I + Ka )abm − bta − ϵm ]
b T
= [(am
2 m1 = (I + Ka )ȧbm , (29)
= abm + (abm )T (KTa + Ka + ∥Ka ∥ )(abm ) + ∥bta ∥
2 2

− (abm )T (I + Ka )T (bta ) − (bta )T (I + Ka )(abm )


m2 = [ω×]bn [(I + Ka )abm − bta ], (30)
− [(abm )T (I + Ka )T − (bta )T ]ϵm
2
− ϵTm [(I + Ka )abm − bta ] + ∥ϵm ∥ . (20)
m3 = m1 + m2 . (31)
Then, the calibration model of the accelerometer can be
expressed as: According to (28), the norm of Cbn ȧnt satisfies:

2 b n 2
Cn ȧt = (Cbn ȧnt )T (Cbn ȧnt )
∥ant ∥ − abm = (abm )T (KTa + Ka + ∥Ka ∥ )(abm )
2 2

2 = (ȧnt )T (Cbn )T Cbn ȧnt


+ ∥bta ∥ − (abm )T (I + Ka )T (bta )
2
− (bta )T (I + Ka )(abm ) + ξ a , (21) = ∥ȧnt ∥ , (32)

2
∥ȧnt ∥ = m1 + m2 − [ω×]bn ϵm − ϵ̇m
2
ξa = [(abm )T (I + Ka )T − (bta )T ]ϵm
2 2
2 = ∥m1 ∥ + ∥m2 ∥ + mT1 m2 + mT2 m1 + ξ g , (33)
− ϵTm [(I + Ka )abm − bta ] + ∥ϵm ∥ . (22)

2
ξ g = [ω×]bn ϵm + ∥ϵ̇m ∥ − mT3 ([ω×]bn ϵm + ϵ̇m )
2
3.2. Gyroscope calibration model

The gyroscope calibration model can be obtained by the accel- − ([ω×]bn ϵm + ϵ̇m )T m3 + ([ω×]bn ϵm )T ϵ̇m
erometer model. According to (17) and (18), we can obtain: + ϵ̇Tm [ω×]bn ϵm . (34)

(I + Ka )abm = Cbn ant + bta + ϵm . (23) From (33), the gyroscope calibration model can be obtained:
2
∥ȧnt ∥ − ȧbm = (ȧbm )T (Ka + KTa + ∥Ka ∥ )ȧbm
2 2
The differential of (23) can be written as:
2
(I + Ka )ȧbm = Ċbn ant + Cbn ȧnt + ϵ̇m . (24) + ∥m2 ∥ + mT1 m2 + mT2 m1 + ξ g . (35)

The gyroscope model in the body frame can be written as:


According to the differential equation of the direction cosine
matrix, as the carriers are slower than the sampling rate of the ω bm = Kg ω bt + btg + ϵg , (36)
gyroscopes, Cbn can be approximately written as:
where ω bt denotes the true angular velocity of the vehicle,
Ċbn ≈ −[ω×]bn Cbn , (25) ω bm represents the gyroscope output after sensor fusion
in section 2, btg is the bias of the gyroscope, and Kg =
where [ω×]bn is an antisymmetric matrix which can be
diag(k− 1 −1 −1
gx , kgy , kgz ) are the scale errors of the gyroscope. The
expressed as:
gyroscope calibration model can be expressed as:
 
0 −ωzb ωyb 2
∥ȧnt ∥ − ȧbm = (ȧbm )T (Ka + KTa + ∥Ka ∥ )ȧbm
2 2
[ω×]bn =  ωzb 0 −ωxb  , (26)
2
−ωyb ωxb 0 + ∥m2 ∥ + mT1 m2 + mT2 m1 + ξ g , (37)

4
Meas. Sci. Technol. 32 (2021) 055108 J Cheng et al

  wm
0 −ωtzb ωtyb 0 = κ/(n + κ), i =0
m2 =  ωtzb 0 −ωtxb  · [(I + Ka )abm − bta ], (38) wc0 = κ/(n + κ) + (1 − α + β), i = 0
2 (45)
−ωtyb ωtxb 0 wm = wci = κ/[2(n + κ)], i = 1, . . . , 2n
i

where ωtxb = kgx (ωmx


b
− btgx ), ωtyb = kgy (ωmy
b
− btgy ), ωtzb = where w denotes the weight coefficient of the sigma point.
kgz (ωmz − bgz ).
b t
(b) State prediction: due to the fact that the dynamic model is
linear, the state prediction process can be calculated as the
4. Parameter estimation KF. Compared with the traditional UKF, the calculation
can be reduced.
4.1. Dynamic model and observation model
xk|k−1 = xk−1|k−1 , (46)
According to the analysis in sections 2 and 3, the bias, scale
errors and alignment errors can be estimated as states. The
dynamic model can be expressed as: Pxx xx
k|k−1 = Pk−1|k−1 . (47)

K̇a = 0, (39) Measurement prediction: the current measurements and


their covariance can be expressed as:

ḃta = 0, (40) zik|k−1 = h(χi,k−1|k−1 ), (48)

ḃtg = 0. (41) X
2n
zk|k−1 = wm i
i zk−1|k−1 , (49)
The observation model can be written as: i=0
2
∥ant ∥ − abm = (abm )T (KTa + Ka + ∥Ka ∥ )(abm )
2 2
X
2n
2 Pzz
k|k−1 = wci (zik−1|k−1 − zk−1|k−1 )·
+ ∥bta ∥ − (abm )T (I + Ka )T (bta )
i=0
− (bta )T (I + Ka )(abm ) + ξ a , (42) (zik−1|k−1 − zk−1|k−1 )T , (50)

2
∥ȧnt ∥ − ȧbm = (ȧbm )T (Ka + KTa + ∥Ka ∥ )ȧbm
2 2
X
2n
2 Pxz
k|k−1 = wci (χi,k−1|k−1 − xk−1|k−1 )·
+ ∥m2 ∥ + mT1 m2 + mT2 m1 + ξg . (43)
i=0
(zik−1|k−1 − zk−1|k−1 )T . (51)
4.2. DUKF algorithm

In the above, it is known that the observation model is a strong (c) Upgrading: the parameters can be estimated as:
non-linear model, and the Jacobian matrix for the EKF is dif- −1
Kk = Pxz xz
k−1|k−1 (Pk−1|k−1 ) , (52)
ficult to calculate. The UKF can estimate parameters by UT
transformation, and the Jacobian matrix is unnecessary for
the UKF [32]. In the UKF algorithm, sigma points and their
xk = xk|k−1 + Kk (zk − zk|k−1 ), (53)
weight coefficients are required during the state prediction and
measurement prediction. According to (39)–(41), it is known
that the dynamic model is linear. The UT transformation for
k = Pk|k−1 − Kk Pk|k−1 Kk .
Pxx xx zz T
(54)
sigma points during the state prediction leads to a great amount
of computation load. To solve the problem, a DUKF algorithm
Repeat the above steps until the estimated values converge.
is proposed for the linear dynamic model and non-linear obser-
Then, the bias, scale errors and alignment errors can be calcu-
vation model.
lated, respectively.
According to the UKF in [33], the recursive process of the
DUKF can be expressed as follows:
5. Simulation and experimental results
(a) Sigma points and their weight
5.1. Simulation results of accelerometer calibration
χ0 = x̂, i =0
p In the real RIMU, the values of the bias, scale errors and align-
χi = x̂ + ( (n + κ)Px )i , i = 1, . . . , n (44) ment errors are unknown. Therefore, the accuracy of the pro-
p
χi = x̂ − ( (n + κ)Px )i−n , i = n + 1, . . . , 2n posed calibration scheme cannot be verified through the exper-
iment alone. Hence, simulation is an effective method to verify
where χ represents the sigma point, κ = α2 (n + λ) − n. the performance of the proposed calibration algorithm.

5
Meas. Sci. Technol. 32 (2021) 055108 J Cheng et al

Table 1. Configuration angles of the RIMU. Table 3. The simulation results of the accelerometers.

Sensor 1 2 3 4 Values bta (m s−2 ) Ka


   
α(deg) 180 70.53 70.53 70.53 0.1 0.1 0 0
β(deg) 0 330 210 90 References  0.2   0.1 0.2 0 
0.3 0.2 0.1 0.3
   
Table 2. The simulation settings of the accelerometers. 0.138 0.096 0 0
Proposed  0.225   0.086 0.19 0 
Items Values
0.293 0.182 0.084 0.299
Frequency (Hz) 100    
[ ]T 0.145 0.120 0 0
bta (m s−2 ) 0.1 0.2 0.3  0.235   0.116
  LS 0.230 0 
0.1 0 0 0.345 0.214 0.114 0.324
Ka  0.1 0.2 0 
0.2 0.1 0.3
 
0 0 −1
 −0.8165 −0.4714 0.333  proposed
 

2
H 1 reference
 0.8165 0.333 

b /m/s
−0.4714 LS
0.5
0 0.9428 0.333

x
0
Noise (m s−2 (rms)) 0.2 0 50 100 150 200 250
t/s
0.2
2

0
b /m/s

-0.2
-0.4
y

-0.6
0 50 100 150 200 250
t/s
1
2
b /m/s

0.5
z

0
0 50 100 150 200 250
Figure 3. The generated accelerometer track: (a) accelerometer t/s
measurements, and (b) accelerometer measurement after sensor
fusion.
Figure 4. Estimated accelerometer bias.

We defined the bias, scale errors, alignment errors and noise 1


proposed
for every sensor first. Table 1 shows the configuration angles reference
Sax

of the RIMU. The configuration matrix H can be calculated by 0.5 LS

the configuration angles. Table 2 shows the values of the sim-


0
ulation settings. Then, the generated accelerometer measure- 0 50 100 150 200 250
ments are shown in figure 3(a). Figure 3(b) shows the gener- t/s
ated accelerometer measurements after sensor fusion. We will 0.2
0
calibrate the accelerometers by using the measurements after
Say

-0.2
sensor fusion. We also try to imitate the traditional LS cal- -0.4
ibration method, and compare the calibration results with the
0 50 100 150 200 250
proposed method. t/s
The calibration results are shown in figures 4–6; the red
0.3
line is the reference value, the blue line is the estimated results
Say

0.2
based on the proposed calibration scheme and the green line 0.1
is the results of the traditional LS method. The estimated res- 0
ults of the proposed scheme and the traditional LS method can 0 50 100 150 200 250
t/s
converge to the reference values. To compare the accuracy of
the proposed scheme and traditional LS method more clearly, Figure 5. Estimated accelerometer scale errors.
the estimated bias, scale errors and alignment errors are given
in table 3. According to table 3, the performance of the pro-
posed scheme is better than that of the LS; the accuracy of the
5.2. Simulation results of gyroscope calibration
estimated bias and scale errors has increased by at least 18 per-
cent. The simulation results show that the accelerometers can The gyroscopes can be calibrated through the calibrated
be calibrated accurately. accelerometers. In the simulation of gyroscope calibration,

6
Meas. Sci. Technol. 32 (2021) 055108 J Cheng et al

(a)
0 10
K xy /deg

ax/m/s2
-0.2 proposed 0

-0.4 reference −10


0 100 200 300 400 500
-0.6 LS t/s
10

ay/m/s2
0 50 100 150 200 250 0

t/s −10
0 100 200 300 400 500
t/s
10
K xz /deg

az/m/s2
-0.2 0

−10
-0.4 0 100 200
t/s
300 400 500

-0.6
0 50 100 150 200 250 Figure 8. Generated measurements after fusion: (a) accelerometer
t/s measurements after fusion, and (b) gyroscope measurements after
0.6 fusion.
K yz /deg

0.4
0.2
0.05 Estimated Reference
0

bgx/rad/s
0 50 100 150 200 250
0
t/s
−0.05
0 100 200 300 400 500
Figure 6. Estimated accelerometer alignment errors. t/s

bgy/rad/s 0.05
Table 4. The simulation settings of the gyroscopes.
0
Items Values
−0.05
0 100 200 300 400 500
Frequency (Hz) 100
[ ]T
t/s

btω (deg s−1 ) 0.1 0.2 0.3 0.05


Noise (m s−2 (rms))
bgz/rad/s

0.2
0

−0.05
(a) 0 100 200 300 400 500
10
a /m/s2

t/s
0
1

−10
0 100 200 300 400 500
t/s
10 Figure 9. Estimated bias of the gyroscopes.
a /m/s2

0
2

−10
0 100 200 300 400 500
t/s
10
a /m/s2

0
1.1
3

−10
0 100 200 300 400 500
kgx

t/s
10
1.05
a /m/s2

0
Estimated Reference
4

−10
0 100 200 300 400 500
t/s 1
0 100 200 300 400 500
t/s

Figure 7. Generated measurements: (a) accelerometer


measurements, and (b) gyroscope measurements. 1
kgy

0.95
0.9
assuming that the accelerometer has been calibrated accur- 0 100 200 300 400 500
t/s
ately, the configuration of the gyroscopes is the same as for
the accelerometers. Table 4 shows the values of the simulation
settings of the gyroscopes. Then, the generated gyroscope and 1.1
gz
k

accelerometer measurements are shown in figure 7. Figure 8


1
shows the generated gyroscope and accelerometer measure- 0 100 200 300 400 500
ments after sensor fusion. The calibration scheme is designed t/s

for the RIMU measurement after sensor fusion.


Figure 10. Estimated scale errors of the gyroscopes.
Figures 9 and 10 show the calibration process of gyroscope
bias and scale errors; the red line is the reference values, and
the blue line is the estimated results. The estimated results con-
5.3. Experimental condition
verge to the reference values in 500 s. Moreover, we try to
imitate the cross product (CP) calibration algorithm in [34]. To verify the practical application of the proposed scheme, an
The estimated values of gyroscope bias and scale errors are RIMU with a tetrahedron structure, presented in figure 11(a),
given in table 5. According to the simulation results, the gyro- is used in the experiment. There are four gyroscopes and four
scopes bias and scale errors can be estimated accurately. Most accelerometers in the RIMU. The configuration angles of the
of the parameters estimated by the proposed scheme are more accelerometers and gyroscopes are the same as the simulation,
accurate than those by the CP method. which is shown in table 1.

7
Meas. Sci. Technol. 32 (2021) 055108 J Cheng et al

Table 5. The simulation results of the gyroscopes. Table 6. The RIMU parameters.

Values btg (deg s−1 ) K gx K gy K gz Items Values


 
0.0017 Frequency (Hz) 100
References  0.0034  1.1 0.9 1.15 Gyros range ±300 (deg s−1 )
0.0051 Gyros resolution 800 (LSB deg−1 s−1 )
  Gyros bias <0.5 (deg s−1 )
0.0016 √ −1
Proposed  0.0035  1.0996 0.9000 1.1498 Gyros noise 0.015 (deg s−1 Hz )
0.0053 Acce range ±2 (g)
  Acce resolution 256 (LSB g−1 )
0.0018
CP  0.0035  1.1102 0.8987 1.1760 Acce bias ±250 (mg)
0.0053 Acce noise  1.1 (LSB rms) 
0 0 −1
 −0.8165 −0.4714 0.333 
H  
 0.8165 −0.4714 0.333 
0 0.9428 0.333

0.8
bx
Accelerometer bias/m/s 2 0.6 by
bz
0.4
(c) zb 0.2

0
g
-0.2
Sensor 3
Sensor 4 -0.4

Sensor 2 -0.6
0 500 1000 1500 2000
yb Sampling point

Figure 12. Estimated bias of the accelerometer.


Sensor 1
xb
0.1
Sax
Say
Figure 11. The experimental equipment: (a) the experimental
Accelerometer scale errors

0 Saz
RIMU, (b) the rotation platform, and (c) the structure of the RIMU.

-0.1
The gyroscopes and accelerometers used in the RIMU
are ADXRS354 and ADXL345, respectively. The paramet- -0.2
ers are given in table 6. Moreover, a rotation platform, shown
in figure 11(b), is used to provide references. As shown in
-0.3
figure 11(b), the RIMU is fixed in a black box and the box
is installed on the rotation platform. The performance of the
proposed scheme can be verified by comparing the calibration -0.4
0 500 1000 1500 2000
results and the reference given by the rotation platform. The Sampling points
structure of the RIMU is described in figure 11(c); xb , yb and
zb are the rotation axis of the rotation platforms. Figure 13. Estimated scale errors of the accelerometer.

5.4. Experimental results


the proposed scheme and LS algorithm are given in table 7.
During the calibration process, the RIMU rotates, as in The calibration results of the gyroscopes by the proposed
figures 3, 7 and 8. Then, we can obtain the estimation results scheme and CP method are shown in table 8.
shown in figures 12–16. We also tried to calibrate the accel- Due to the fact that the real values of bias, scale errors
erometers using the traditional LS algorithm, and calibrated and alignment errors of accelerometers and gyroscopes are
the gyroscopes using the CP method. The values of bias, scale unknown, we cannot verify the performance of the calibration
errors and alignment errors of the accelerometers estimated by methods. To solve the problem, we calibrated the RIMU using

8
Meas. Sci. Technol. 32 (2021) 055108 J Cheng et al

2 Table 7. The experimental results for the accelerometers.


Kxy
bta (m s−2 )
Accelerometer align errors/deg

1.5 Kxz Values Ka


Kyz
1    
0.208 −0.047 0 0
0.5 Proposed  0.130   0.342 −0.058 0 
0.234 0.316 0.695 −0.041
0
   
-0.5 0.176 −0.0295 0 0
LS  0.122   0.538 −0.035 0 
-1 0.271 0.371 0.645 −0.051
-1.5

-2
0 500 1000 1500 2000
Table 8. The experimental results for the gyroscopes.
Sampling points
Values Proposed CP
Figure 14. Estimated alignment errors of the accelerometer.    
−0.169 −0.219
btg (deg s −1
)  −0.439   −0.414 
b b b −0.337 −0.393
gx gy gz
0.5
kx 0.992 0.979
ky 0.994 0.995
kz 0.984 0.970
Gyro bias/deg/s

-0.5

ω̃ bm = Kg (ω bm − btg ). (56)
-1
where ãbm is the triaxial accelerometer output after calibration,
and ω̃ bm is the triaxial gyroscope output after calibration.
-1.5
By substituting (2) into (55) and (56), respectively, the four-
0 100 200 300 400 500
dimensional accelerometer output and gyroscope output after
t/s
calibration can be expressed as:
Figure 15. Estimated bias of the gyroscopes.
ã = Hãbm , (57)

1.2
k gx k gy k gz
ω̃ = Hω̃ bm . (58)
1.15

1.1 Then, two groups of experiments using the rotation plat-


form and experimental RIMU are given. The experimental
Gyro scale error

1.05
RIMU is installed on the rotation platform, as shown in
1 figures 11(b) and (c). In the first accelerometer experiment,
the RIMU rotates 90 deg around yb . Then, the RIMU remains
0.95
stationary. Figure 17(a) shows the orientation of the acceler-
0.9 ometers in the first experiment. In the second accelerometer
experiment, the RIMU rotates 90 deg around xb ; then, the
0.85
RIMU rotates around zb , and the angular velocity is 10 deg s−1 .
0.8 Figure 17(b) shows the orientation of the accelerometers in the
0 100 200 300 400 500 600
t/s second experiment. In the first gyroscope experiment, the axes
of the rotation platform coincide with the local geographic
Figure 16. Estimated scale errors of the gyroscopes. coordinate system, and the system is stationary. The orient-
ation of the gyroscopes is shown in figure 11(c). In the second
gyroscope experiment, the orientation is the same as the first
the estimated parameters, and compared the accuracy before gyroscope experiment. Then, the system rotates around the
and after calibration. z-axis of the rotation platform, and the angular velocity is
Accelerometer output and gyroscope output after calibra- 20 deg s−1 .
tion can be expressed as: Figures 18 and 19 show the results before and after cal-
ibration. The reference value is given by the rotation plat-
form. The four-dimensional outputs of the accelerometers and
ãbm = (I + Ka )abm − bta , (55) gyroscopes are calculated by (57) and (58). We can obtain

9
Meas. Sci. Technol. 32 (2021) 055108 J Cheng et al

uncalibrated Proposed Reference LS

f m1 /m/s 2
0.4
0.2
0
-0.2
-0.4
-0.6
0 50 100 150
t/s

f m2 /m/s 2
-8
-8.5
-9
0 50 100 150
t/s
9

f m3 /m/s 2
8.5
8
0 50 100 150
t/s
0.4

f m4 /m/s 2
0.2
0
-0.2
-0.4
-0.6
0 50 100 150
t/s

uncalibrated Proposed Reference LS

f m1 /m/s 2
0.2
0
-0.2
-0.4
-0.6
0 50 100 150
t/s
10 10
f m2 /m/s 2

9
0 8
7
-10 25 30 35
0 50 100 150
t/s
10 9
f m3 /m/s 2

Figure 17. Orientation of the accelerometers: (a) the first 5 8


accelerometer experiment, and (b) the second accelerometer 0
-5 7
experiment. 25 30 35 25
0 50 100 150
t/s
10
f m4 /m/s 2

9
Table 9. The accelerometer RMSE before and after calibration. 5
0 8
-5
◦ −1 7
Angular velocity ( s ) 0 Z-axis 20 25 30 35
0 50 100 150
t/s
Acce 1 Uncalibrated 0.4262 0.3600
LS 0.2131 0.1714
Figure 18. Accelerometer data before and after calibration: (a) the
Proposed 0.0772 0.0956 rotation velocity is 0, and (b) the rotation velocity of axis zb is
Acce 2 Uncalibrated 0.5611 0.4372 10 deg s−1 .
LS 0.2345 0.2736
Proposed 0.0721 0.1573
Acce 3 Uncalibrated 0.5919 0.3557 Table 10. The gyroscope RMSE before and after calibration.
LS 0.1999 0.1418
Proposed 0.0693 0.1116 Angular velocity (◦ s−1 ) 0 Z-axis 20
Acce 4 Uncalibrated 0.3961 0.5692
Gyro 1 Uncalibrated 0.2268 0.4945
LS 0.1786 0.2632
CP 0.1362 0.3860
Proposed 0.0964 0.1136
Proposed 0.1280 0.1324
Gyro 2 Uncalibrated 0.2006 0.1666
CP 0.1323 0.1379
the calibration accuracy by comparing the reference values Proposed 0.1297 0.1299
with the sensor measurement values. In figures 18 and 19, Gyro 3 Uncalibrated 0.1326 0.1956
the blue line represents the measurement value before calib- CP 0.1288 0.1804
ration, the orange line is the measurement value after calibra- Proposed 0.1279 0.1254
Gyro 4 Uncalibrated 0.3348 0.4209
tion via the proposed scheme, the yellow line is the reference
CP 0.1294 0.1696
given by the rotation platform and the green line represents the Proposed 0.1290 0.1307
estimated results of the traditional method. To compare the
calibration results more clearly, the root mean square errors
(RMSEs) of each calibration method are shown in table 9 and
table 10. Compared to the RMSEs of the accelerometers and of the proposed scheme is better than those of the LS method
gyroscopes before calibration, we can see that the RMSEs of and CP method. Compared with the LS method, the accelero-
the accelerometers and gyroscopes after calibration reduce at meter accuracy of the proposed method improved 20% at least.
least 65% and 20%, respectively. In addition, the performance Compared with the CP method, the gyroscope accuracy of the

10
Meas. Sci. Technol. 32 (2021) 055108 J Cheng et al

Uncalibrated CP Proposed References accurately and effectively. Compared with the RIMU which
0.6
0.4
0.2
has not been calibrated, the accuracy of the calibrated accel-
0
-0.2 erometers and gyroscopes is improved by 65% and 20%,
-0.4
0 50 t/s 100 150 respectively.
0.6
0.4
0.2
0
-0.2
-0.4
Acknowledgments
0 50 100 150
t/s
0.4 This work was supported in part by the National Nature
0.2
0 Science Foundation of China under Grant No. 61633008, in
-0.2
-0.4 part by Heilongjiang Outstanding Youth Foundation under
0 50 t/s 100 150 Grant No. JJ2018JQ0059, in part by Fundamental Research
0.2 Funds Central Universities HEUCFP201768, in part by the
0
-0.2 National Nature Science Foundation of China under Grant No.
-0.4 61773132, and in part by the 7th Generation Ultra Deep Water
0 50 100 150
t/s Drilling Unit Innovation Project.
Uncalibrated CP Proposed References
21
20 ORCID iD
19
0 50 100 150
t/s Ping Liu  https://orcid.org/0000-0001-6692-6066
-6
-6.5
-7 References
0 50 100 150
t/s
-6 [1] Qin Y, Zhang H and Wang S 2015 Theory of Kalman Filter
-6.5 and Integrated Navigation (Xi‘an: Northwestern
-7 Polytechnical University)
0 50 100 150 [2] Cheng J, Dong J, Landry R and Chen D 2014 A novel optimal
t/s
configuration form redundant MEMS inertial sensors based
-6.5 on the orthogonal rotation method Sensors 14 13661–78
-7 [3] Cheng J, Sun X, Liu P and Mou H 2019 An improved residual
-7.5 chi-square test fault isolation approach in four-gyro SINS
0 50 t/s 100 150 IEEE Access 7 174400–11
[4] Iuri F, Federico P and Alberto B 2012 Autocalibration of
Figure 19. Gyroscope data before and after calibration: (a) the triaxial MEMS accelerometers with automatic sensor model
rotation velocity is 0, and (b) the rotation velocity of axis zb is selection IEEE Sens. J. 12 2100–08
20 deg s−1 . [5] Dai X 2017 Research on the key technology of
disassembly-free calibration method for missile-borne SINS
MS Thesis Southeast Univ., Nanjing, China
[6] Cao T 2012 On-line alignment and calibration technique of
proposed method improved about 5%. The experimental res-
fiber optic gyroscope SINS PhD Dissertation Harbin Eng.
ults show that the proposed scheme can estimate bias, scale Univ., Heilongjiang, China
errors and alignment errors accurately and can calibrate the [7] Cheng J, Liu P, Gao P, Zou M and Fu W 2019 High-precision
RIMU effectively. calibration scheme for RIMU IEEE Access 7 72376–86
[8] Liang H 2011 Key technique of micro inertial system based on
redundant gyroscopes PhD Dissertation Harbin Eng. Univ.,
6. Conclusion Heilongjiang, China
[9] Liang H, Liang J, Meng G, Xu H and Lv Z 2016 Method of
In this paper, a novel RIMU self-calibration method is laboratory calibration for redundant gyroscope unit Aerosp.
Control 34 9–14
proposed, which requires neither any navigation equipment [10] Zhang H, Wu Y, Wu W, Wu M and Hu X 2010 Improved
nor a rotation platform. Firstly, a redundant sensor fusion multi-position calibration for inertial measurement units
algorithm is proposed to obtain the RIMU measurement Meas. Sci. Technol. 21 015107
value in the body frame, which can simplify the calibration [11] Yang H, Zhang S and Cai H 2010 An in-flight calibration for
model. Then, the calibration models of gyroscopes and accel- redundant inertial measurement gyroscope J. Astronaut.
31 104–10
erometers based on gravity vector observation are presen- [12] Wu Z and Wang W 2018 Magnetometer and gyroscope
ted. Based on this model, the attitude information is not calibration method with level rotation Sensors 18 748–65
required; therefore, a rotation platform is not required in [13] Wu Y and Pei L 2017 Gyroscope calibration via magnetometer
the proposed scheme. Finally, the DUKF is proposed to IEEE Sens. J. 17 5269–75
estimate the parameters of the RIMU, which can reduce [14] Wu Z 2019 Research on INS/magnetometer integrated
positioning technology for land vehicle PhD Dissertation
the computation load and improve the accuracy. To verify Harbin Eng. Univ., Heilongjiang, China
the performance of the proposed method, simulation and [15] Ye Y, Ahmadreza A, Branko G C, Hung T N and Steven W S
experimental results were given. The RIMU can be calibrated 2017 Online auto-calibration of triaxial accelerometer with

11
Meas. Sci. Technol. 32 (2021) 055108 J Cheng et al

time-variant model structures Sensors Actuators A [25] Manuel G, Dayo O, Patrick S and Yiannos M 2014 Real-time
266 294–307 autocalibration of MEMS accelerometers IEEE Trans.
[16] Ye Y, Guo Y and Steven W S 2017 An efficient autocalibration Instrum. Meas. 63 96–105
method for triaxial accelerometer IEEE Trans. Instrum. [26] Hu G, Gao S and Zhong Y 2015 A derivative UKF for tightly
Meas. 66 2380–90 coupled INS/GPS integrated navigation ISA Trans.
[17] Fong W T, Ong S K and Nee A Y C 2018 Methods for in-field 56 135–44
user calibration of an inertial measurement unit without [27] Shim D and Yang C 2009 Optimal configuration of redundant
external equipment Meas. Sci. Technol. 19 085202 inertial sensors for navigation and FDI performance Sensors
[18] Jia C and Brian L E 2014 Online camera-gyroscope 10 6497–512
autocalibration for cell phones IEEE Trans. Image Process. [28] Dai X 2015 The key technology research on redundant inertial
23 5070–81 navigation system PhD Dissertation Harbin Eng. Univ.,
[19] Iuri F, Federico P and Alberto B 2009 Autocalibration of Heilongjiang, China
MEMS accelerometers IEEE Trans. Instrum. Meas. [29] Zhao L, Guan D, Rene L, Cheng J and Kostyantyn S 2015 An
58 2034–41 accurate and fault-tolerant target positioning system for
[20] Syed Z F, Aggarwal P, Goodall C, Niu X and El-Sheimy N buildings using laser range NDERS and low-cost
2007 A new multi-position calibration method for MEMS MEMS-based MARG Sensors Sensors 15 27060–86
inertial navigation systems Meas. Sci. Technol. [30] Peng X, Chen Y, Li J, Yan G and Zhang T 2013 Study on
18 1897–07 calibration method of MEMS 3-axis digital gyroscope
[21] Ren C, Liu Q and Fu T 2015 A novel self-calibration method Transducer Microsyst. Technol. 32 63–8
for MIMU IEEE Sens. J. 15 5416–22 [31] Martin S, Pavel P, Jan R and Petr N 2012 Analyses of triaxial
[22] Fang J and Yang S 2011 Study on innovation adaptive EKF for accelerometer calibration algorithms IEEE Sens. J.
in-flight alignment of airborne POS IEEE Trans. Instrum. 12 1157–65
Meas. 60 1378–88 [32] Xiong K, Zhang H and Chan C 2006 Performance evaluation
[23] Zhang X 2015 The modification of second-order extended of UKF-based nonlinear filtering Automatica 42 061103
Kalman filter for non-linear system MS Thesis Harbin Inst. [33] Cheng J, Liu P and Kang Y 2019 A novel calibration
Technol., Heilongjiang, China algorithm for RIMU based on derivative UKF Proc. 38th
[24] Katarina K, Miroslav L, Dusan P and Viktor K 2014 Chinese Conf. pp 4032–6
Attitude-independent 3-axis accelerometer calibration [34] Xiang L and Zhi L 2014 Vector-aided in-field calibration
based on adaptive neural network Proc. Eng. 87 method for low-end MEMS gyros in attitude and heading
1255–8 reference systems IEEE Trans. Instrum. Meas. 63 2675–80

12

You might also like