IEEE - Aerospace and Electronic Systems - July 2021 - 10

Machine Learning Support for Radar-Based Surveillance Systems
Table 1.
Data Formats
Field
ADS-B
AIS
Identifier ICAO Code
Position GPS
Velocity GPS
Timestamp GPS
Class./Ident. Aircraft Type
Class./Ident. Military Flag
Class./Ident. Ground Flag
MMSI/IMO
GPS
GPS
GPS
Ship Type
Width/Length
Class./Ident. Airline/Operator Cargo
Track ID
Radar
Tracking
GPS
DopplerClassification
HRR Classification
Navigation Status IFFIdentification
Environment Classification
Class./Ident. Source/Destination Source/Destination BehaviorAnalysis
Italicfields are optional.
ofit. Additionally, data can be accessed via databases, which
are optimized for the distributed handling ofbig data. In the
world of No-SQL databases, there are manifold solutions
that all have focused on different storage mechanism and
data types, e.g., wide-column storage, graph databases, or
document databases. Examples are Apache Cassandra [26],
SAP OrientDB [27], and Apache MongoDB [28], which
were also used here.
CLEANINGANDSMOOTHING
Data streams from ADS-B and AIS receivers often contain
impurities and spikes, which are caused by receiver
inaccuracies or time misalignments of the receivers,
e.g., ADS-B Exchange provides " short trail " messages
next to default messages, which contain incomplete
information [8].
Also time delays are very often observable so that
tracks jump back, which leads to inconsistent track spikes
[29]. And since not every field in both protocols is mandatory,
information can be missing in messages. Finally data
often contains duplicates of messages. These issues complicate
the extraction of a representative learning input
and demand a preprocessing of the data, which even might
be the biggest challenge.
For cleaning, spikes were detected and removed and
smoothed with a Kalman filter afterward [30]. Figure 1
visualizes the different steps for one single track.
SLIDINGWINDOWS
Some machine learning methods are based on features that
consist of several parameters of one update, but others of
several track updates. Since the data contains very long
trajectories of individual objects, an effective processing
is required to cut out these segments of trajectories-the
so-called tracklets. Spark offers these possibilities by
extracting these data with a sliding window. The left side
of Figure 2 is intended to explain the principle schematically
especially for trajectories as it was used for the following
examples. Segments with a fixed length are copied
from the trajectory and saved separately. Since each trajectory
can be uniquely assigned to an object, processing takes
place in parallel. In order to work with representative data,
it must be standardized in time, i.e., the time interval
Radar
Figure 1.
Example for an ADS-B track. From left to right: Raw, cleaned,
and smoothed.
10
Figure 2.
Principle of the sliding window applied to trajectories and time
normalization especially for ADS-B Exchange data
IEEE A&E SYSTEMS MAGAZINE
JULY 2021
IEEE - Aerospace and Electronic Systems - July 2021

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - July 2021
