1. Introduction
High dynamic vehicles require resilient navigation solutions. Indeed, the Global Navigation Satellite System (GNSS), which is the leading technology of positioning navigation and timing, is sensitive to jamming and spoofing. To cope with such a situation, the GNSS community has developed interference mitigation techniques, but another complementary solution is that which uses an alternative technology to GNSS when it becomes inoperative.
This work deals with the navigation using signals of opportunity (SOP), more specifically the signals of the Iridium Next Communication System. SOP are all radio frequency (RF) signals that are not intended for navigation purposes. They are signals of communication systems (mobile networks, satellite-based communication systems), TVs, AM/FM radios, radar, etc. The advantages of SOP in navigation are the existing infrastructure that is free of use, the high signal power level compared to that of GNSS signals, and the high frequency diversity if many systems are used. The disadvantages of SOP stem from the fact that the signals are not optimized for navigation purposes. That is, the signals’ availability is not guaranteed everywhere, signals from different transmitters (including those of the same system) are not synchronized in time and their clock stability is lower than that of GNSS satellites, and transmitter positions are unknown. In addition, if different systems are coupled, one needs multi-band antennas, multi-band RF front-end and a sufficient computing power.
Despite these constraints, finding an alternative or an augmentation technology to GNSS is crucial. The research in the topic of navigation based on SOP essentially started two decades ago, and has gained interest over the last few years, especially with the announcement of the advent of future new mega LEO satellite (LEO SV) constellations, Starlink and OneWeb, and the modernization of the Iridium constellation (which became Iridium Next in 2017). The popularity of LEO SV systems for navigation is mainly thanks to their global coverage not only on Earth, but also in its surrounding space, which allows high-altitude vehicles to be potential users.
In this paper, we study the potential of the Iridium Next LEO SV signals for the navigation of a high dynamic vehicle. The Iridium Next system offers voice and data communication. The company Satelles used the messaging to transmit bursts designed specifically for the navigation in a PNT solution called Satellite Time and Location (STL) [
1]. Since positioning and namely timing capabilities have been demonstrated based on the STL technology, in this work, we will use the Iridium Next signals as signals of opportunity—that is, no navigation specific data are extracted from the signal in space.
Navigation using LEO SV signals as SOP has been the focus of many papers. In [
2], a differential positioning using time difference of arrival (TDOA) and frequency difference of arrival (FDOA) is implemented. Experiments showed a positioning accuracy of 25 m with 5 min data collected in a static location. In [
3], the authors studied a positioning using LEO SV Doppler shifts that are loosely fused with an altimeter height. Experiments, using 1 min data from two Orbcomm satellites and by assuming a known altitude, exhibited a 2D positioning error of 358 m for a static observer. In [
4,
5], the authors derived the performance of LEO SV measurements fused with an inertial navigation system (INS). A range of information from LEO SVs is assumed to be available in addition to the Doppler shifts. The algorithm refines the satellite position and velocity estimates by including them into the state in addition to satellite clock biases and drifts. The global positioning system (GPS) is only used at the beginning, and then the navigation filter is updated by LEO SV and altimeter height measurements. An experiment using 4 min data of 2 Orbcomm SV led to a positioning accuracy with a root mean square error (RMSE) of a few hundred meters for a land vehicle. In [
6], positioning using Iridium signals in forest canopy is demonstrated. Experiments showed a height aided static positioning accuracy of a few hundred meters obtained with 30 min data. In [
7], the authors showed that the satellite position accuracy, based on the Simplified Perturbation model (SPG4) fed with Two Line Element (TLE) files, can be as high as 3 km, and the velocity accuracy can be as high as 3 m/s. They showed that the two-body model using the second gravitational zonal coefficient J2 allows for a better satellite positioning, if a good initial satellite position is available. In [
8], the authors presented a Doppler-based positioning using Ku-band Starlink signals. A 3D positioning better than 23 m was shown by the experiment in which the TLE epochs were adjusted to cope with the ephemeris errors. In [
9], a Multi-Constellation Software-Defined Receiver was designed to measure the Doppler shifts of LEO SV downlink signals. The paper showed by means of the experiment the benefits of the multi-constellation (Orbcomm and Iridium systems) for the positioning. A multi-constellation LEO SV signal receiver was also presented in [
10]. The demonstrated positioning accuracy is 22.7 m, obtained by assuming a known altitude of the user, and when up to four LEO SVs were tracked from Iridium and Orbcomm systems. In [
11], the Doppler shift is measured on signals with orthogonal frequency-division multiplexing (OFDM) modulation, which is expected for the deployment of future LEO SV. Experiments using new 5G terrestrial radio signals showed an RMSE of 6.45 Hz in the estimation of the Doppler shift. In [
12], a carrier phase-based positioning on LEO SV signals was designed. Then a 3D positioning accuracy of 33.5 m was achieved by the experiment using six Starlink SV signals.
The synthesis of this existing work shows that the navigation is typically performed based on the Doppler shift measurement, which is a straightforward observable that can be obtained from a radio signal transmitter with a known carrier frequency. A first category of LEO SV-based navigation techniques uses only LEO SV signals. While the positioning is usually standalone, the differential mode involving two observers has also been explored. A second category performs the navigation based on the coupling between LEO SV and an INS (based on gyroscopes and accelerometers). In general, the tight coupling is considered due to the poor availability of the LEO SVs. In both categories, the navigation is either single- or multi-constellation (involving two or more LEO SV systems) and is eventually height-aided.
These developments and similar work [
13,
14,
15,
16] demonstrated the potential of LEO SV signals for the opportunistic positioning. In this work, we will address the following complementary tasks
Evaluate the potential of LEO SV SOP for a high dynamic vehicle with a short-duration mission. The short duration is expected to prevent achieving an optimal navigation performance.
Design a filter measurement model based on magnetometer outputs and LEO SV Doppler shift that allows for observing the position, velocity, attitude, and receiver clock bias and drift.
Use of representative Iridium data for performance simulation. Indeed, the Doppler shift measurements are obtained by processing Iridium-like signals by an Iridium signal-processing tool. These Doppler shift data are more representative than those obtained by simply adding an error distribution to the geometric Doppler shift between a satellite and a user terminal.
Evaluate the performance of attitude estimation in addition to that of positioning. Note that the attitude and especially the roll angle can be used to roll-rate stabilize the vehicle, and in guidance [
17].
Evaluate the influence of the individual error sources on Doppler-based navigation. The errors consist mainly of the satellite position and velocity errors, the satellite clock drift, and thermal noise.
Benchmark the performance of the system against an INS based on a tactical-grade IMU and magnetometers.
This paper is organized as follows. After this introductive section (
Section 1),
Section 2 defines the tight coupling filter between IMU, LEO SV Doppler shift and magnetometer data.
Section 3 presents the simulation and signal processing tools developed to conduct this study.
Section 4 studies the performance of the positioning and the attitude estimation in which the effect of each error source is determined, and the performance in a typical intermediate case and worst case are derived and compared to an INS solution.
Section 5 concludes the paper.
2. The Navigation Filter
The navigation filter implements a tight coupling between IMU, the Iridium Doppler shift and the magnetometer (
Figure 1). Here, we take advantage of the complementarity between the inertial navigation that has a good performance in the short term but diverges with time, and the Doppler measurements from Iridium satellites that are very noisy in the short term but should be relatively stable at the long term. This algorithm is different from the usual GNSS/IMU tight coupling in three aspects: (1) it uses only LEO SV Doppler shift measurements, not pseudoranges; (2) the measurements from satellites in view over the same period of time come at different instances as the corresponding bursts do, and thus the innovation at each time update contains usually a single Doppler shift measurement, in addition to the Earth’s magnetic field measurements; and (3) the use of magnetometers for a direct attitude observation, especially when dealing with a spinning vehicle.
We define the local frame (
O,
x,
y,
z) such that the
x-y plane is tangent to the Earth ellipsoid at point
O, with (
Ox) towards a given direction (the geographic north for instance), (
Oz) is the vertical downward direction and (
Oy) completes the direct coordinate system. The body frame (
,
,
,
) is centred on the vehicle’s center of mass, with (
) being the vehicle longitudinal axis oriented towards the movement direction, and the
-
plane being perpendicular to (
). Both frames are visible in
Figure 2. The attitude angles (roll:
, pitch:
, yaw:
) are defined as the angles that allow the transformation between the body frame and the local frame. The three-axis accelerometer, the three-axis gyroscope and the three-axis magnetometer are supposed to be aligned with the body frame axes. The L1 band (1.626 GHz) antenna is supposed to be located at
. These assumptions are intended to simplify the filter model.
The navigation algorithm implements an extended Kalman filter (EKF). The state vector is formed by the position error
and the velocity error
, both in the navigation frame, the attitude error
, the accelerometer biases error
and the gyroscope biases error
, both in the body frame, the receiver clock bias error
and the receiver clock drift error
converted in m and in m/s, respectively:
is the signal propagation speed.
2.1. Dynamic Model
The filter dynamics can be written as [
18]:
where
is the matrix that allows the transformation between the body frame and the local frame.
where
and
are the cosine and sine functions, respectively.
is the specific force measured by the accelerometer and corrected for the accelerometer biases . is the skew-symmetric matrix operator. and are, respectively, the accelerometer and gyroscope noise processes. Sensor biases are modelled by a first order Gauss–Markov process with correlation times and and noises and , respectively, for the accelerometer and the gyroscope. The time derivative of the receiver clock bias error equals the clock drift error plus a Gaussian noise . The receiver clock drift error is a constant random process, and its time derivative is a Gaussian noise .
The system defined in (2) can be written as , with being the dynamic matrix and being the system noise matrix with covariance . The attitude is predicted by the integration of the differential equation , where = is the angular velocity vector in the body frame provided by the gyroscope and corrected for the gyroscope biases. The user acceleration in the local frame is obtained by , with being the local gravity vector. The receiver clock bias is predicted by the integration of the receiver clock drift . The velocity in the local frame is obtained by the integration of the acceleration . Then, the position in the local frame is obtained by the integration of the velocity . Finally, the state covariance matrix is computed according to , where is the state transition matrix and is the time step of the prediction.
2.2. Observation Model
The measurements consist of the Doppler shifts from LEO SVs (converted to range rates) and the Earth’s magnetic field from a triad of magnetometers.
2.2.1. Doppler Shift Model
The Doppler shift measurement is subjected to many error sources. The Multipath error is negligible for a vehicle travelling in an open sky. The Doppler shift induced by the ionosphere delay variation is higher than that of GNSS due to the faster variation of the ionosphere pierce point of the LEO SVs. Nevertheless, simulations we carried out using ionosphere grid maps showed that the ionosphere-induced Doppler shift stays largely below 1 Hz at 1.626 GHz. The tropospheric delay variation is dominated by the fast height variation of the vehicle rather than the elevation variation of the satellite. The usage of a tropospheric delay correction model (for instance, the MOPS [
19]) is known to compensate accurately for this error. The expected residual is therefore small enough compared to the Doppler shift measurement noise. In this study, we will ignore the Doppler shift induced by the atmosphere and by the multipath propagation.
The
kth-satellite range rate measurement,
(in m/s), is given by the opposite sign of the Doppler frequency shift
multiplied by the carrier wavelength
. It can be written as the relative velocity between the receiver’s antenna and the transmitter’s antenna projected onto the line of sight (LOS) unit vector
between them, plus the receiver clock drift (converted in m/s),
, minus the satellite clock drift,
(converted in m/s), and an additive Gaussian thermal noise
[
20]
The LOS unit vector
is written as:
where
and
are, respectively, the position and the velocity of the
kth-satellite at the time of burst transmission
, where
is the burst reception time and
is the propagation delay of the signal between the satellite antenna and the receiver antenna.
Usually, if the satellite clock drift
is known, it is used to correct the range rate measurement
. In our case, it is unavailable. The predicted range rate computed at the burst reception time can be written as:
In (6), we assumed that the reconstructed propagation time is close to the actual propagation time once the filter has converged, i.e., .
In order to linearize the range rate model, we define the following perturbations of the user position, user velocity, user time and user clock drift, respectively:
By inserting (7) into (6), we obtain
The satellite position and velocity at the biased emission time can be written as:
Inserting (9) and (10) into (8) yields
In (11), the reference to time
in
,
,
, and
is omitted for clarity. To develop the predicted range rate (11), we can write
In (12), we neglected the quadratic terms (i.e., and ) and we limited the development to the first order for small values of and .
Substituting
by
in (12), and then inserting (12) into (11), we obtain the following approximation of the range rate
The development of (13) by neglecting the second order terms yields
Therefore, the range rate error defined by
can be written as:
Finally, the range rate error of (15) can be written in the following form:
where the coefficients
,
and
can be easily deduced by identification between (15) and (16).
2.2.2. Magnetic Field Measurement Model
The earth’s magnetic field in the local frame is obtained from the magnetometer measurement
as follows
The magnetometer is supposed to be calibrated so that the measurement error is dominated by the sensor noise
as
We chose this model for the sake of simplicity. Note that the Earth’s magnetic field measured by a magnetometer is distorted, in addition to the sensor errors and their misalignment with the body frame axes, by the perturbations due to the ferromagnetic material composing the vehicle and by the current induced by the rotation of the metallic vehicle. In practice, those errors can be calibrated [
21,
22,
23,
24] and the residuals will add to the sensors bias and scale factor errors. A model that includes the magnetometer bias and scale factor in the filter can be found in [
25].
For small attitude errors
,
can be written as:
Inserting (18) and (19) into (17) yields
The earth’s magnetic field error in the local frame
is therefore
The reference Earth’s magnetic field in the local frame,
, can be obtained for instance from the World Magnetic Model [
26].
2.2.3. Update
Equations (16) and (21) allow us to express the observation errors
as a linear function of the error state
as
In (22),
is the observation model matrix given by
and
is the observation model noise given by
The noise covariance, , is a diagonal matrix formed by the variance of the Doppler shift noise and the variances of the three-axis magnetometer noises.
The predicted state
and the covariance matrix
are then updated as follows:
where
is the EKF gain given by
5. Conclusions
In this paper, we have studied the potential of Doppler shifts of LEO SV signals for navigation, when used as signals of opportunity. We defined a filter based on the tight coupling between a consumer-grade IMU, magnetometers and Doppler shifts. The filter was fed with representative sensors data, simulated for a high dynamic vehicle. We have characterized both position and attitude performances subject to thermal noise, satellite position and velocity errors, and satellite clock drift. This allowed for assessing and understanding the contribution of each error source to the navigation error. We have shown that, while the positioning error is large due to large orbit errors or high SV clock drifts, it becomes competitive with that of an INS based on a better quality IMU, if precise satellite orbits are available.
Contrary to the positioning, which is highly affected by large error sources, we have shown that the attitude estimation is less sensitive to large orbit errors and high SV clock drifts. For typical values of the error sources (the intermediate case), the defined SOP-based filter, which uses the SPG4 model fed with the broadcast NORAD TLE files, is able to track the vehicle’s attitude with a roll angle RMSE of ~1.6°, a pitch angle RMSE of ~2.2°, and a yaw angle RMSE of ~4.8°. These results illustrate the potential of the LEO SV signals of opportunity for the navigation of high dynamic vehicles in a GNSS-denied environment. Recently, the European Union (EU) announced that it will launch its own system of satellite-based high-speed internet [
31]. Since large satellite orbit errors and clock drifts are the main factors limiting the performance of the Doppler shift-based positioning, if such a system provides accurate satellite positions and clock drifts, then it could in addition offer a navigation service that will be very useful in GNSS-denied environments.