589
1 INTRODUCTION
Paradoxically, the ubiquity of and trust in GNSS makes
these systems vulnerable to interference. Signals
transmitted by satellites reach Earth with a power of
the order of picowatts - making them easily drowned
out by stronger radio emissions [4]. In addition, open
civilian GNSS signals (GPS, Galileo, GLONASS,
BeiDou) are not encrypted or authenticated, which
means there are no built-in safeguards against
imitation. This raises two main threats: jamming and
spoofing. Jamming involves intentionally generating
radio interference on GNSS frequencies so that the
receiver is unable to receive the true signal from the
satellites [10]. The result of such an action is the loss or
severe degradation of satellite navigation performance
- the receiver may report a missing signal or determine
a position with very low accuracy. Spoofing, on the
other hand, is a more sophisticated attack in which an
attacker broadcasts falsified signals that mimic real
GNSS messages so that the receiver assumes them to
be authentic [10]. In this way, erroneous location or
time information can be fed to the receiver, potentially
directing the “victim” to the wrong location or
misinforming time-dependent systems (such as power
grids or financial networks). Interference with GNSS
signals is not just a theoretical threat, but a real
problem noted with increasing frequency around the
world. Reports from the maritime and aviation services
indicate an increasing number of jamming and
spoofing incidents, especially in geopolitically unstable
regions [10]. For example, numerous GPS signal
disruptions have been reported in recent years for
ships in the Eastern Mediterranean, the Black Sea or the
Persian Gulf. Civil aviation has also seen an alarming
increase in spoofing attacks, with pilots reporting
The Use of Artificial Intelligence in Enhancing
Navigation Safety in Case of GNSS Signal Fading
or Interference
M. Chrzan
Kazimierz Pulaski University in Radom, Radom, Poland
ABSTRACT: Global Navigation Satellite Systems (GNSS) play a key role in modern navigation and transportation
service delivery. Today it is difficult to imagine the daily operation of land, sea or air transportation without the
positioning and timing provided by satellite systems [6, 8, 11]. In civil aviation, GNSS provides precision landing
approaches and support for safety systems, in maritime navigation it enables open sea course determination and
maneuvering, and on land it directs vehicle navigation and synchronization of telecommunications networks.
Many of these applications are so-called PNT (Positioning, Navigation and Timing) systems, where uninterrupted
and accurate position and timing information is critical to safe operations [8]. Unfortunately, satellite systems
have recently become the target of hacking attacks. This paper will present AI methods for detecting GNSS
anomalies, the role of inertial systems as an independent navigation source, and position correction techniques
(e.g., using Kalman filters) supported by intelligent algorithms. An overview of current research and experiments
in this field will also be presented, as well as conclusions on the effectiveness and development prospects of the
technologies discussed.
http://www.transnav.eu
the International Journal
on Marine Navigation
and Safety of Sea Transportation
Volume 19
Number 2
June 2025
DOI: 10.12716/1001.19.02.31
590
instances where onboard systems gave incorrect
coordinates or altitude, posing an obvious risk to flight
safety. Spoofing applied to an aircraft can result in
erroneous navigation, and in extreme cases - if the crew
does not figure it out in time - even lead to a violation
of airspace or an attempted landing on a false path. In
military applications, the consequences are equally
serious: a spoofed signal can result in military units
losing orientation or taking control of the unmanned
aircraft, which is a significant breach of operational
security [8, 10].
The reason for the vulnerability of GNSS systems to
attacks is the lack of mechanisms to verify the
reliability of the signal - most receivers trust the data
received from satellites implicitly. Unlike, for example,
computer networks, where there are firewalls and
authentication systems or protocol verification
procedures, a typical GNSS receiver has no way to
distinguish a genuine signal from a crafted one until
additional detection methods are put in place. The fact
that full documentation of GNSS protocols is publicly
available makes it easier for potential attackers to
prepare hardware and software for an attack. What's
more, advances in technology and engineering have
made the “barrier to entryfor attackers significantly
lower - nowadays, for a few hundred dollars, one can
get a software-defined radio (SDR) capable of
generating signals that interfere with or mimic GNSS
[11]. While this used to require specialized knowledge
and expensive equipment, today there are open-source
tools and instructions available on the Internet. All of
this demonstrates the importance of ensuring
navigation safety despite GNSS interference.
Improving the resilience of navigation systems to
GNSS interference has thus become a research and
engineering priority in recent years. Multi-layered
strategies are being introduced: from improving the
receivers themselves (e.g., anti-jam antennas, signal
authentication in new systems) to system approaches
using redundant sources of position information and
artificial intelligence algorithms to detect and
compensate for anomalies. Particularly promising is
the use of artificial intelligence (AI) - machine learning
and deep learning techniques - to analyze navigation
data to quickly catch unusual system behavior and
sustain navigation in the event of satellite signal loss.
2 LITERATURE RESEARCH AND REVIEW
2.1 Application of artificial intelligence in GNSS anomaly
detection
Effectively recognizing that a GNSS signal has been
disturbed or falsified is the first step to ensuring
navigation safety. Traditional GNSS receivers have
limited mechanisms to warn of anomalies - Receiver
Autonomous Integrity Monitoring (RAIM) systems,
for example, can detect that a pseudo-distance
measurement deviates from others, but against
specialized spoofing attacks they are sometimes
powerless. Therefore, more advanced detection
methods are being developed, often using multi-sensor
analysis and intelligent algorithms. In general, GNSS
anomaly detection approaches can be divided into
several categories: signal-based methods, bit/data-
based methods, positional methods, and methods
using machine learning [11].
Signal methods involve monitoring the parameters
of the GNSS signal received by the antenna - such as
signal power, signal-to-noise ratio (C/N0), or reception
characteristics on multiple antennas. Jamming
interference often manifests itself as a sharp drop in
C/N0 across all channels and loss of satellite tracking,
while spoofing can cause unnatural changes in power
(e.g., a sudden increase in signal strength seemingly
from all satellites simultaneously) or abnormal
behavior of the receiver's synchronization block. Signal
detection can use simple threshold tests or more
complex signal processing algorithms - such as
observing the signal spectrum in the GNSS band to
identify additional unwanted signals.
Bitwise methods focus on the content of navigation
data demodulated from the signal. The receiver can
verify that the almanac/efemerid data or time marks
are consistent and as expected. False signals may
contain minor deviations in the data (e.g., invalid
satellite identifier, repeated data string - characteristic
of meaconing, i.e., retransmission of a real signal with
a delay). Analysis of this data, supported, for example,
by correlation techniques between multiple signals,
makes it possible to detect certain types of attacks. For
example, one method uses monitoring the cross-
correlation of multiple GNSS observations (pseudo-
distance, carrier phases, etc.) and looking for statistical
outliers, which then serve as input features to a
machine-learned detector [12]. In experiments, it has
been shown that a trained classifier (e.g., SVM -
Support Vector Machine) fed with such features can
distinguish spoofing/meaconing situations from
normal ones, effectively alerting to an attack.
Importantly, this approach was verified on real
incident data - the algorithm learned on simulation
data also recognized recordings of real disturbances
with high efficiency [12].
Positional methods are based on checking the
consistency of the determined position with other
information. For example, a vessel equipped with
GNSS and radar or lidar can compare independent
position measurements: if the GPS shows that the
vessel has moved a few hundred meters, while the
radar still sees the shore at a constant distance - that's a
sign of a problem. Another approach is dynamic
control: monitoring whether the calculated movement
trajectory is physically possible (e.g., no sudden
“jumps” in position that exceed the vehicle's
achievable acceleration). In practice, many such
symptoms are used: a loss-of-position alarm, sudden
course deviations observed on an electronic map
compared to the course from radar, jumps in speed or
coordinate readings while the HDOP is low (which
normally means good satellite geometry) - all of these
warning signals can indicate GNSS interference [3].
Operators and systems can also cross-check position
from different systems (e.g., GNSS vs. local systems
like LORAN or cellular network position) to look for
discrepancies.
Methods using artificial intelligence (AI) often
combine elements of the above approaches,
simultaneously analyzing multiple parameters and
signals to detect subtle anomalies invisible to the naked
eye. Machine learning, particularly deep learning (DL),
591
has proven to be extremely effective in GNSS
interference detection tasks in recent years [8, 11].
These algorithms can learn from large data sets
characteristic patterns indicative of jamming or
spoofing. For example, a neural network can
simultaneously take into account temporal changes in
signal power, pseudo-distance errors, trajectory
deviations, and other characteristics, and on this basis
classify the state of the system as normal or attacked.
Various machine learning (ML) models have been
compared in the literature for spoofing detection -
results indicate that already relatively simple models,
such as decision trees (CART), achieve very good
performance in identifying attacks, outperforming
classical methods based on rigid thresholds [11]. More
advanced approaches use convolutional or recurrent
Long Short-Term Memory (LSTM) networks to analyze
raw streams of signals or strings of observations. The
latest experiments using deep networks have already
achieved nearly 99% success rate in detecting jamming
attacks on available datasets, a few percent
improvement over earlier algorithms. Such high
results show that AI is capable of very accurately
distinguishing between real and jamming signals
based on even small deviations in patterns.
In practice, AI-based GNSS anomaly detection
systems can operate in continuous receiver monitoring
mode. For example, on board an aircraft, an intelligent
module analyzes the data stream from a GPS receiver
(position, speed, time, signal parameters) and
compares it with data from an inertial navigation
system or other sensors. If the AI detects discrepancies
beyond normal conditions (e.g., much larger
differences between the position from GPS and inertial
than usual, or the simultaneous shift of all pseudo-
distances characteristic of spoofing), it can alert the
system to a suspected attack in a fraction of a second.
This approach is much faster and more reliable than
manual observation by an operator, especially since
combined attacks (e.g., simultaneous jamming and
spoofing) can manifest themselves in complex ways
that are difficult to catch with traditional methods [11].
Artificial intelligence enables learning by example in
this context - the system can be trained on a variety of
attack scenarios, so it will know what the symptoms of
a given threat look like.
It is worth noting that the effectiveness of AI
methods strongly depends on the availability of
adequate training data. The scientific community has
several public GNSS attack datasets (e.g., TEXBAT and
OAKBAT containing recordings of various spoofing
scenarios) used for algorithm validation [11]. However,
there is still a lack of open datasets covering a wider
range of situations (different types of simultaneous
interference, other environmental conditions, etc.),
which is the subject of active work. Despite these
challenges, current research results show that the
application of AI significantly improves the ability of
navigation systems to quickly detect and classify GNSS
anomalies, thus providing a foundation for activating
emergency mechanisms to protect traffic safety.
2.2 Role of the inertial systems in navigation
When a GNSS signal fades, the natural salvation is to
resort to other sources of motion information. For
decades, a standard component of professional
navigation systems has been an inertial system (INS) -
that is, a device that measures vehicle acceleration and
angular velocity based on a set of motion sensors [6].
A typical INS consists of an inertial measurement
unit (IMU) containing a three-axis accelerometer and a
three-axis gyroscope, whose readings are processed in
real time by a navigation computer. Operating
completely independent of external signals, the INS
performs so-called dead-reckoning navigation - from a
known initial position, using measurements of
acceleration and rotation, it calculates the new position,
speed and orientation of the vehicle in real time
through a process of integrating (integrating) these
accelerations over time [2]. As a result, INS provides
continuous navigation information even in the
complete absence of external signals (GPS, radio
beacons, etc.), making it invaluable in conditions of
interference or where satellite systems do not reach
(e.g. underwater, in tunnels) [6].
Inertial systems come in different classes of
accuracy. High-end INS, used, for example, in
commercial aviation or submarines, are based on laser
or fiber-optic gyroscopes and have very low drift - their
position error builds up slowly, on the order of a meter
every few minutes or an hour. Low-cost INS based on
MEMS technologies (micro-mechanical sensors used,
for example, in smartphones or drones) unfortunately
have much worse stability - errors in acceleration
measurements mean that after just a few tens of
seconds without correction, the position can have an
error of several meters, and after a few minutes even
hundreds of meters. In general, a fundamental
limitation of any INS is that its accuracy decreases over
time due to the accumulation of measurement errors
during integration [2]. Even the most accurate inertial
systems suffer from a certain level of bias (drift) of
accelerometers and gyroscopes, which, if not
periodically corrected, leads to an increasing deviation
of the calculated position from the actual position. In
the literature, this phenomenon is referred to as
unbounded growth of error - without external
correction, the INS will sooner or later lose its way” to
the point where its indications become navigationally
useless [2].
Despite this disadvantage, autonomy and
immunity to external interference make INS serve as a
key backup navigation source. For example, in modern
passenger aircraft, several inertial navigation
platforms (IRS) are installed to determine the
machine's movement parameters throughout the
flight. When a GPS signal is available, the data from the
INS is calibrated and corrected on an ongoing basis, but
if it is lost, it is the INS that assumes the burden of
maintaining knowledge of the flight's position and
direction for the time needed, for example, to return
safely to the airport [9]. Similarly, in maritime
navigation - modern ships are equipped with
AHRS/INS systems that can navigate without GPS for
a period of time, relying solely on gyroscopes and
speed logs. In autonomous vehicles, INS (sometimes in
a reduced form, the so-called RISS - Reduced Inertial
Sensor System, e.g., with the replacement of some of
the sensors with a wheel odometer) assists in
maintaining traffic awareness in dense urban areas
where the GNSS signal is weak or fades among tall
buildings [9].
592
However, practice shows that GNSS and INS are
complementary - they have complementary properties
[2]. GNSS provides absolute position reference with a
current time-independent error (e.g., a few meters
horizontally), but is susceptible to interference and
fading. INS, on the other hand, is continuous and
difficult to disturb, but its relative error increases over
time. For this reason, the integrated use of GNSS and
INS has become standard, so as to combine the
advantages of both. In a typical navigation system,
GNSS provides periodic accurate position and velocity
corrections, resetting INS's growing error, while INS
provides a bridge during periods when current GNSS
data is unavailable [2]. This approach creates a
navigation system that is both accurate in the long run
(thanks to GNSS) and resilient to short-term reception
problems (thanks to INS).
The disadvantage of integration based solely on
occasional “resetting” of the INS with GPS data (the so-
called loosely coupled system) is that when the GNSS
signal interruption lasts too long or the number of
available satellites falls below the minimum (e.g., less
than 4, making it impossible to determine the position),
the system switches to a standalone INS anyway and
the error starts to grow rapidly [2]. This is why tightly
integrated GNSS/INS systems are used in practice,
where the two systems work together continuously -
the INS continuously corrects itself even with partial
GNSS data, and in return can help track satellite signals
during difficult conditions (e.g., in tightly coupled
systems, the INS can help sustain satellite tracking
when the signal is on the verge of fading). Regardless
of the detailed integration architecture, however, the
role of the INS is crucial as a navigation safeguard: it is
through it that the system is able to sustain position
determination when GNSS signals are disrupted or
lost. Sections 2.3 and 2.4 discuss how artificial
intelligence can further enhance this integration,
reducing INS errors and effectively extending the time
for safe navigation without GNSS.
2.3 Position correction mechanisms using AI and filtering
For years, the Kalman filter has been the primary
mathematical tool used to combine GNSS and INS data
into coherent navigation information. A Kalman filter
is a state estimation algorithm that, based on a motion
model and a measurement model, can optimally
estimate the current error and correction based on the
available data, even if it is subject to noise [6]. In the
context of GNSS/INS integrated navigation, the
Kalman filter usually acts as an INS error estimator - it
compares the position and velocity calculated by the
INS with GNSS measurements and calculates
corrections to the inertial solution based on this [1, 2].
In other words, when a GNSS signal is available, the
filter continuously corrects INS drift, minimizing the
differences between the two sources. At times when
the GNSS signal dies, the filter enters predictive mode:
it continues to estimate the state based only on the INS
model, and its uncertainty gradually increases, but
thanks to the earlier calibration of INS errors, it slows
the accumulation of these errors compared to the
stand-alone INS [15].
A key challenge is maintaining accuracy during
longer GNSS signal gaps. The standard Kalman filter
assumes a specific INS error model (e.g., gyro drift
modeled as a first-order Gauss-Markov process, etc.)
and uses limited information for prediction, so that at
long gaps (tens of seconds or more) its predictions
begin to diverge from reality. In recent years, however,
ideas have emerged to support classical filtering with
artificial intelligence methods that can learn more
complex error and motion patterns than the simple
linear model used in the Kalman filter. AI can be
involved in several ways: to predict corrections during
periods of GNSS absence, to adaptively tune filter
parameters, or as a separate sensory fusion module
parallel to the classical approach.
One interesting strategy is to predict GNSS
observations from INS data using machine learning
models. Such a mechanism has been proposed, among
others, by Chen et al. (2022), where they used a hybrid
method based on partial least squares regression
(PLSR) and regression with Gaussian process (GPR)
[5]. The model learns the relationship between INS
errors and GNSS observations (when these are
available) to predict, during satellite fading, what the
GNSS data (e.g., pseudo-distance or position) would
look like based on INS alone. These “synthetic”
observations are then fed into the Kalman filter instead
of real GNSS data, allowing it to continue to correct INS
errors despite the absence of an actual satellite signal.
In experiments with four GNSS fades during a test
flight, it was shown that the use of such prediction
significantly improved navigation accuracy compared
to the usual mode without corrections - the Kalman
filter “supported” by PLSR/GPR prediction
maintained smaller trajectory deviations than the
traditional one, proving the effectiveness of the
approach. In other words, the intelligent module
learned to predict certain GNSS receiver behavior
based on traffic history and INS errors, thus filling in
data gaps during disturbances.
Another line of research focuses on directly
correcting INS readings using AI algorithms. An
example is the use of adaptive neuro-fuzzy systems
(ANFIS) to calibrate lower-cost inertial systems. In the
work of Mahdi et al. (2024) presented a system in which
a reduced RISS inertial system (using, among other
things, an odometer and one gyroscope) was
integrated with GNSS using ANFIS as an error
correction learning block [9]. ANFIS, combining
elements of neural network and fuzzy logic, was
trained to predict INS-generated position errors based
on vehicle movement patterns and previous
differences between INS and GNSS. As a result, the INS
so calibrated provides much more accurate
information during interruptions in GPS reception.
Tests using actual vehicle runs, during which GNSS
signal fades lasting between 50 and 150 seconds were
simulated, showed a reduction in 2D position error
(RMSE) of about 43.8% relative to traditional
RISS/GNSS integration without such calibration [9].
Maximum position deviations were also reduced by
about 47%, meaning that the vehicle, even for two
minutes without GNSS, stayed much closer to the
actual trajectory than would have been possible with
the classical system [9]. This is a big step forward, given
that we're talking about a system made of low-cost
MEMS sensors - thanks to AI, precision close to that
which would be provided by much more expensive
inertial systems was achieved.
593
Also, deep neural networks are finding applications
in improving inertial navigation in the absence of
GNSS. Long Short-Term Memory (LSTM) models, a
variation of recurrent neural networks capable of
learning temporal dependencies, have been used to
predict vehicle movements and trajectories based on
history from inertial sensors. Research indicates that
LSTM networks can effectively stabilize INS drift error:
in field tests, it was shown that the use of such a model
can significantly improve positioning accuracy during
GNSS fading (e.g., 30 or 90 seconds) compared to
unassisted stand-alone INS. A network learned from
data from periods when GNSS is operational can then
predict on-the-fly corrections to the INS-calculated
position when GNSS fades. In practice, this can look
like this: an autonomous vehicle equipped with an
LSTM network, when the GPS signal is lost, switches
to a mode where each new position from the INS is
corrected by a neural network predicting how much
error is likely to have occurred. It's a bit like a virtual
GPS acting on past experience. Such an approach,
while sounding futuristic, is already yielding tangible
results, and further development of the models (e.g.,
combining LSTM with convolutional networks that
analyze camera data - which combines inertial and
visual navigation) could further improve the reliability
of GNSS-free navigation.
Artificial intelligence can also support position
correction more indirectly - for example, through
adaptive tuning of the Kalman filter. Filter parameters
(covariances of model and measurement noise) are
often fixed rigidly or according to simple adaptations,
which do not always reflect the actual error dynamics.
Methods of using learning algorithms to continuously
adapt these parameters on the fly are being
investigated, so that the Kalman filter can better
respond to changes in conditions (e.g., rapid vehicle
maneuvers that increase INS model error). For
example, a neural network could learn from observed
deviations between INS and GNSS to adjust the filter's
Q and R covariance matrices, making it more resilient
to unusual situations. Such approaches are at the
research stage, but early results suggest that combining
a physics-based model (Kalman filter) with a learning
model yields better results than each alone. In other
words, AI does not have to replace proven navigation
algorithms - it can complement them, compensating for
their limitations (e.g. nonlinearities, model
uncertainty).
In summary, GNSS interference position correction
mechanisms are increasingly using a hybrid approach:
classical filters and models are combined with AI
components that provide additional information or
adaptivity. Experimental results prove that this makes
it possible to significantly improve the accuracy and
reliability of navigation during prolonged
interruptions in satellite signal reception. As a result,
the navigation system becomes more resilient - it can
rely on itself for longer without external data,
increasing overall safety.
2.4 Overview of actual research, implementation and
experiments
The topic of enhancing the safety of navigation using
AI under GNSS interference is being intensively
researched around the world. A large number of
experimental works, prototype implementations and
simulation studies have appeared in the literature of
the last decade, which confirm the usefulness of the
methods described. The following is an overview of
selected developments and initiatives in this area,
divided into the issues of anomaly detection and
maintaining navigation during signal fading.
GNSS interference detection: As early as around
2015-2020, work began to appear demonstrating the
effectiveness of machine learning in spoofing
detection. The authors in [12] presented a system that
used supervised machine learning algorithms
(including SVM) to detect false GNSS signals. A unique
element was the use of real incident data - the authors
used sets of meaconing and spoofing records from real
tests, which made it possible to test the algorithm
under near-real conditions. The results showed that the
classifier learned from laboratory data recognizes
attacks in field data with high efficiency, which is
significant evidence of the practical usefulness of AI.
Other research teams have also confirmed the
effectiveness of the ML/DL approach: in 2021 and 2022,
research results were published in which various
models (from decision trees, to convolutional neural
networks analyzing the signal, to hybrid methods
based on ensemble learning) achieved efficiencies of
95-99% in detecting GNSS interference on shared
datasets [8]. The ~99% accuracy result mentioned
earlier was achieved on two different jamming and
spoofing datasets, demonstrating the versatility of the
approach used [8]. It is interesting to note that the best
results were obtained by combining different
techniques - the aforementioned work used both deep
learning methods and elements of signal analysis and
even vision techniques (e.g., a camera analyzing the
environment for sources of interference). This indicates
a trend that multi-sensor fusion involving AI can yield
the most reliable sensing systems - for example,
combining data from a GNSS receiver, inertial sensors,
cameras and other sources into a single platform
supervised by learning algorithms. Such solutions are
currently being tested in aeronautical or maritime
applications: next-generation prototype GNSS
receivers can collect dozens of diagnostic parameters,
which are then analyzed by an embedded AI module
and signal the crew or autopilot in real time when a
potential anomaly is detected. For example, research
projects co-funded by European Union agencies are
testing systems where the aircraft is equipped with a
dual GNSS set (one primary, the other monitoring) and
INS - data from both receivers and INS is compared by
an intelligent algorithm that can detect even subtle
attacks (e.g., those that slowly move position, hoping
not to be noticed by standard alarms).
Importantly, the aviation and transportation
industries recognize the seriousness of the problem.
Organizations such as EASA (European Union
Aviation Safety Agency) and IATA are sounding the
alarm about the increasing number of GNSS jamming
incidents and are launching initiatives to make systems
more resilient. Post-conference recommendations
point to the need to equip aircraft with improved GNSS
jamming detection and warning mechanisms, as well
as to maintain traditional navigation aids as back-ups
(e.g., aviation VOR/DME beacons) [7].
As for maintaining navigation during GNSS fading,
numerous experiments have been reported in this area
594
as well, confirming the effectiveness of AI methods.
Chen, L., et. all [5] conducted flight tests in which the
aircraft experienced four targeted GNSS signal outages
at different phases of flight. The navigation accuracy of
the GNSS/INS system with a classical Kalman filter and
the system assisted by the AI prediction module
(PLSR/GPR) were compared. The results clearly
showed that the AI-assisted version maintained the
correct trajectory much better - the average position
error during GNSS fading was significantly smaller
than in the traditional system. What's more, the AI
system showed greater stability under different
maneuvers (the flight included both calm stretches and
sharp turns); the classical filter gave a larger deviation
at certain points, while the system with AI prediction
still kept close to the true path. This is important
evidence that intelligent methods improve not only
static accuracy, but also robustness to different
dynamic conditions.
In the field of land navigation, especially for
autonomous vehicles, a number of tests have been
conducted in urban conditions. For example, the
authors [9], in their experiment, drove a vehicle along
urban routes, where GNSS signals were deliberately
turned off for set periods of time (from tens to over a
hundred seconds). The car then relied only on its own
sensors (IMU, odometer) - once with and once without
ANFIS correction enabled. The comparison showed
that without AI the vehicle “lost position” from the real
one quite quickly (the error grew exponentially with
time), while with ANFIS the trajectory reconstructed
by the system remained close to the real one - the
differences in distances of tens of meters were able to
be reduced to a few tens of meters, which can be
decisive for maintaining continuity in the lane or
avoiding a collision at a critical moment [9]. It is worth
mentioning that these tests took into account different
scenarios: straight driving, sharp turns (where INS
errors usually increase faster), and speed changes. In
all cases, neuro-fusion integration improved the
results, making it a promising technique for the
automotive industry. Already, some so-called dead
reckoning GPS systems in cars use simple forms of
machine learning for self-calibration (e.g., learning the
bias of a MEMS gyroscope when the car is at a traffic
light). Future on-board navigation systems can be
expected to incorporate even more advanced learning
algorithms, especially in autonomous vehicles, where
full redundancy (GNSS, INS, lidar, camera) is standard
- AI will play the role of integrator of all these sources.
For ships, systems are being developed that
integrate INS, GNSS and local sensors like logs and
gyrocompasses, where AI manages how much weight
to assign to each source at any given time. Under
conditions of GNSS interference, the system can
automatically switch to INS + optical” mode, using,
for example, cameras and lidar to track the shoreline
(known as terrestrial navigation), and machine
learning helps compare this information with the map.
These are, however, solutions just crawling in research.
Military projects are more advanced - armies have long
used high accuracy INS, and are now testing AI to
correct these INS without GNSS, such as by matching
observed maneuver runs to known models. The details
are often not public, but it is conceivable that military
vehicle navigation systems will become a testbed for
many of the techniques described (AI for spoofing
detection, fusion with magnetometers, IR cameras,
etc.).
Despite numerous successes, operational
implementations of these technologies are just
beginning. So far, most of the AI systems described
have been demonstrated in test or prototype
conditions. However, there are early swallows: for
example, Honeywell has announced work on a GNSS
receiver integrated with an AI module capable of self-
learning the interference environment for general
aviation. In automotive, manufacturers of GPS systems
for the automotive class are beginning to add learning-
based features (like the aforementioned INS self-
calibration). Commercial software platforms are also
emerging that allow the implementation of custom AI
algorithms on GNSS receiver data (e.g., receivers with
access to raw observations and the ability to upload
user software) - this is important because it will allow
the user community to create dedicated anti-spoofing
solutions for their own needs.
In summary, research and experiments conducted
over the past few years clearly confirm the
effectiveness of using AI both in detecting GNSS
threats and in maintaining navigation when they
occur. Impressive results have been achieved in
laboratory conditions (close to 100% attack detection,
significant increase in safe navigation time without
GNSS). In real-world conditions, the prototypes are
also performing very well, although they are, of course,
still subject to intensive testing in various scenarios.
The next step is to translate these achievements into
industrial standards and procedures, which is already
beginning to happen with the cooperation of scientific
institutions, manufacturers and regulators.
3 MATHEMATICAL MODEL OF GNSS/INS/AI
SYSTEM
GNSS systems are crucial for precise navigation, but
their signals are susceptible to interference (e.g.,
spoofing, jamming) and atmospheric disturbances. To
increase the reliability of position estimation,
additional inertial systems (INS) and adaptive filtering
methods are used. Integration of these systems while
monitoring anomalies in the GNSS signal enables
dynamic error correction. This analysis presents
mathematical modeling of the state estimation process,
an adaptive anomaly detection algorithm and detailed
simulation results.
3.1 The extended state model
Assume that the state of the system is described by a
vector:



=




k
k
a
k
g
k
p
v
b
b
k
x
where:
pk position,
vk speed,
595
bk
a
drift (bias) of the accelerometer,
bk
g
gyro drift.
For a discrete dynamic system, the model can be
written as:
whereby:
Fk state transition matrix, taking into account the
dynamics of movement and evolution of drifts,
Bkuk Control model, where uk to controlled
acceleration (e.g., correction from INS),
wk~N(0,Qk) process noise.
An example of the form of the Fk matrix (for a 1D
model with Δt sampling) might look as follows:
2
1 0
2
0 1 0
0 1 1 0
0 0 0 1

−



−
=






t
t
t
k
F
where the negative members associated with drifts
model the systematic effect of INS errors on the
estimation [14].
3.2 Measurement model
The GNSS signal provides the position measurement
directly. We write the measurement model as:
=+
kk
xv
k
zH
where:
H=[1 0 0 0] observation matrix,
vk~N(0,Rk) measurement noise, whose covariance Rk
may undergo adaptive modifications under
anomalous conditionsi.
3.3 Anomaly detection using an AI system
3.3.1 Innovation and statistical test
The basis of anomaly detection in the Kalman filter
is innovation vector analysis:
|1
ˆ
=−
k k k k
v z xH
where:
|1
ˆ
kk
x
is the predicted state. For the standard
Kalman filter, the innovation covariance matrix is:
|1
=+
T
k k k k
SPH H R
Anomaly detection test is based on statistics:
21
=
T
k k k k
d v S v
which under normal conditions has a chi-square
distribution with mmm degrees of freedom (where m
is the measurement dimension). Threshold γ is selected
to take care of a certain level of significance of the test
(e.g. α=0.05):
2
anomaly detected
k
d
If an anomaly is detected, the system can modify the
matrix Rk or initiate a filter reset procedure.
3.3.2 Adaptive AI-based mechanisms
Machine learning algorithms (e.g., deep neural
networks, SVMs) can be trained on historical data to
classify signals as normal or disturbed. In a
hybrid approach, the result of the classification
(denoted, for example, as k {0,1}) affects the filter
parameters:
if δk=1 (anomaly), element of the matrix increases
Rk, which reduces the weight of measurement
GNSS,
additionally, it is possible to use adaptive
algorithms (e.g., drift recalibration) to minimize
system errors.
The adaptation model can be formalized by:

=+
adapt
kk
k
R R I
where:
λ is an adaptive parameter,
I unit matrix [10].
3.4 Implementation and simulation
The following simulation parameters were assumed:
Sampling time: Δt=0.1 s,
Number of steps: 1000 iterations (simulation lasting
100 s),
Noise parameters:
The noise of the process: Qk = diag(qp, qv, qb
a
, qb
g
)
with typical values e.g.. qp=10
4
, qv = 10
-4
, qb
a
=10
6
, qb
g
=10
6
Measurement noise GNSS: Rk=
2
GNSSzGNSS1
Interference conditions: Random interference was
introduced in the simulation to simulate spoofing -
in random iterations, a large deviation was added
to the zk.
3.5 Estimation algorithm
A classical Kalman filter was used for GNSS/INS data
fusion. Stages of the filter:
1. Prediction:
| 1 1| 1 1
ˆˆ
=+
k k k
F B u
k k k k
xx
| 1 1| 1
=+
T
k k k k k k k
P F P F Q
2. Update (with adaptation):
( )
( )
( )
1
| 1 | 1
| | 1 | 1
| | 1
ˆˆ
ˆ
−−
−−
=+
= +
=−
adapt
TT
k k k k k
k
k k k k
k k k k k
zx
k k k k
K P H HP H R
x x K H
P I K H P
The anomaly detection module monitors dk
2
i and, if
the threshold is exceeded, modifies Rk according to the
adaptive model presented earlier.
596
3.6 Simulation results
Simulation conducted using the Monte Carlo method
(500 runs) made it possible to determine the statistical
properties of the estimation. The results are shown in
Table 1:
Table 1. Simulation results
Scenario
Mean position
error [m]
Standard
deviation [m]
GNSS without correction
3.5
1.2
GNSS/INS + classical Kalman filter
1.2
0.5
GNSS/INS + Kalman filter with AI
anomaly detection
0.7
0.3
In addition, analysis of the covariance matrix Pk|k
showed a significant reduction in estimation
uncertainty in the hybrid system. Graphical
presentation of the estimated trajectory as a function of
time showed that at moments of disturbance (detected
by the AI module) there is immediate adaptation of the
system, which minimizes the impact of anomalies on
the estimation.
3.7 Discussion and interpretation of results
INS and GNSS data integration: Data fusion makes
the system resilient to short-term GNSS signal
interference - INS provides continuity of motion
dynamics information.
Adaptive Rk matrix correction: The inclusion of an
anomaly detection module allows the dynamic
increase of measurement uncertainty when a
disturbance is detected, resulting in less impact of
erroneous data on the estimation.
Effectiveness of AI algorithms: Trained models
detect abnormal patterns in the innovation vector,
enabling the system to respond quickly. The use of a
hybrid approach reduces the average position error by
more than 40% compared to a classic filter.
4 CONCLUSIONS
The mathematical analysis and simulations conducted
indicate that:
GNSS/INS integration using an extended state
model and an adaptive Kalman filter significantly
improves navigation precision even under signal
interference conditions.
Adaptive anomaly detection mechanisms based on
innovation vector analysis and complemented by
artificial intelligence algorithms enable dynamic
correction of filter parameters.
Simulation results confirm that the hybrid approach
reduces position error to levels on the order of
0.7 m, which is important for critical applications
(e.g., autonomous systems, aviation, shipping, rail
transportation).
The analyses and examples presented here clearly
indicate that artificial intelligence has great potential
for enhancing navigation safety under unreliable
GNSS systems. AI methods, particularly machine and
deep learning, have proven capable of detecting
satellite interference with a speed and precision
unattainable by traditional algorithms [8]. At the same
time, intelligent data integration techniques allow the
system to maintain correct navigation for much longer
in the absence of a signal - thanks to AI, the system can
adapt to changing conditions and quickly correct
errors before they grow to dangerous proportions. In
other words, AI makes the navigation system more
aware: it can recognize on its own that “something is
wrong” (e.g., that the data being used may be false or
faulty) and react accordingly, whether by alerting the
operator or switching to an alternative mode.
The effectiveness of existing AI solutions has been
confirmed in numerous studies. In the context of GNSS
anomaly detection, the attack classification accuracies
achieved often exceed 95%, and in the best cases
approach 99%. This means that these systems very
rarely mistake normal conditions for an attack (low
false alarm rate) and at the same time almost always
capture the actual threat (high sensitivity). This is
important because in safety-critical applications, while
a false alarm is undesirable, failing to notice an actual
attack is absolutely unacceptable. AI seems to be able
to meet these requirements better than classical
methods - by analyzing multiple sources and features
at once, it is less likely to be fooled by interference
masquerading as a normal signal. In the field of
navigation maintenance (fusion of GNSS/INS and
other sensors), on the other hand, AI has proven that it
can drastically reduce positioning error during long
gaps. Reductions of 30-50% in error achieved in
automotive and aviation tests translate into real
improvements in safety - for example, a ship will stay
in a safe navigation corridor, an aircraft will stay closer
to its intended path, and an autonomous vehicle will
not lose its lane. All this brings us closer to the goal of
resilient navigation (resilient PNT).
It should be noted, however, that despite promising
results, large-scale implementation of this technology
comes with challenges. First, AI methods must be
properly validated and certified, especially in fields
such as aviation, where safety requirements are very
stringent. It is necessary to prove that the learning
algorithm works reliably under all predicted
conditions, and to understand its behavior in
unforeseen situations. This is not always easy, because
AI models (especially deep neural networks) are
sometimes treated as “black boxes.” That's why one of
the directions of development is so-called XAI
(Explainable AI) - methods to look inside the model
and explain why it made such a decision and not
another. In the context of navigation, this can increase
trust in the system and facilitate certification. Second,
current solutions often require large training data sets
representing a variety of disturbance scenarios.
BIOBLIOGRAPHY
[1] Bonar B., Buratowski T.: Extended Kalman filter
supported by logical conditions for inertial navigation.
W: Systemy mechatroniczne : zagadnienia wybrane :
praca zbiorowa / pod red. Michała Mańka. Kraków :
Katedra Robotyki i Mechatroniki. Akademia Górniczo-
Hutnicza w Krakowie, 2022. ISBN: 978-83-965598-2-1.
S. 5159.
[2] Boguspayev, N., Akhmedov, D., Raskaliyev, A., Kim, A.,
Sukhenko, A. (2023): A Comprehensive Review of
GNSS/INS Integration Techniques for Land and Air
Vehicle Applications. Applied Sciences, 13(8), 4819.
https://doi.org/10.3390/app13084819
597
[3] Britannia P&I Club - Global Marine P&I Insurance.
https://britanniapandi.com/wp-
content/uploads/2024/10/Navigational-Risks-at-Sea-
Jamming-and-Spoofing.pdf [C: 16 October 2024]
[4] Clausnitzer, Ch., Alexander K,.(2024): GNSS
Jamming/Spoofing and Continued Operational Safety
International Committee on GNSS Interference Detection
&Mitigation Workshop 15 April 2024
[5] Chen, L., Liu, Z., Fang, J.(2022): A novel hybrid
observation prediction methodology for bridging GNSS
outages in INS/GNSS systems. Journal of Navigation.
2022;75(5):1206-1225. doi:10.1017/S037346332200025X
[6] Chrzan, M., Jackowski, S. (2016): Współczesne systemy
satelitarne w transporcie kolejowym. Politechnika
Radomska im. Kazimierza Pułaskiego w Radomiu
[7] EASA partners with IATA to counter aviation safety
threat from GNSS spoofing and jamming | EASA. (b. d.).
EASA. https://www.easa.europa.eu/en/newsroom-and-
events/press-releases/easa-partners-iata-counter-
aviation-safety-threat-gnss-
spoofing#:~:text=“GNSS%20systems%20offer%20tremen
dous%20advantages,is%20a%20priority%20for%20the]
[C: 10.03.2025]
[8] Ghanbarzade, A., & Soleimani, H. (2025): GNSS/GPS
Spoofing and Jamming Identification Using Machine
Learning and Deep Learning. arXiv preprint
arXiv:2501.02352.
[9] Mahdi AE, Azouz A, Noureldin A, Abosekeen A.(2024):
A Novel Machine Learning-Based ANFIS Calibrated
RISS/GNSS Integration for Improved Navigation in
Urban Environments. Sensors (Basel). 2024 Mar
20;24(6):1985. doi: 10.3390/s24061985.
[10] Pardhasaradhi, B., Yakkati, R. R. Cenkeramaddi, L.
R.(2022): Machine Learning-Based Screening and
Measurement to Measurement Association for
Navigation in GNSS Spoofing Environment in IEEE
Sensors Journal, vol. 22, no. 23, pp. 23423-23435, 1 Dec.1,
2022, doi: 10.1109/JSEN.2022.3214349
[11] Radoš, K., Brkić, M., & Begušić, D. (2024): Recent
Advances on Jamming and Spoofing Detection in GNSS.
Sensors, 24(13), 4210. https://doi.org/10.3390/s24134210
[12] Semanjski, S., Semanjski, I., De Wilde, W., & Muls, A.
(2020): Use of Supervised Machine Learning for GNSS
Signal Spoofing Detection with Validation on Real-World
Meaconing and Spoofing DataPart I. Sensors, 20(4),
1171. https://doi.org/10.3390/s20041171
[13] Simon, D., "Optimal State Estimation: Kalman, H
Infinity, and Nonlinear Approaches", John Wiley & Sons
2006.
[14] Titterton, D. H. i Weston, J. L., (2004): Strapdown Inertial
Navigation Technology, Radar , Sonar and Navigation
Series 17. 2004.
[15] Zhang, Y., A (2019): Fusion Methodology to Bridge GPS
Outages for INS/GPS Integrated Navigation System. IEEE
Access. 2019;7:6129661306. doi:
10.1109/ACCESS.2019.2911025