IEEE - Aerospace and Electronic Systems - June 2022 - 6

Feature Article::
DOI. No. 10.1109/MAES.2021.3053108
Using MLSTM and Multioutput Convolutional LSTM
Algorithms for Detecting Anomalous Patterns in
Streamed Data of Unmanned Aerial Vehicles
Ahmad Alos and Zouhair Dahrouj, Higher Institute for Applied Sciences and
Technology, Damascus, Syria
INTRODUCTION
During flights, the unmanned aerial vehicle (UAV) sends
back large amounts of telemetry data to a ground station.
The telemetry data are stored in rows, where each data
row contains the values of the UAV attributes. These
attributes are either commands (altitude command, the
rudder command, and the aileron command) or sensor
readings (airspeed, pitch, roll, and yaw). This incoming
data are heterogonous and grows in size very rapidly.
Human monitoring of the collected data is not sufficient to
identify anomalous cases instantly, which could indicate a
potential fault of the system. A possible fault might result
in a system failure, where the chances of catastrophic situations
are high. Hence, it is a matter of paramount necessity
to design efficient algorithms to extract abnormal
instances from the data flow of the concerned sensors and
label these extracted instances as potential faults [2]. To
foresee system failure, we use anomaly detection algorithms.
Anomaly detection algorithms find patterns in data
that do not follow an expected behavior [5]. Usually,
anomaly detection algorithms operate in one of three
modes [5], [19]: Supervised, semisupervised, and unsupervised.
The supervised anomaly detection algorithms
assume the availability of training data with given labels
for the Normal class as well as for the Anomalous class.
The semisupervised algorithms assume the training data
Authors' current addresses: Ahmad Alos, Zouhair Dahrouj,
Higher Institute for Applied Sciences and Technology,
Damascus, Syria (e-mails: ahmad.alos@hiast.edu.
sy; zouhair.dahrouj@hiast.edu.sy).
Manuscript received June 7, 2020, revised September 5,
2020; accepted January 5, 2021, and ready for
publication January 18, 2021.
Review handled by Tai-Hoon Kim.
0885-8985/21/$26.00 ß 2021 IEEE
6
have labeled instances for the Normal class only. The
unsupervised anomaly detectors do not involve any labeling
for the training data, where they make the implicit
assumption that the Normal instances are far more frequent
than the Abnormal ones.
In complex systems such as the UAV, the sensor faults are
categorized into three types: point, contextual, and collective
[5]. A point fault is observed as a single data instance that is
anomalous concerning the rest of the data [25]. A contextual
fault occurs when the sensor shows an invalid value concerning
the current context, such as a zero altitude value while the
UAV being airborne. A collective fault means that the sensor
shows a sequence of invalid values when put collectively
together. For example, the UAV is climbing down, but the
altimeter shows a fixed value [24].
The streamed data of the UAV are considered multivariate
temporal data that are received at a fixed rate of
time. Some explored algorithms consider the temporal
dependencies of the sequences [8]. However, the majority
of the existing algorithms do not contemplate the complex
relationships among the different UAV attributes, which
leaves a potential opportunity for new ideas.
Our contribution involves a comparative study of two
supervised approaches, which are state of the art, which use
existing deep learning tools: long short-term memory
(LSTM) and convolutional LSTM (ConvLSTM). We use
these two methods in a novel way to detect the potentially
defected sensors while analyzing the streamed data of the
UAV. The first approach involves using multiple long shortterm
memory (MLSTM) networks, where each LSTM network
is responsible for detecting anomalies in the continuous
values of one attribute. The input for MLSTM is reshaped
using a sliding window technique. The sliding window
reflects the effect of the previous values of each attribute.
Afterward, the data are separated and distributed to the multiple
LSTM networks so they can be analyzed, and then the
result of each LSTM is labeled as (Normal/Abnormal)
IEEE A&E SYSTEMS MAGAZINE
JUNE 2022
https://orcid.org/0000-0003-4549-780X
IEEE - Aerospace and Electronic Systems - June 2022

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - June 2022
