Original Paper

Measurement and Control

1–13
Solar vector solving method based on Ó The Author(s) 2019
Article reuse guidelines:
photovoltaic array and its application sagepub.com/journals-permissions
DOI: 10.1177/0020294019887487

in attitude estimation of navigation journals.sagepub.com/home/mac

system

Hongmei Zhang, Huaqing Zhang , Guangyan Xu and Hao Liu

Abstract
Solar-powered unmanned aerial vehicles usually fly at high altitudes, and they are mainly powered by the photocells cov-
ering the body of unmanned aerial vehicles. Considering that the solar vector cannot be affected by the disturbing mag-
netic field and harmful acceleration, a unit solar vector solving method based on photovoltaic array is proposed in this
paper. The photocells with different installation angles are selected to form the photovoltaic array. The solar vector is
solved by the least-squares method on the basis of normalization by using the output currents of the photovoltaic array.
For eliminating the influence of faults and reflected light on the solving of the solar vector, an adaptive least-squares unit
solar vector solving method is proposed. In addition, a solar vector measuring device is designed in order to verify the
effectiveness of the proposed methods. By employing the structural advantages of the device, the current generated by
the reflected light of the sky can be solved according to the currents generated by all photocells of this device. Thus, the
solved current generated by the reflected light of the sky is more accurate. Moreover, strapdown inertial navigation sys-
tem/solar vector/global positioning system integrated navigation Kalman filtering algorithm is proposed, in which the
adaptive least-squares unit solar vector solving method is applied to the measurement update of the filter. The effective-
ness of the methods proposed in this paper is illustrated by some numerical and physical simulations.

Keywords
Photovoltaic array, photocell current, solar vector, attitude estimation, Kalman filtering, solar-powered UAV

Date received: 4 August 2019; accepted: 17 October 2019

Introduction anti-disturbance algorithm for attitude determination

of UAVs.
For most unmanned aerial vehicles (UAVs), improving In the paper by Hemerly et al.,5 an extended Kalman
the accuracy and stability of the navigation system filter for implementing an attitude and heading refer-
using relatively low-cost sensors is one of the crucial ence system (AHRS) with acceleration compensation
issues to be considered. The MARG (magnetic, angular by the global positioning system (GPS) was introduced.
rate, and gravity) sensors are widely used for attitude In the filter, the derivative of the velocity obtained by
determination of UAVs or other vehicles.1–4 An inertial GPS is removed from the acceleration measured by the
measurement unit (IMU) is usually composed of gyro- accelerometer which is also used to sense gravity field.
scope and accelerometer which are used to measure the For exploring anti-disturbance attitude measurement
angular rate and acceleration of a UAV, respectively. methods, an improved antenna array–based communi-
In addition, the accelerometer is also used to sense cation system6 was applied for attitude estimation by
the gravity field simultaneously. The magnetometer is computing the line-of-sight (LOS) path between the
used to measure the geomagnetic field. However, accel-
erometers and magnetometers are sensitive to harmful
acceleration and disturbing magnetic field in the pro- School of Automation, Shenyang Aerospace University, Shenyang, China
cess of measuring the gravity field and geomagnetic
Corresponding author:
field, respectively. Therefore, it is necessary to develop Guangyan Xu, School of Automation, Shenyang Aerospace University,
sensors which are free from harmful acceleration Shenyang 110136, China.
and disturbing magnetic field or corresponding Email: [email protected]

Creative Commons CC BY: This article is distributed under the terms of the Creative Commons Attribution 4.0 License
(http://www.creativecommons.org/licenses/by/4.0/) which permits any use, reproduction and distribution of the work without
further permission provided the original work is attributed as specified on the SAGE and Open Access pages (https://us.sagepub.com/en-us/nam/
open-access-at-sage).
2 Measurement and Control

base station and the UAV. In addition, a complete generated by two rotating photocells. This method
framework for attitude estimation by exploiting three- demands a rotating device of photocells and has no
dimensional (3D) LOS vector obtained from the fault tolerance for photocell failure.
antenna array system was proposed. In order to Considering that the solar vector is free from harm-
improve the capability of resisting harmful acceleration, ful acceleration and disturbing magnetic field, a unit
an attitude estimation method which is robust for solar vector solving method based on photovoltaic
harmful accelerative environment was designed using array is proposed in this paper. The solar vector is
MARG sensors and GPS receiver in the work of No solved by the least-squares method on the basis of nor-
et al.7 The navigation system excludes accelerometer malization using the output currents of the photovoltaic
and estimates pseudo-attitude information from the array, which significantly simplifies the solving process
velocity measured by the GPS receiver. In addition, a of the solar vector. Moreover, an adaptive least-squares
magnetic-pitch angle determination method which unit solar vector solving method is also proposed,
extracts pitch angle and yaw angle information from which greatly improves the fault tolerance of the solar
magnetic vector was proposed in order to complement vector solving. In order to verify the effectiveness of the
the deviation of pseudo-attitude. The paper by Tong proposed methods, a physical device with photovoltaic
et al.8 proposed a novel adaptive extended Kalman fil- array is designed in this paper. The current generated
ter which includes a multiplicative extended Kalman fil- by the reflected light is solved according to the currents
ter (MEKF) and a hidden Markov model (HMM) generated by all photocells of this device. Thus, the
recognizer. Moreover, the HMM recognizer can obtain solved current generated by the reflected light is more
precise attitude information using MARG sensors in reliable. In simulations of the proposed strapdown iner-
the circumstance of 3D motion tracking. tial navigation systems (SINS)/solar vector/GPS inte-
For the problem of attitude estimation of UAV grated navigation algorithm, the effectiveness of the
using Kalman filter, the disturbance of harmful accel- proposed solar vector solving method and its auxiliary
eration on the gravity field measured by accelerometer effect for attitude estimation are verified.
can be reduced by constructing an adaptive observation
noise variance matrix. The adaptive adjustment of the
observation noise variance matrix requires the estima- Mathematical model of solar vector
tion of disturbance. Disturbance-observer-based con-
trol (DOBC) and related methods have been researched The relationship between two coordinate frames which
and applied in various industrial sectors in the last four refer to ‘‘B’’ and ‘‘N’’ is used to represent the attitude of
decades. An observation mechanism is employed to vehicles:
estimate the total disturbance in the DOBC method.
The paper proposed by Chen et al.9 reviews many 1. ‘‘B’’ represents the body-fixed frame which is a
widely used linear and nonlinear disturbance/uncer- sensor frame with each axis pointing along each
tainty estimation techniques and discusses various com- of the IMU axes.
pensation techniques. For enhancing the performance 2. ‘‘N’’ represents the local earth-fixed reference
of disturbance rejection, the modeling error is actively coordinate frame taken from the East-North-Up
estimated by a Kalman filter using an active-model- (ENU) reference system. The XYZ axes are pos-
based control scheme.10 itive to geographic East, North, and Up direc-
Celestial navigation system (CNS) has been widely tions, respectively.
used for attitude and position estimation in the naviga-
tion system of UAVs.11,12 In recent years, research Hence, RBN and RN B are the rotation matrixes con-
works on the solar vector measurement using atmo- structed by the quaternions of the real attitudes of the
spheric polarization modes for attitude determination UAV dynamic model in the numerical simulation.
have achieved fruitful results.13–15 Information on Correspondingly, the rotation matrixes CBN and CN B are
atmospheric polarization modes can also be obtained constructed by estimated quaternions of the filter.
in the field of views obscured by barriers. For this rea-
son, using the atmospheric polarization mode to obtain
the solar vector has become a research direction. As The model of solar vector in the ‘‘N’’ coordinate
shown in Figure 1, solar-powered UAV usually flies
frame
above clouds. With the continuous advancement of
photovoltaic technology, solar-powered UAV can The angle between the line which connects the earth
obtain more energy from the sun for flight. In addition, and the sun and the equatorial plane of the earth will
solar-powered UAVs have broad application prospects. vary when the earth rotates around the sun. The rota-
Hence, it is necessary to develop the techniques of tion of the earth causes the point of direct sunray to
acquiring the solar vector through photocells for atti- move around the surface of the earth. In order to con-
tude determination of UAVs. The paper by Yang struct the model of the solar vector in the navigation
et al.16 proposed a method of solving the angle of the coordinate frame, assume that the sun rotates around
solar ray using the alternating short-circuit currents the earth. According to the results proposed by Chong
Zhang et al. 3

Figure 1. Solar-powered UAV flying at high altitudes.

Figure 3. The projection of the unit solar vector in the ‘‘B’’

coordinate frame.

earth passes perihelion, and Os is the earth revolution

angular velocity.
Assume that the latitude and longitude of the local posi-
tion are la and lo , respectively. Consequently, the model of
unit solar vector in the ‘‘N’’ coordinate frame is
p p
sN = Ry Rx Ry (  la )Rz (lo + Oe t)sI ð3Þ
2 2
Figure 2. The solar vector in geocentric inertial coordinate
frame.
where Oe is the earth rotation angular velocity. Thus,
as long as the current date, time, and the latitude and
and Wong,17 the model of the solar vector in the ‘‘N’’ longitude can be determined, the unit solar vector in
coordinate frame can be derived. the navigation frame can be obtained.
The solar vector in geocentric inertial coordinate
frame O  XI YI ZI is depicted in Figure 2, and the
model of the solar vector in geocentric inertial coordi- The model of the unit solar vector in the ‘‘B’’
nate frame is given by coordinate frame
The azimuth angle and elevation angle of the unit solar
1
sI = Rx (  is )Rz (  vs )so ð1Þ vector sB in the ‘‘B’’ coordinate frame are defined as j
1  e cos E and z. The projection of the unit solar vector in the ‘‘B’’
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
where so = ( cos E  e sin E (1  e2 ) 0 )T , is is the coordinate frame is shown in Figure 3.
obliquity of the ecliptic, vs is the argument of perihe- According to the projection of the unit solar vector
lion of the earth revolution orbit, E is the eccentric shown in Figure 3, the unit solar vector in the ‘‘B’’ coor-
anomaly of the earth revolution orbit, e is the eccentri- dinate frame17 is given by
city of the earth revolution orbit, and R() is the rota- 2 3 2 3
tion matrix of rotating the corresponding angle around sx cos z cos j
the corresponding axis. sB = 4 sy 5 = 4 cos z sin j 5 ð4Þ
According to Kepler’s equation sz sin z

E  e sin E = M = Os (t  ts ) ð2Þ The unit solar vector is usually obtained by solving

the azimuth angle j and elevation angle z of the solar
the real-time solution of the eccentric anomaly E can vector in the ‘‘B’’ coordinate frame.16 This paper pro-
be obtained. In equation (2), M is the mean anomaly of poses a method of directly solving the unit solar vector
the earth revolution orbit, ts is the moment when the by using the output currents of the photovoltaic array.
4 Measurement and Control

The solving progress of the solar vector is significantly The expression of current generated by the sunray18
simplified. is approximately

Ii = I? cos ui ð8Þ
Design of the unit solar vector solving
method where I? is the current generated by the sunray when
the sunray is incident vertically to the photocell of a cer-
Photocell is a special semiconductor diode which can tain area. Its value is mainly related to the light inten-
convert light energy into electrical energy. In order to sity. For each photocell in the photovoltaic array, the
improve the anti-disturbance capability of attitude esti- value of I? is treated as the same.
mation, a unit solar vector solving method is proposed According to equations (6)–(8), the current output
using the structural advantages of photovoltaic array by photocell i can be expressed as
and the principle of energy conversion. The short-
circuit current output by the photovoltaic array is col- ISi = IF + I? hRi ð3Þ, sB i
lected through the corresponding circuit. Usually, ð9Þ
= fi ðIF , I? , sB (1), sB (2), sB (3)Þ
solar-powered UAVs are covered with photocells hav-
ing the same area. Each of these photocells has a fixed where constants IF and I? , and the vector sB are all
installation angle. Thus, there is a fixed and known unknown quantities, and fi () is a function with five vari-
rotation matrix Ri from the coordinate frame of photo- ables. It takes only five photocells in the photovoltaic
cell i to the ‘‘B’’ coordinate frame. In order to reduce array that the variable IF can be solved. Thus, the cur-
the influence of wing vibration on solar vector solving, rent generated by the sunray of photocell i is known,
the photocells with different installation angles at the and the value is
wing root and airframe are selected to form the photo-
voltaic array. In addition, the photocells are labeled as Ii = ISi  IF ð10Þ
1 – N, and the rotation matrix Ri is calculated accord-
ing to the installation angle of photocell i. According to equations (6) and (8), the expression of
current generated by the sunray of photocell i is
Least-squares unit solar vector Ii = I? hRi ð3Þ, sB i
ð11Þ
The rotation matrix Ri from the coordinate frame of = I? ½ Ri ð1, 3Þ Ri ð2, 3Þ Ri ð3, 3Þ sB
photocell i to the ‘‘B’’ coordinate frame is known. Thus,
the normal unit vector of photocell i in the ‘‘B’’ coordi- Assume that the photocells in the photovoltaic array
nate frame can be presented as are all exposed to the sunray and generate currents nor-
mally. Thus, the current vector output by the photovol-
ni = R i z i ð5Þ taic array is
T
where zi = ½ 0 0 1  is the normal unit vector of
photocell i in the coordinate frame. IS = ½ I1 I2    IN  T = I? R S sB
2 3
In the ‘‘B’’ coordinate frame, the angle between the R1 ð1, 3Þ R1 ð2, 3Þ R1 ð3, 3Þ
normal unit vector ni of photocell i and the unit solar 6 R2 ð1, 3Þ R2 ð2, 3Þ R2 ð3, 3Þ 7
6 7 ð12Þ
vector sB is defined as ui . Thus, the direction cosine of = I? 6
6 .. .. ..
7 sB
7
the vector ni and sB can be given by 4 . . . 5

cos ui = hni , sB i = hRi ð3Þ, sB i ð6Þ RN ð1, 3Þ RN ð2, 3Þ RN ð3, 3Þ

where the symbol hi represents the inner product of where RS is the rotation matrix of the photovoltaic
two vectors, and Ri (3) is the third column of rotation array.
matrix Ri . There is a linear relationship between the current
The current output by photocell i contains the cur- vector IS and the unit solar vector sB . However, matrix
rents generated by the reflected light and the sunray, RS is irreversible. The unit solar vector sB cannot be
respectively. Assume that all photocells of the photo- directly solved. The installation angles of photocells in
voltaic array are exposed to the sky. Thus, define IFi the photovoltaic array are different from each other. It
and Ii as the currents generated by the reflected light of can be inferred that RS must be a matrix with full col-
the sky and the sunray, respectively. In addition, the cur- umn rank. In addition, I? cannot be directly measured.
rents generated by the reflected light of the sky can be Hence, the least-squares unit solar vector can be solved
treated as the same for different photocells. Thus, the according to RS and IS using the normalized idea. The
current18 output by photocell i can be represented as least-squares unit solar vector is
 1  T 1 T
ISi = IF + Ii ð7Þ RTS RS RTS IS RS RS RS IS
sB = =
 1 
 ð13Þ
I?  RT RS
where IF is the current generated by the reflected light
 S RS IS 
T

of the sky.
Zhang et al. 5

According to equation (13), the relation of I? and where

jj(RTS RS )1 RTS IS jj satisfies 2 3
  Rm1 ð1, 3Þ Rm1 ð2, 3Þ Rm1 ð3, 3Þ
 1  6 Rm2 ð1, 3Þ Rm2 ð2, 3Þ Rm2 ð3, 3Þ 7
I? =  RTS RS RTS IS  ð14Þ 6 7
Rseq, 5 = 6
6 Rm3 ð1, 3Þ Rm3 ð2, 3Þ Rm3 ð3, 3Þ 7
7
4 Rm4 ð1, 3Þ Rm4 ð2, 3Þ Rm4 ð3, 3Þ 5
Rm5 ð1, 3Þ Rm5 ð2, 3Þ Rm5 ð3, 3Þ
Adaptive least-squares unit solar vector
s5B is the least-squares unit solar vector solved by
During the flight of UAV, the angles between the nor-
Iseq, 5 and Rseq, 5 .
mal vectors of some photocells and the solar vector
According to equations (5) and (18), the angle umi
may be greater than 90°, which prevents the sunray
(i . 5) between the normal unit vector nmi of photocell
from reaching the photocells. Moreover, each photocell
mi and the solved unit solar vector s5B can be deter-
still outputs a certain amount of current due to the irra-
mined. Thus, according to equations (11) and (14), the
diation of the reflected light. Thus, it is necessary to
theoretical value of the current generated by photocell
determine whether the photocell is exposed to the sun-
mi with umi \ 908 is given as
ray and adaptively construct the current vector IS and
the rotation matrix RS of the photovoltaic array. In 1  T
Ipmi = kððRseq, 5 ÞT Rseq, 5 Þ Rseq, 5 Iseq, 5 k
addition, the current IF generated by the reflected light ð19Þ
of the sky can be obtained only under the condition 3½ Rmi ð1, 3Þ Rmi ð2, 3Þ Rmi ð3, 3Þ s5B
that the photocells are exposed to sunray.
In addition, the criterion of determining whether the
Suppose that the number N of photocells constitut-
photocell labeled mi (i . 5) is available is
ing the photovoltaic array is sufficiently large. The cur-

p

rent ISi (i = 1, 2, . . . , N) output by the photocell array

I  Im
4t
mi i
ð20Þ
ranges from large to small using a sorting algorithm.
Thus, the ordered current vector of the photovoltaic where t is an empirically given threshold. When the cur-
array is rent generated by photocell mi does not satisfy equation
 T (20), the serial number of photocell mi1 is saved as mh .
ISseq = ISm1 ISm2    ISmN ð15Þ Therefore, an adaptive photovoltaic array can
be established according to the photocells labeled
where the serial numbers m1 , m2 , . . . , mN correspond to m1 , . . . , mh . In addition, the current vector generated
the labels of the photocells. Similarly, the ordered rota- by the photocells is defined as Iseq, h . According to equa-
tion matrix of the photovoltaic array is given as tion (13), the adaptive least-squares unit solar vector is
2 3  1 
Rm1 ð1, 3Þ Rm1 ð2, 3Þ Rm1 ð3, 3Þ T T
6 Rm2 ð1, 3Þ Rm2 ð2, 3Þ Rm2 ð3, 3Þ 7 Rseq, h Rseq, h Rseq, h Iseq, h
6 7 a
sB =   ð21Þ
Rseq, S = 6
4
.. .. .. 7 ð16Þ
5   1  T 
. . .  Rseq, h T Rseq, h 
Rseq, h Iseq, h 

RmN ð1, 3Þ RmN ð2, 3Þ RmN ð3, 3Þ
Take the first five large currents from current vector where the rotation matrix Rseq, h of the photovoltaic
ISseqto form the current vector ISseq, 5 as follows array is given as
2 3
 T Rm1 ð1, 3Þ Rm1 ð2, 3Þ Rm1 ð3, 3Þ
ISseq, 5 = ISm1 ISm2    ISm5 ð17Þ 6 Rm2 ð1, 3Þ Rm2 ð2, 3Þ Rm2 ð3, 3Þ 7
6 7
Rseq, h = 6 .. .. .. 7 ð22Þ
Under the condition that the photocells work nor- 4 . . . 5
mally and the number of photocells N is sufficiently Rmh ð1, 3Þ Rmh ð2, 3Þ Rmh ð3, 3Þ
large, the photocells outputting the current vector ISseq, 5
must be exposed to the sunray. Thus, the current IF
generated by the reflected light of the sky can be Solar vector measuring device and system
solved according to equations (9) and (17). Moreover, model
the current vectors Iseq = ½ Im1 Im2    ImN T and
Iseq, 5 = ½ Im1 Im2    Im5 T are obtained after the The solar vector and the gravity field are used to assist
current IF is subtracted from the components of current SINS for attitude estimation of UAV. GPS is used to
vectors ISseq and ISseq, 5 . assist SINS for position and velocity estimation of
According to equation (13), the least-squares unit UAV.
solar vector satisfies
 T 1  T Solar vector measuring device and its model
Rseq, 5 Rseq, 5 Rseq, 5 Iseq, 5
5
sB =   1  
 ð18Þ The temperature drift of each photocell and the differ-
 T
 Rseq, 5 T Rseq, 5 
Rseq, 5 Iseq, 5  ence between each photocell are ignored. In order to

verify the validity of the designed unit solar vector
6 Measurement and Control

solving method, a solar vector measuring device shown

in Figure 4 is designed in this paper. In addition, the
device is placed on the roof, and all photocells are
exposed to the clear sky. Seven photocells with the same
area are uniformly laid on seven surfaces of the six frus-
tum planes and are labeled as 1–7, respectively. The
angle between each side face and the bottom surface is
a = 408. The angle between each side of the bottom sur-
face is b = 608. Moreover, the acquisition circuit of
short-circuit current is designed for the photocells.
Since the light is too strong during the test, the short-
circuit currents are somewhat distorted. Thus, half of
each photocell is blocked with white paper.
Note that cos b = cos 2b and sin b = sin 2b.
Under the condition that the sunray can reach all the
Figure 4. Solar vector measuring device placed on the roof.
photocells of the solar vector measuring device, the
rotation matrix of the photovoltaic array is constructed
as
According to equations (7), (8), and (25), the follow-
2 3 ing equation can be obtained
R1 ð1, 3Þ R1 ð2, 3Þ R1 ð3, 3Þ
6 R2 ð1, 3Þ R2 ð2, 3Þ R2 ð3, 3Þ 7 8 S S S
6 7
RS, 7 = 6 7 < I1 + I4  2IF = 2(I7  IF ) cos a
>
6 .. .. .. 7
4 . . . 5 IS2 + IS5  2IF = 2(IS7  IF ) cos a ð26Þ
>
: S S S
R7 ð1, 3Þ R7 ð2, 3Þ R7 ð3, 3Þ I3 + I6  2IF = 2(I7  IF ) cos a

2 3 The average value of IF obtained from the three

0  sin a cos a ð23Þ expressions in equation (26) is
6 7
6 sin a sin b  sin a cos b cos a 7
6 7 IS1 + IS2 + IS3 + IS4 + IS5 + IS6  6IS7 cos a
6 sin a sin b sin a cos b cos a 7 IF = ð27Þ
6 7 6(1  cos a)
6 7
=6 0 sin a cos a 7
6 7 In addition, the model of the solar vector measuring
6  sin a sin b sin a cos b cos a 7
6 7 device can be simply constructed as
6 7
4  sin a sin b  sin a cos b cos a 5
0 0 1 IS, 7 = ½ I1 I2    I7 T = I? RS, 7 sB, N + nI ð28Þ

According to equation (6), the direction cosine of the where sB, N = RBN sN ; nI represents the random white
normal unit vector ni (i = 1, 2, . . . , 7) and sB is given as noise drift vector with variance s2I .
8
> cos u1 = sB (2) sin a + sB (3) cos a
>
> The model of IMU sensors and GPS
>
> cos u2 = sB (1) sin a sin b  sB (2) sin a cos b
>
>
>
> It is assumed that the IMU sensors have been cali-
>
> + sB (3) cos a
>
> brated. Thus, the models of gyroscope and acceler-
>
> cos u3 = sB (1) sin a sin b + sB (2) sin a cos b
>
> ometer can be established as follows
>
>
>
< + sB (3) cos a
> 
vB = vt + evr + nv
cos u4 = sB (2) sin a + sB (3) cos a ð29Þ
>
> aB = RBN gN + av + ba + ear + na
>
> cos u5 = sB (1) sin a sin b + sB (2) sin a cos b
>
>
>
> where vB and aB are measuring outputs of gyroscope
> + sB (3) cos a
>
>
> and accelerometer in the ‘‘B’’ coordinate frame, respec-
>
> cos u6 = sB (1) sin a sin b  sB (2) sin a cos b
>
> tively; vt is the actual angular velocity of the UAV
>
>
>
> + sB (3) cos a dynamic model; gN is the gravity field in the ‘‘N’’ coor-
>
:
cos u7 = sB (3) dinate frame; av and ba are vehicle linear acceleration
ð24Þ and bias of accelerometer obtained by the acceler-
ometer calibration, respectively; nv and na are the ran-
According to equation (24), the following equation dom white noise drift vectors of gyroscope and
can be obtained accelerometer, respectively; and evr and ear are first-
8 order Markov process drifts of the gyroscope and
< cos u1 + cos u4 = 2 cos a cos u7 accelerometer, respectively. The dynamic models of evr
cos u2 + cos u5 = 2 cos a cos u7 ð25Þ and ear are as follows
:
cos u3 + cos u6 = 2 cos a cos u7
Zhang et al. 7
8
> 1 v According to equations (33), (35), and (36), the state
> v
< e_r = e + vr
Tv r equation of UAV based on sensor models can be
ð30Þ
>
> 1 obtained as
: e_ar = ear + va
Ta
x_ = Ax + BaN ð37Þ
where Tv and Ta are the correlation time, and vr and
where
va are the random white noise drifts of the first-order
8 2 3
Markov process. > 1=2O 0433 0433
>
>
The position and velocity model in the ‘‘N’’ coordi- >
< A = 4 0334 0333 I333 5
nate frame obtained by GPS is established as follows
>  0334 0333 0333 10310 ð38Þ
( >
> 0
>
: B = 733
pG = pG
t + np I333 1033
ð31Þ
vG = vG
t + nv

where pG G
t and vt are the actual position and velocity of Observation matrix of the state
the UAV dynamic model in the ‘‘N’’ coordinate frame,
respectively, and np and nv are the random white noise The observed unit of the solar vector and gravity field is
drift vectors with variances of s2p and s2v , respectively.  B 
CN sN
yk = = h(q) ð39Þ
CBN gN
State equation of UAV dynamic model Expand h(q) by Taylor at the state of quaternions.
The motion states of UAV usually include attitude qua- The observation matrix of the solar vector and gravity
ternions, positions, and velocities in 3D space. Thus, field is
the solar vector of the UAV dynamic model is con-

∂h(q)

structed as Hsg, k = ð40Þ

∂q
q = q^kjk1
 T
x = qT pT vT 1031
ð32Þ where q^kjk1 is the estimated quaternion after time
update.
where q, p, and v are attitude quaternions, positions, Since the filter also observes the position and velo-
and velocities of the UAV, respectively. city information obtained by GPS, the observation
In order to make the numerical simulation more matrix of the filter is constructed as
close to the actual situation of hardware, the dynamic

model of UAV is established according to the measure- Hsg, k 0636
ment data of sensor models which have been estab- Hk = ð41Þ
0634 I636 12310
lished in equation (29). Thus, the dynamic model of
attitude of UAV is given by

1
SINS/solar vector/GPS integrated
q_ = Oq ð33Þ navigation Kalman filtering algorithm
2
Before the measurement update of the Kalman filtering
where
algorithm, the unit solar vector needs to be solved in
2 3 real time through the output current vector of the
0 vxB vyB vzB
6 vx 0 vzB y 7
vB 7 photovoltaic array. The detailed steps of the algorithm
6
O = 4 By ð34Þ are given as follows:
vB vzB 0 vxB 5
vzB vyB vxB 0
Time update
The linear acceleration of UAV measured by the
accelerometer model in the ‘‘N’’ coordinate frame is Since the state equation (37) is a continuous time sys-
given as tem, equation (37) is discretized by fourth-order
Runge–Kutta method in order to update the state in
aN = CN
B (aB  ba )  gN ð35Þ discrete time. Then the time update of the state can be
obtained
which is taken as the input of the second-order integral 8
dynamic model of UAV. The second-order integral > S1 = A^xk1 + BaN
>
>
dynamic model is given as < S = A(^ xk1 + S1 Th =2) + BaN
2
ð42Þ
 >
> S3 = A(^xk1 + S2 Th =2) + BaN
p_ = v >
:
ð36Þ S4 = A(^xk1 + S3 Th ) + BaN
v_ = aN
8 Measurement and Control

dx = (S1 + 2S2 + 2S3 + S4 )Th =6
ð43Þ
x^k=k1 = x^k1 + dx

where Th is the sampling period, x^k1 is the state esti-

mation at the moment k  1, and x^k=k1 is the state esti-
mation after the time update of the filter.
The time update of state estimation covariance
matrix is given as

dP = (APk1 + Pk1 AT + BRu BT + Q)Th
ð44Þ
Pk=k1 = Pk1 + dP

where Ru is the input noise variance matrix, Q is pro-

cess noise variance matrix of state equation (37), Pk1
is the state estimation covariance matrix at the moment
k  1, and Pk=k1 is the covariance matrix after the time
update of the filter.
Figure 5. The simulation curves of pitch and roll angles of UAV.

Solving of the unit solar vector

According to equations (15)–(20), the ordered current
vector Iseq, h and the ordered rotation matrix Rseq, h of
the photovoltaic array can be obtained. Thus, accord-
ing to equation (21), the adaptive unit solar vector is
given as
 T 1  T
Rseq, h Rseq, h Rseq, h Iseq, h
a
sB =   1  
 ð45Þ
 T
 Rseq, h T Rseq, h 
Rseq, h Iseq, h 


Measurement update
The measurement update is given as
8  1
< K = Pk=k1 HTk Hk Pk=k1 HTk + R
x^ = x^k=k1 + K Zk  Hk x^k=k1 ð46Þ
: k
Pk = ðI  KHk ÞPk=k1 Figure 6. Position estimation errors of the filter in numerical
simulation.
where K is the gain matrix of the Kalman filtering algo-
rithm and R is the measurement noise variance matrix. pffiffiffi T
In addition, the measurement information Zk is sN = ½ 1=2 0 3=2  and gN = ½ 0 0 1 T ,
   T  G T T respectively. The value of I? is given as 10 mA. The
Zk =½ saB T ðaB  ba  av ÞT pG v  ð47Þ standard deviation of the random white noise of the
solar vector measuring device is given as sI = 0:03 mA.
where the motion acceleration av is solved by the mea-
The random white noise and first-order Markov pro-
sured velocity differentiation of UAV according to the
cess drift of the gyroscope are given as 0.002 and
method proposed by Hemerly et al.,5 and saB is the
0.001 rad/s, respectively, and the correlation time of the
adaptive least-squares unit solar vector in the ‘‘B’’ coor-
Markov process drift is given as Tw = 7200=s. The ran-
dinate frame.
dom white noise and first-order Markov process drift
of the accelerometer are given as 0.8mg and 2mg,
Numerical simulation and physical respectively, and the correlation time of the Markov
simulation process drift is given as Ta = 3600 s. The random white
noise drifts of position and velocity obtained by GPS
Numerical simulation are given as 2 m and 0.3 m/s. Assume that the UAV is a
The algorithms designed in this paper are simulated by multirotor, and the multirotor flies at a fixed altitude.
MATLAB. The simulation step size is given as For the convenience of simulation, suppose that the
Dt = 102 s. The simulation time is given as t = 200 s. motion accelerations in the x-direction and the y-direc-
The unit solar vector and the unit gravity field vector tion are proportional to the pitch and roll angles of
in the ‘‘N’’ coordinate frame are given as UAV, respectively. Assume that the yaw angle of UAV
Zhang et al. 9

Figure 9. Attitude estimation errors assisted only by the

gravity field in numerical simulation.
Figure 7. Velocity estimation errors of the filter in numerical
simulation.

Figure 8. Attitude estimation errors of the filter assisted by Figure 10. The collected currents of photocells through the
the gravity field and the solar vector in numerical simulation. solar vector measuring device.

is fixed; the simulation curves of pitch and roll angles Physical simulation
of UVA are given in Figure 5.
The simulation step size and simulation time are also
The position and velocity estimation errors of the fil-
given as Dt = 102 s and t = 200 s, respectively. On a
ter are shown in Figures 6 and 7; the attitude estimation
bright afternoon, the solar vector measuring device
errors of the filter assisted by the gravity field and the
shown in Figure 4 is placed horizontally on the roof.
solar vector are shown in Figure 8; and the attitude esti-
Moreover, the photocells of the device are all exposed
mation errors of the filter assisted only by the gravity
to the sky and the sunray. The collected currents of
field are shown in Figure 9. According to the results
photocells are also filtered by a digital low-pass filter,
shown in Figures 6 and7, the estimation errors of posi-
and then the currents are sent to MATLAB. The col-
tion and velocity converge to a relatively small range.
lected currents of photocells through the solar vector
As shown in Figure 8, the attitude estimation errors of
measuring device are shown in Figure 10. The current
the filter are within 0.5°. As shown in Figure 9, the atti-
generated by the reflected light of the sky is solved
tude estimation errors of the pitch and roll angles are
according to equation (27). The currents of photocells
all within 0.5°, but the estimation errors of the yaw
generated by the sunray are shown in Figure 11.
angle are divergent. Thus, according to the simulation
Figures 10 and 11 show that the seven photocells of the
results of Figures 8 and 9, we can see the assistant effect
solar vector measuring device have the same
of the solar vector on attitude estimation.
10 Measurement and Control

Figure 11. The currents of photocells which have removed the

current IF . Figure 13. Attitude estimation errors obtained by gyroscope
integration with IMU data coming from test 1.

Figure 12. The UAV with Pixhawk flight controller.

characteristics according to the variation trend of the

current curves. Therefore, it can be inferred that the Figure 14. Attitude estimation errors obtained by gyroscope
design of the solar vector measuring device is reason- integration with IMU data coming from test 2.
able to some extent.
Since the IMU sensors and magnetometer in Since various errors exist in the solar vector measur-
Pixhawk flight controller of an UAV shown in Figure ing device, the projection in the ‘‘N’’ coordinate system
12 have been corrected, export data of these sensors for of the solved unit solar vector by the solar vector mea-
physical simulation. The UAV with Pixhawk flight suring device at the initial moment is used as the unit
controller is placed on the ground, and it is assumed solar vector in the ‘‘N’’ coordinate frame. In addition,
that the UAV model in numerical simulation is also in the Euler angles of UAV can be solved according to
a static state. The transmission speed of IMU and mag- the collected data of accelerometer and magnetometer
netometer data in the MAVLink module of the PX4 when the UAV is still. Hence, the data of IMU and
source code of the Pixhawk flight controller is changed magnetometer are collected once at test 1 when the
to 100 Hz. The data of IMU and magnetometer are col- yaw, pitch, and roll angles of UAV are c1 = 0:5078,
lected to log file through ground control station for u1 = 0:7188, and g 1 = 1:4018, respectively, and the
more than 200 s, and then the log file is read through data of IMU and magnetometer are collected
MATLAB for 200 s. The collected IMU data are simu- once again at test 2 when the yaw, pitch, and roll angles
lated together with the current vector output by the of UAV are c2 = 17:3218, u2 =  22:0238, and
solar vector measuring device. g 2 = 14:2488, respectively.
Zhang et al. 11

Figure 15. Attitude estimation errors assisted by the gravity

field and the solar vector with IMU data coming from test 1.
Figure 17. Attitude estimation errors assisted only by the
solar vector with IMU data coming from test 1.

Figure 16. Attitude estimation errors assisted by the gravity

field and the solar vector with IMU data coming from test 2.

Figure 18. Attitude estimation errors assisted only by the

For determining the assistant effect of solar vector solar vector with IMU data coming from test 2.
on attitude estimation, the following experiments are
carried out. The attitude estimation errors obtained by
gyroscope integration are shown in Figures 13 and 14, It can be seen from Figures 17 and 18 that the cumu-
when the collected IMU data come from tests 1 and 2, lative errors of attitude estimation assisted by the solar
respectively. The attitude estimation errors of the filter vector are smaller than those of gyroscope integration
assisted by the gravity field and the solar vector are given in Figures 13 and 14. As presented in Figures 15
shown in Figures 15 and 16, when the collected IMU and 16, the cumulative errors of yaw angles are smaller
data come from tests 1 and 2, respectively. The attitude than those shown in Figures 19 and 20. Therefore,
estimation errors of the filter assisted only by the solar according to the comparison of attitude estimation
vector are given in Figures 17 and 18, when the col- errors illustrated in Figures 13– and 20, we can see that
lected IMU data come from tests 1 and 2, respectively. the solar vector has an assistant effect on attitude esti-
The attitude estimation errors of the filter assisted only mation. Consequently, the numerical and physical
by the gravity field are presented in Figures 19 and 20, simulations have verified that the solar vector solving
when the collected IMU data come from tests 1 and 2, method is effective and has an assistant effect on atti-
respectively. tude estimation.
12 Measurement and Control

to the UAV. Moreover, the solar vector cannot be

affected by the disturbing magnetic field and harmful
acceleration. Thus, it can be concluded that the unit
solar vector solving methods proposed in this paper
have certain value in engineering application.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with
respect to the research, authorship, and/or publication of this
article.

Funding
The author(s) disclosed receipt of the following financial
support for the research, authorship, and/or publication of this
article: This work is partially supported by the National Natural
Science Foundation of China (61703286) and the series project
Figure 19. Attitude estimation errors assisted only by the of Liaoning Province Education Department (L201711).
gravity field with IMU data coming from test 1.

ORCID iD
Huaqing Zhang https://orcid.org/0000-0002-4218-7542

References
1. Li W and Wang J. Effective adaptive Kalman filter for
MEMS-IMU/magnetometers integrated attitude and
heading reference systems. J Navig 2013; 66(1): 99–113.
2. Zhang RH, Jia HG, Chen T, et al. Attitude solution for
strapdown inertial navigation system based on quater-
nion algorithm. Optic Precis Eng 2008; 16(10): 1963–
1970 (in Chinese).
3. Grivon D, Vezzetti E and Violante MG. Development of
an innovative low-cost MARG sensors alignment and
distortion compensation methodology for 3D scanning
applications. Robot Auton Syst 2013; 61(12): 1710–1716.
4. Su Y, Liu Y, Li XK, et al. Attitude measurement and
optimization for horizontal directional drilling based on
MARG sensor. Adv Mat Res 2013; 616–618: 455–460.
5. Hemerly EM, Maciel BCO, Milhan AP, et al. Attitude
Figure 20. Attitude estimation errors assisted only by the and heading reference system with acceleration compen-
gravity field with IMU data coming from test 2. sation. Aircr Eng Aerosp Tec 2012; 84(2): 87–93.
6. Cordeiro TFK, Da Costa JPCL, de Sousa RT, et al.
Improved Kalman-based attitude estimation framework
Conclusion for UAVs via an antenna array. Digit Signal Process
2016; 59: 49–65.
MARG sensors in navigation system of UAV are vul- 7. No H, Cho A and Kee C. Attitude estimation method
nerable to disturbing magnetic field and harmful accel- for small UAV under accelerative environment. GPS
eration. Considering that the solar-powered UAVs Solut 2015; 19(3): 343–355.
usually fly at high altitudes, the unit solar vector and 8. Tong X, Li Z, Han G, et al. Adaptive EKF based on
adaptive unit solar vector solving methods are pro- HMM recognizer for attitude estimation using MEMS
posed in this paper. The unit solar vector is solved by MARG sensors. IEEE Sens J 2018; 18(8): 3299–3310.
9. Chen WH, Yang J, Guo L, et al. Disturbance—obser-
the least-squares method. Thus, the obtained unit solar
ver—based control and related methods—an overview.
vector has higher precision and fault tolerance when
IEEE Trans.Ind Electron 2016; 63(2): 1083–1095.
there exist errors in photovoltaic array and acquisition 10. Yi K, Liang X, He Y, et al. Active-model-based control
circuit of short-circuit currents. For solar-powered for the quadrotor carrying a changed slung load. Electro-
UAVs flying at high altitudes, photocells selected from nics 2019; 8(4): 461.
the surface of the body can be used to assist attitude 11. Ushaq M, Cheng FJ and Jamshaid A. A fault tolerant
estimation of the UAV. This method has an assistant SINS/GPS/CNS integrated navigation scheme realized
effect on attitude estimation without adding too much through federated Kalman filter. Appl Mech Mater 2013;
extra weight while the photocells are providing energy 332: 104–110.
Zhang et al. 13

12. Min YL, Xiong Z, Xing L, et al. An improved SINS/ 16. Yang P, Shi W, Liu S, et al. Research on measurement
GNSS/CNS federal filter based on dual quaternions. Acta method of the large range sunray incident angle. Acta
Armament 2018; 39(2): 315–324 (in Chinese). Energ Solar Sin 2011; 32(2): 216–219 (in Chinese).
13. Wang Y, Chu J, Zhang R, et al. A novel autonomous 17. Chong KK and Wong CW. General formula for on-axis
real-time position method based on polarized light and sun-tracking system and its application in improving
geomagnetic field. Sci Rep 2015; 5: 9725. tracking accuracy of solar collector. Sol Energy 2009;
14. Ren J, Jun L, Tang J, et al. Polarized skylight pattern- 83(3): 298–305.
based approach to attitude determination. IEEE Sensors 18. Santoni F and Bolotti F. Attitude determination of
J 2015; 15(9): 4917–4927. small spinning spacecraft using three axis magnetometer
15. Fan ZG, Chen ML, Wang B, et al. Three-dimensional and solar panels data. In: Proceedings of the 2000 IEEE
attitude information obtained by the skylight polariza- aerospace conference, Big Sky, MT, 25 March 2000,
tion pattern. Optic Precis Eng 2016; 24(6): 1248–1256 (in pp. 127–133. Piscataway, NJ: IEEE.
Chinese).