GPS Solutions (2019) 23:88

https://doi.org/10.1007/s10291-019-0877-4

ORIGINAL ARTICLE

INS/magnetometer integrated positioning based on neural network

for bridging long‑time GPS outages
Zongkai Wu1 · Wei Wang1

Received: 11 October 2018 / Accepted: 13 June 2019 / Published online: 19 June 2019
© Springer-Verlag GmbH Germany, part of Springer Nature 2019

Abstract
In global position system (GPS) and inertial navigation system (INS) integrated navigation systems, the positioning method
based on artificial intelligence (AI) learning algorithms has the disadvantage of position diverging as the GPS outages time
continues. To solve this problem, INS/magnetometer integrated positioning based on neural network is proposed for bridg-
ing GPS outages over a long time. First, the possibility of using magnetometer for positioning is verified by analyzing the
international geomagnetic reference field model and the magnetometer measurement model. Then, a magnetometer-assisted
positioning solution is proposed. This solution includes four parts: the magnetic fields update module with adaptive extended
Kalman filter (AEKF), predictor by AI, INS/GPS integration with KF, and INS/position integration with AEKF. The simula-
tion and driving test results show that the proposed method can keep most of the position errors within a certain range and
there is no tendency of divergence at all.

Keywords  Inertial navigation system (INS) · Global position system (GPS) · Magnetometer · Positioning · Artificial
intelligence (AI) · Adaptive extended Kalman filter (AEKF)

Introduction Many existing methods focus on training the relation

between the INS measurements and position errors using
The inertial navigation system (INS) and global position artificial intelligence (AI) learning algorithms when GPS
system (GPS) integrated navigation system is widely used is available. Then, the position errors can be predicted
in many applications, such as airborne navigation system by trained AI for bridging GPS outages. The radial basis
(Fang and Yang 2011; Abdolkarimi et al. 2018) and vehicle function neural networks (RBFNN) can be used to predict
navigation system (Wang et al. 2018). However, the GPS INS position and velocity errors for correcting position
signal is easy to be interrupted and cannot be used indoor (Semeniuk and Noureldin 2006). This technique is called
environments in the city, such as tunnel, underground park- the AI-based segmented forward predictor (ASFP). It trig-
ing, and so on. In the past, the high-precision INS provided gers the research about AI learning algorithms in the inte-
an accurate position to bridge GPS outages. The accuracy of grated navigation system, especially during GPS outages,
the position depends on the precision of INS and the length but it ignores the relation between the position errors and
of GPS outages. With the development of micro-electro the past INS outputs. Considering this relation, INS/GPS
mechanical system (MEMS) technology, the volume and integration using dynamic neural networks is provided
cost of the INS system reduce effectively, but accuracy also (Noureldin et al. 2011). This method uses the current and
is reduced. Therefore, with MEMS-INS, it is hard to provide past INS outputs, which lead to a more reliable position
an accurate position in GPS outages. result. For more accurate positioning results during GPS
outages, random forest regression (RFR) is first used to
* Wei Wang predict the position errors (Adusumilli et al. 2013). RFR
wangwei407@hrbeu.edu.cn can provide a more accurate model for high nonlinear INS
Zongkai Wu errors. Compared with the neural networks, RFR reduces
wuzongkai407@hrbeu.edu.cn the position errors by 24–56%. However, RFR has a dis-
advantage that it cannot handle both linear and nonlin-
1
College of Automation, Harbin Engineering University, No. ear input–output relationships effectively. Thus, a hybrid
145 Nantong Street, Harbin 150001, China

13
Vol.:(0123456789)
88 
Page 2 of 11 GPS Solutions (2019) 23:88

method of RFR and principal component regression (PCR) The next section gives an overview of the proposed method.
is proposed for fusing INS and GPS to bridge the period Then, every part of the proposed method is introduced clearly,
of GPS outages (Adusumilli et al. 2015). The multiple- and the results of the simulation and the actual driving test are
decrease factor cubature Kalman filter (MDF-CKF) is pro- presented. Finally, conclusions are discussed.
posed to assist RFR for improving the robustness during
GPS outages (Zhang et al. 2018). But, RFR may have a
long training time. Extreme learning machine (ELM) is Overview of the proposed method
designed to predict the position errors as it just needs a
shorter training time (Zheng et al. 2016; Xu et al. 2018). For the accurate position during GPS outages, we propose a
From the above methods, we can conclude that most of the positioning scheme based on INS/magnetometer integration.
positioning methods using AI learning algorithms concen- However, the premise of this scheme is that position and mag-
trate on the relation between INS measurements and INS netometer must have a relation. The analysis process of the
position errors. But due to the characteristic of INS, it is relation between magnetometer and position will be introduced
inevitable that these methods will still lead to position as follows.
error accumulation and even divergence. First, there are two important frames that need to be defined.
With the development of MEMS technology, the low- The navigation frame (n-frame) is located on the vehicle. It
cost MEMS-INS module or chip usually includes a mag- points to east, north, and upward. The body frame (b-frame) is
netometer. The existing studies about magnetometer mainly fixed on the vehicle. Its x-axis and y-axis point to the right and
focus on calibration and correcting heading angle, but they forward of the vehicle, and z-axis follows the right-hand rule.
ignore the relation between magnetometer and position. Based on the 12th generation of international geomagnetic
Hence, we propose an INS/magnetometer integrated posi- reference field (IGRF) (Thébault et al. 2010), the internal geo-
tioning method for bridging the long-time GPS outages. magnetic field B(r,
⃗ 𝜃, 𝜙, t) and its annual rate of change have
This method mainly includes four modules: the magnetic been set. On the surface of the earth, the geomagnetic fields B ⃗
fields update module with adaptive extended Kalman filter are defined to satisfy B = −∇V  , in which the spherical polar
⃗
(AEKF), predictor by AI, INS/GPS integration, and INS/ coordinates V are defined as follows.
position integration. The INS/magnetometer integration can N n ( )
a n+1
provide the stable geomagnetic fields by adaptive extended
∑ ∑
V(r, 𝜃, 𝜙, t) = a
r
Kalman filter (AEKF), and then, the relation between geo- n=1 m=0
magnetic fields and position from INS/GPS integration can × gm + hm m
[ ]
n (t) cos (m𝜙) n (t) sin (m𝜙)Pn (cos 𝜃)
be built based on predictor by AI when GPS is available. (1)
During GPS outages, the position can be predicted by the where r denotes the radial distance from the center of the
trained AI, but this position may be discrete and have a large earth, a = 63, 712.2  km is the mean reference spherical
noise. Hence, the predicted position and INS are integrated radius of the geomagnetic earth. 𝜃 and 𝜙 are latitude and lon-
by AEKF for more continuous and accurate position result. gitude, respectively. The function Pm n (cos 𝜃) is the Schmidt
The main contributions can be summarized as follows. quasi-normalized associated Legendre functions with degree
n and order m. gm n and hn are the Gauss coefficients with time
m

(1) Different from the traditional method based on the t  . The maximum spherical harmonic degree of the expan-
relation between the INS measurements and position sion N is related to time, and its value is generally taken as
errors, we build the relation between the geomag- 13 in recent years.
netic fields and the absolute position. This scheme can From (1), we can see that the geomagnetic fields only relate
ensure that the position errors will not accumulate or with the position information and time, that is, to say that the
diverge regardless of duration of GPS outage if suf- position can be obtained if the geomagnetic fields and time
ficient datasets for training are ensured. The datasets are known. The geomagnetic fields can be calculated by the
should collect geomagnetic fields observations and magnetometer measurement model. The magnetometer meas-
positions as much as possible, and the data should be urement model is shown as follows (Wu and Pei 2017):
uniformly distributed over the target area. If possible,
the resolution of data should be smaller than 1 m. 𝐁bk = 𝐂sc 𝐂no (𝐂si 𝐂bnk 𝐘nk + 𝐛bi ) + 𝐛b + 𝛆
k
(2)
(2) We employ the cascading framework of AEKF-AI-
AEKF. This framework can provide a stable input into where 𝐁bk is the output from magnetometer at step k in the
b-frame. 𝐂nkk is the rotation matrix from n-frame to b-frame
b
the predictor by AI under training mode. In predicting
mode, the position predicted by AI is integrated with at step k. 𝐛bi and 𝐛b are hard-iron effect and bias vectors. 𝛆
INS by AEKF for the stable, continuous, and accurate is the measurement noise, and it follows the Gaussian white
position. noise. 𝐂si , 𝐂sc , and 𝐂no represent soft-iron interference

13
GPS Solutions (2019) 23:88 Page 3 of 11  88

GPS
INS/GPS integra on During GPS outages, the attitudes calculated by INS and
with KF
magnetic fields in b-frame are integrated to get the geomag-
Velocity
A tude update netic fields in n-frame. The geomagnetic fields can be put into
Posi on
A tude
the predictor by AI to predict the position. But, this position
Posi on update
Posi on may be discrete. Hence, the predicted position and INS are
MEMS-INS Integra on
algorithm with
fused for more stable and continuous position information.
Accelerometer KF

Magne c fields
Gyroscope update with AEKF Predictor by AI Method formulation
Magne c fields Training
Calibrated
Magnetometer
update by Gyro model Based on Fig.  1, the main parts of the proposed method
Integra on
algorithm with
include: the magnetic fields update module with AEKF, pre-
AEKF dictor by AI, INS/GPS integration with KF, and INS/position
integration with AEKF. The details are described in sequence.
Fig. 1  Flow diagram of the proposed method in training mode
Magnetic fields update module with AEKF

INS/posi on Because magnetic fields from magnetometer have errors

MEMS-INS integra on with AEKF
Posi on including noises and disturbances, even if it has been cali-
Acclerometer
A tude
update
A tude brated, the magnetometer is integrated with INS for the stable
update Integra on Posi on
Gyroscope
Predictor by AI algorithm geomagnetic fields in this module. For convenience, the (2)
Magne c fields with AEKF
update with AEKF can be rewritten as follows:
Calibrated
Magnetometer Posi on is
predicted by
Trained AI 𝐁bk = 𝐑𝐂bnk 𝐘nk + 𝐛 + 𝛆 (3)
k

Fig. 2  Flow diagram of the proposed method in predicting mode

where
𝐑 = 𝐂sc 𝐂no 𝐂si (4)

matrix, scale factor matrix, and non-orthogonal matrix, 𝐛 = 𝐂sc 𝐂no 𝐛bi + 𝐛b (5)
respectively. 𝐘nk is the geomagnetic field at step k in n-frame, Based on (3), the model at step k − 1 can be written as
and it can be calculated by transforming B ⃗ from spherical
polar to rectangular coordinates. 𝐁bk−1 = 𝐑𝐂bnk−1 𝐘nk−1 + 𝐛 + 𝛆 (6)
So far, the above analysis gives the relation between mag-
k−1

netometer and position. However, according to (1), it is dif- From (1), we know that 𝐘nk is related to position and time.
ficult to calculate the position from the geomagnetic field. In the actual MEMS-INS system, the sampling frequency is
Based on this situation, we propose the AI learning algorithm about 200 Hz or even more. During the process from step
to build the position model by geomagnetic fields. The flow k − 1 to k, the position of vehicle system will not change
diagram of the proposed method is shown in Figs. 1 and 2. much. So, 𝐘nk is assumed to a constant during the process
from step k − 1 to k. Hence, 𝐂nk−1 𝐘nk−1 can be transformed
b
First, the magnetometer should be calibrated before use. k−1

The calibration methods have been introduced in many pub- by (6) as follows.
lications. For example, Wu et al. (2018) propose a dynamic
𝐂bnk−1 𝐘nk−1 = 𝐑−1 𝐁bk−1 − 𝐛 − 𝛆
( )
magnetometer calibration method to calibrate the mag- k−1
(7)
netometer by EKF, and at the same time, the method can
align with inertial sensors. Wu and Wang (2018) provide a Due to the prior knowledge of attitude update by a gyro-
scope, the relation between 𝐂nkk and 𝐂nk−1 can be obtained as
b b
magnetometer and gyroscope-integrated calibration method k−1

based on level rotation. Second, the gyroscope and a cali- (Wu et al. 2018)
brated magnetometer from MEMS-INS are integrated for
𝐂bnk = 𝐈3 − 𝛚bk × ts 𝐂bnk−1
( [ ] )
the accurate and stable magnetic fields in the b-frame with k k−1
(8)
AEKF. Then, the attitudes from the INS/GPS integration can
where 𝐈3 is an identity matrix of dimension three. 𝛚 k × is
[ b ]
be used to calculate the geomagnetic fields in the n-frame.
an anti-symmetric matrix from the output of gyroscope 𝛚bk
Finally, the geomagnetic fields and accurate position are put
at step k, and ts is the interval of data update. Combining (3),
into the predictor by AI for training.

13
88 
Page 4 of 11 GPS Solutions (2019) 23:88

(7), and (8), the magnetometer model using angular velocity Input Layer Hidden Layer Output Layer
from gyroscope can be obtained as
ϕ
𝐁bk = 𝐑 𝐈3 − 𝛚bk × ts 𝐑−1 𝐁bk−1 − 𝐛 + 𝐛 + 𝛆 (9) W1
( [ ] ) ( )
x1
Because the magnetometer has been calibrated in this mod- Wj Y
ϕ
ule, 𝐑 and 𝐛 in (9) are considered known constants. Then, 𝛚bk Σ
in (9) can be obtained by the gyroscope measurement model.
The simple gyroscope measurement model can be given as xp WN
(Wu et al. 2018) ϕ W0

̃ bgk = 𝛚bk + 𝐛bgk + 𝛆g

𝛚 (10)
Fig. 3  Block diagram of the RBFNN
where 𝛚 ̃ gk is the output of gyroscope and 𝐛gk is the gyroscope
b b

bias in the b-frame at the step k. 𝛆g is the measurement noise,

which follows zero-mean Gaussian white noise. 𝐱k = 𝐱k∕k−1 + 𝐊k 𝐕k (19)
Based on the above deduction, the magnetic fields and the
gyroscope’s bias are selected as the states of the filter in the 𝐏k = (𝐈 − 𝐊k 𝐇k )𝐏k∕k−1 (20)
gyroscope and]magnetometer integration system, that is, The state transition function f (⋅) can be provided by (11). 𝐇k
T
is the measurement matrix, and it is based on [measurement
[
𝐱k = 𝐁bk 𝐛bgk  . Hence, the state transition model and meas-
model in (13). The expression of 𝐇k is 𝐇k = 𝐈3 𝟎3  , 𝐈3 is
]
urement model of filter are as follows.
identity matrix with three dimensions, and 𝟎3 is zero matrix
with three dimensions. 𝚽 is the Jacobian matrix of f (⋅) ,
(a) State transition model
and 𝐐k is the process noise covariance matrix. The meas-
urements of the filter satisfy 𝐳k = 𝐁 ̃ bmk . For adaptively over-
( ( ) ) coming the random magnetic disturbances, the innovation
𝐁bk = 𝐑 𝐈3 − 𝛚̃ bgk − 𝐛bgk−1 × ts 𝐑−1 𝐁bk−1 − 𝐛 + 𝐛 + 𝛆
( )
covariance matrix 𝐂k will be calculated by the innovation
(11) vector 𝐕k and it is the average result of the sliding window
(b) Measurement model 𝐕k 𝐕Tk  . The window size is M  , which can improve the adap-
𝐛bgk = 𝐛bgk−1 (12) tive ability to overcome the inaccurate measurements. The
size M of the sliding window is usually set within 50–100.
̃ bk = 𝐁bk
𝐁 (13)
where 𝐁 ̃ bk is the output from the magnetometer. In this mod- Predictor by AI
ule, the AEKF is chosen as adaptive fuse algorithm to filter
disturbances (Fang and Yang 2011). The key of the proposed positioning method is to predict posi-
The process of AEKF is as follows. tion using the magnetometer during GPS outages rather than
an advanced AI algorithm. For the sake of generality, we adopt
(a) Time update a classic neural network (NN) to verify the effect. There are
two main types of models in common NN, back propagation
neural network (BPNN) and radial basis function neural net-
( ) work (RBFNN). Yang et al. (2010) analyzed the characteris-
𝐱k∕k−1 = f 𝐱k−1 (14)
tic of RBFNN and BPNN. They obtained a conclusion that
RBFNN-based model has more high fitting ability than BPNN.
(b) Innovation covariance estimation Because a better fitting ability is mostly needed in predicting
𝐏k∕k−1 = 𝚽k,k−1 𝐏k−1 𝚽Tk,k−1 + 𝐐k (15) position, RBFNN is selected to fit the model between geomag-
netic fields and position. Figure 3 shows the block diagram of
𝐕k = 𝐳k − 𝐇k 𝐱k∕k−1 (16) RBFNN.
As shown in Fig. 3, RBFNN includes an input layer, a hid-
(c) Measurement update
M den layer, and an output layer. Each hidden layer is used to
1 ∑
𝐂k = 𝐕 𝐕T (17) satisfy a radial basis function. The most common radial basis
M k=k k k function is the Gaussian function, and its representation is as
0

follows:

𝐊k = 𝐏k∕k−1 𝐇Tk 𝐂−1 (18) ‖x−𝜇‖2

(21)
k 𝜑(x, 𝜇) = e− 2d2

13
GPS Solutions (2019) 23:88 Page 5 of 11  88

where x is the input of the network, 𝜇 is the center of Gauss- GPS is available INS/GPS integra on
ian function, and d is the radius from the center of 𝜑(x, 𝜇) . Velocity with KF
A tude
GPS Predictor by AI
Through the training, x and 𝜇 can be determined. Posi on
update
available Predictor for
The output O(x) of RBFNN is given as Posi on Training A tude
MEMS-INS update Trained Posi on
Predictor
N Accelerometer Integra on
∑
Algorithm with KF GPS outages
O(x) = W0 + Wj 𝜑j (x) (22) Gyroscope
j=1

where W is the vector of weight between the input layer and Fig. 4  Diagram of the comparison method
output layer, and W0 is the bias.

INS/GPS integration with KF (a) State vector

When GPS is available, INS needs to be integrated with GPS

for more accurate attitude and position. In this module, the 𝐱=
[
𝜙E 𝜙N 𝜙U 𝛿VE 𝛿VN 𝛿L 𝛿𝜆 𝜀x 𝜀y 𝜀z ∇x ∇y
]T
KF is employed. The state vector includes navigation errors
and inertial sensor errors. The navigation errors consist of (26)
attitude errors 𝜙 , velocity errors 𝛿V  , and position errors (lati- (b) Measurement vector
tude 𝛿L , longitude 𝛿𝜆 ). The inertial sensor errors include the
gyroscope drifts 𝜀 and accelerometer bias ∇ . The state and ]T
measurement vectors can be written as follows: (27)
[
z = 𝛿L 𝛿𝜆

(a) State vector

[ ] [ ]
𝛿L LINS − LNN
[ ]T 𝛿𝜆
=
𝜆INS − 𝜆NN (28)
𝐱 = 𝜙E 𝜙N 𝜙U 𝛿VE 𝛿VN 𝛿L 𝛿𝜆 𝜀x 𝜀y 𝜀z ∇x ∇y
(b) Measurement vector (23) where the subscript INS expresses that the value is cal-
culated by the INS and the subscript AI represents that
the value is provided by AI. Then, AEKF is employed to
𝐳=
[
𝛿VE 𝛿VN 𝛿L 𝛿𝜆
]T
(24) increase adaptability, and the form of AEKF has been intro-
duced in the module of magnetic fields update with AEKF.
[ ] [ ] [ ] [ ]
𝛿VE VEINS − VEGPS 𝛿L LINS − LGPS
= , =
𝛿VN VNINS − 𝜆NGPS 𝛿𝜆 𝜆INS − 𝜆GPS
(25) Results
where E, N, U represent east, north, and upward in n-frame.
The subscripts x, y, z denote the right, forward, and upward In this section, the proposed method will be verified by
in b-frame. The subscript INS expresses that the value is simulation data and an actual driving test. Meanwhile, an
calculated by the INS and the subscript GPS shows the value existing method will be introduced and compared with the
is provided by GPS. Because we focus on the vehicle naviga- proposed method.
tion system, the height, vertical velocity, and accelerometer
bias in z-axis can be ignored.
Comparison method
INS/position integration with AEKF
In INS/GPS integration, the existing methods concentrate on
Because the positions from trained AI may be discrete and the relation between position errors and the outputs of INS to
have a large noise, it needs to be integrated with INS for bridge GPS outages (Semeniuk and Noureldin 2006). In the
more accurate positions. This module is similar to the INS/ comparison method, the state and measurement vectors are
GPS integration with the KF module, but the accuracy of the same as in the module INS/GPS integration with AEKF
position decreases and the measurements do not include described above. RBFNN is still selected as the AI learning
velocity. So, the state vector includes navigation errors and method. The flow diagram is shown in Fig. 4. When GPS is
inertial sensor errors in this module. The measurements only available, the velocity and position are calculated by INS,
consist of position errors. The state and measurement vec- and then, the measurements of INS and the errors of posi-
tors can be written as follows: tion and velocity are put into AI for training. During GPS

13
88 
Page 6 of 11 GPS Solutions (2019) 23:88

Fig. 5  Simulation route Fig. 7  Simulation attitude

Fig. 6  Simulation position Fig. 8  Simulation velocity

outages, the position and velocity errors can be obtained by Table 1  Simulation parameters of MEMS-INS and GPS system
the trained AI, and then, these errors will be put into KF as Items Parameters
measurements to compensate position and velocity.
Sampling rates 205 (Hz)
Simulation results Gyroscope output noise 0.01(°/s) (rms)
Accelerometer output noise 2 (mg) (rms)
In this part, we will use simulation to verify the effectiveness Magnetometer output noise 0.5 (mGauss) (rms)
of the proposed method. First, we will introduce the simula- GPS output noise 3 (m) (rms)
tion data and the parameters of the simulation system. Then,
we will show the simulation results.
of the first two laps is available, which is about 500 s, and
then, GPS will become invalid. The parameters of MEMS-
Simulation setting INS and GPS are set based on the actual system; the detail
is summarized in Table 1.
The simulation is employed to evaluate the effect of the
proposed method. The simulation route is shown in Fig. 5. Simulation results
Figure 6 indicates the position information of simulation.
Figures 7 and 8 show the attitude and velocity of simulation. Figures 9 and 10 show the results of the simulation. In
From these figures, we can see that the total time is 3600 s Fig. 9, the red line is the reference route, the brown line is
and the vehicle moves around a circle in the first 2300 s and the route calculated only by INS, and the blue line repre-
then stops until the end of the test. Such movement pattern sents the result by the comparison method. From Fig. 9, we
can verify the static and dynamic positioning accuracy at know that the position errors will be accumulated if INS
the same time. In this simulation data, the GPS information is directly used to calculate position during GPS outages.

13
GPS Solutions (2019) 23:88 Page 7 of 11  88

Fig. 9  Simulation results of the comparison method

Fig. 11  Results of position errors in the test (top) with partial

enlargement (bottom)

accumulate. At the end of the simulation, the position

Fig. 10  Simulation results of the proposed method
error of the comparison method has diverged. By contrast,
although the positioning error will increase a little with the
proposed method, the position errors will not accumulate.
On the contrary, the comparison method can restrain error
accumulation effectively by using RBFNN to predict the Actual test results
errors of velocity and position in a short time, but it still
leads to divergence for a long time. In Fig. 10, the red line In this part, we will use the actual test to verify the effec-
still shows the reference route, the green points are the tiveness of the proposed method. First, we will introduce
predicted result by RBFNN, and the blue line represents the route of test and the parameters of the actual integrated
the results of the proposed method. We can see the truth navigation system. Then, we will show the test results. By
that RBFNN can predict the positions, but these predict comparing the results of the comparison method, the pro-
results are discrete and contain the noise. After fusing the posed method will be verified.
predicted results of RBFNN and INS by AEKF, the pro-
posed method can obtain stable, accurate, and continu- Test setting
ous positions. Figure 11 shows the comparison diagram
of position errors. An actual MEMS-INS integration is employed to verify the
When GPS is available, the two methods have a similar effect of the proposed method. In this test, an ADIS16488
result in Fig. 11. During GPS outages, the comparison unit from Analog Devices company is taken as the MEMS-
method can provide an accurate position within a short INS system. U-blox LEA-M8S is chosen as the GPS module
time. But as time goes on, the position errors would in Fig. 12. Their features are shown in Table 2.

13
88 
Page 8 of 11 GPS Solutions (2019) 23:88

Fig. 14  Attitude of the system

Fig. 12  ADIS16488/U-blox LEA-M8S integration system

Table 2  ADIS16488/U-blox LEA-M8S integration features

Items Parameters

Sampling rates 205 (Hz)

Gyroscope in-run bias stability 6.25 (°/h)
Gyroscope output noise 0.16 (°/s) (rms)
Accelerometer in-run bias stability 0.1 mg
Accelerometer output noise 1.5 (mg) (rms)
Magnetometer output noise 0.45 (mGauss) (rms)
Position accuracy 2.5 m (CEP)

Fig. 15  Position of the system

From Figs. 13, 14 and 15, we can know that the system

moves around a closed loop on campuses. In the test, we
set GPS available during the first two circles, just like in
the simulation, that is, about 1000 s. Then GPS will be set
invalid until the end of the test.

Test results

The results of the driving test are summarized here. Because

the environments of the actual test are very complex and
there is more noise and errors, the accuracy of the actual test
may be worse than simulation.
Fig. 13  Actual driving test route Figures 16 and 17 show the route results. Figures 18 and
19 show the position results. From Fig. 16, the route cal-
culated only by INS diverges rapidly. In contrast, the route
The driving test route is shown in Fig. 13. Figure 14 of the comparison method restrains the divergence to some
shows the attitude of the system. The changes in position extent at the early stage of GPS outages. But, since the
are shown in Fig. 15. prediction of the comparison method is the position error
rather than the position, the inaccurate position errors will

13
GPS Solutions (2019) 23:88 Page 9 of 11  88

Fig. 16  Route results of the test by the comparison method Fig. 19  Position results of the test by the proposed method

be accumulated. Hence, it diverges eventually from Fig. 18.

Compared with Fig. 16, the proposed method has a good
positioning accuracy in Fig. 17. The position can be pre-
dicted by RBFNN, but it may have errors and noises. After
integrating with INS in the proposed method, the positions
basically follow the GPS positions and there is no tendency
to diverge at all in Fig. 19.
Figure  20 shows the position errors. From these two
figures, we can know that both the proposed method and
comparison method have a good positioning accuracy dur-
ing GPS is available. During GPS outages, the comparison
method can provide an accurate position for a short time, but
it finally diverges. By contrast, the proposed method basi-
cally maintains the accurate position for a long time. Also,
the errors of position are mostly kept within 200 meters,
Fig. 17  Route results of the test by the proposed method even if there are some errors caused by large magnetic dis-
turbances. To summarize, the test result is similar to the
simulation, which can prove the effectiveness of the pro-
posed method.

Conclusion

For reducing the position errors during GPS outages in the

GPS/INS integrated system, the existing methods mainly
focus on the advanced filter or AI learning methods to pre-
dict the more accurate position errors during GPS outages.
However, sooner or later, that will lead the position errors to
diverge as the GPS outages time continues, no matter how
Fig. 18  Position results of the test by the comparison method
advanced the intelligent algorithm.
In order to get a long-time non-diverging position when
GPS outages, we propose a positioning method based on
INS/magnetometer integration during GPS outages as
they are usually integrated into a MEMS-INS unit. Since
the existing usage of magnetometer mainly focuses on

13
88 
Page 10 of 11 GPS Solutions (2019) 23:88

training is ensured. The datasets should collect geomagnetic

fields and positions as much as possible and the data should
be uniformly distributed over the target area. Simulation and
actual driving test verify the effectiveness of the proposed
method.
Above all, we propose a positioning scheme using mag-
netometer and NN. The role of NN is to build a relationship
between the geomagnetic field and positions. However, some
problems exist in practice. The proposed method mainly
focuses on some indoor environments in the city, such as
tunnel, underground parking and so on. An NN correspond-
ing to a target area can be used for different INS/magnetome-
ter system entering the area, but it is hard to be used for other
area. The training process requires a large amount of data.
These problems will be subject of research in the future.

Acknowledgements  This work is supported by the National Natural

Science Foundation (61571148, 61871143), Fundamental Research
for the Central University (HEUCFG201823, 3072019CF0402),
Heilongjiang Natural Science Foundation (LH2019F006), Research
and Development Project of Application Technology in Harbin
(2017R-AQXJ095).

References
Abdolkarimi ES, Abaei G, Mosavi MR (2018) A wavelet-extreme
learning machine for low-cost INS/GPS navigation system in
high-speed applications. GPS Solut 22(1):1–13
Fig. 20  Results of position errors in the test (top) and partial enlarge- Adusumilli S, Bhatt D, Wang H, Bhattacharya P, Devabhaktuni V
ment (bottom) (2013) A low-cost INS/GPS integration methodology based on
random forest regression. Expert Syst Appl 40(11):4653–4659
Adusumilli S, Bhatt D, Wang H, Devabhaktuni V, Bhattacharya P
(2015) A novel hybrid approach utilizing principal component
assisting in calculating attitude especially heading rather regression and random forest regression to bridge the period of
than position, we first verify the possibility of using mag- GPS outages. Neurocomputing 166:185–192
Fang J, Yang S (2011) Study on innovation adaptive EKF for in-
netometer for positioning by analyzing the IGRF model
flight alignment of airborne POS. IEEE Trans Instrum Meas
and magnetometer measurement model. At the same 60(4):1378–1388
time, the conclusion is obtained that it is difficult to use Noureldin A, El-Shafie A, Bayoumi M (2011) GPS/INS integration
the magnetometer directly for positioning since it is easy utilizing dynamic neural networks for vehicular navigation. Inf
Fusion 12(1):48–57
to be disturbed and it is hard to calculate positions form
Semeniuk L, Noureldin A (2006) Bridging GPS outages using neural
the geomagnetic fields using IGRF. Hence, we propose a network estimates of INS position and velocity errors. Meas Sci
magnetometer-assisting positioning solution. This solu- Technol 17(10):2783–2798
tion includes four parts, magnetic fields update module Thébault E et al (2010) International geomagnetic reference field: the
eleventh generation. Geophys J Int 183(3):1216–1230
with AEKF, predictor by AI, INS/GPS integration with Wang W, Wu Z, Zhang H (2018) An adaptive cascaded Kalman fil-
KF, and INS/position integration with AEKF. In predictor ter for two-antenna GPS/MEMS-IMU integration. In: IEEE/ION
by AI, the classical RBFNN is selected as an AI learning PLANS 2018, Institute of Navigation, Monterey, CA, USA, 23–26
algorithm for generalization. April, pp 833–837
Wu Y, Pei L (2017) Gyroscope calibration via magnetometer. IEEE
In the simulation and driving test, the GPS/INS integrated Sens J 17(16):5269–5275
positioning solution based on RBFNN is taken as the com- Wu Z, Wang W (2018) Magnetometer and gyroscope calibration
parison method. The results of simulation and driving test method with level rotation. Sensors 18(3):748
show that the comparison method only holds positioning Wu Y, Zou D, Liu P, Yu W (2018) Dynamic magnetometer calibration
and alignment to inertial sensors by Kalman filtering. IEEE Trans
accuracy over a short time and finally diverges. By contrast, Control Syst Technol 26(2):716–723
the proposed method can provide the position that shows no
trend of divergence for a long time if a sufficient dataset for

13
GPS Solutions (2019) 23:88 Page 11 of 11  88

Xu Q, Li X, Chan CY (2018) Enhancing localization accuracy of Wei Wang  received a PhD in

MEMS-INS/GPS/In-vehicle sensors integration during GPS out- Navigation, Guidance, and Con-
ages. IEEE Trans Instrum Meas 67(8):1966–1978 trol from Harbin Engineering
Yang B, Lu H, Chen L (2010) BPNN and RBFNN based modeling University, China, 2005. He is a
analysis and comparison for cement calcination process. In: professor at Harbin Engineering
IWACI 2010, 3rd international workshop on advanced compu- University and published more
tational intelligence, Suzhou, Jiangsu, China, 25–27 August, pp than 100 refereed journal and
101–106 conference papers. He is a Senior
Zhang Y, Shen C, Tang J, Liu J (2018) Hybrid algorithm based on Member of the Institution of
MDF-CKF and RF for GPS/INS system during GPS outages Electrical and Electronic Engi-
(April 2018). IEEE Access 6:35343–35354 neers (MIEEE) and Member of
Zheng J, Zhao W, Han B, Wen Y (2016) Integrating extreme learning the Society of Aerospace and
machine with Kalman filter to bridge GPS outages. In: ICISCE Electronic Systems. His current
2016, 3rd international conference on information science and research interests include Inte-
control engineering, Beijing, China, 8–10 July, pp 420–424 grated Navigation and Signal
Processing for MIMO Radar.
Publisher’s Note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.

Zongkai Wu  is currently pursuing

the PhD Degree in the School of
Control Science and Engineer-
ing, Harbin Engineering Univer-
sity, Harbin, China. His current
research interests include an
inertial navigation system, inte-
grated navigation system, and
multi-sensor information fusion.

13