IEEE - Aerospace and Electronic Systems - June 2022 - 7

Credit: Image licensed by Ingram Publishing
instance. This model does not consider the relationships
among the different attributes, but it only analyzes the temporal
sequence of the data. Besides, using multiple LSTM networks
for analyzing multivariate and correlated data would
result in toomuch redundancy. To solve the redundancy issue,
we propose a different approach. The second approach
depends mainly on multioutput ConvLSTM network. This
approach tries to detect abnormalities in the values of a concerning
attribute by considering the effect of all the other
attributes. To use this approach, we propose that the incoming
input is iteratively collected and reshaped as a sequence of3D
tensors. The elements ofa 3-D tensor are composed of two
channeled vectors. The two channels contain the values
of a sliding window for two attributes. This way, each
row of the input will be reshaped as if it was a multidimensional
raster image. This architecture enables the
convolutional LSTM to analyze the temporal relationships
among the attributes. A separation layer is added
to this model to separate the results among the concerning
attributes, and each part of the output will be
labeled as (Normal/Abnormal) instance.
To conduct experiments, we used a real-life dataset,
which consisted of four flights for a fixed-wing aircraft.
The two approaches detected different types of faults,
such as (sensor-impulse, sensor-drift, and sensor-cut).
However, the multioutput ConvLSTM was faster and presented
superior results in most cases.
The rest of the article is arranged as follows. The
" Literature Review " provides a briefreview ofrelated works.
" Methodology " provides some definitions and explains the
background of the proposed MLSTM and multioutput
ConvLSTM algorithms. " Experimental Results " presents the
used dataset and the results of experiments. Finally, we
present our conclusions and future suggestions.
LITERATURE REVIEW
Many algorithms were designed to detect potential faults
by extracting anomalies from telemetry data. These algorithms
are categorized into three types: Knowledge-based,
JUNE 2022
model-based, and data-driven-based algorithms [11]. The
knowledge-based algorithms depend on predefined rules
(if-then sentences) and can be used in case of the availability
of comprehensive and accurate knowledge of the
domain. The model-based algorithms use an underlying
model, which is extracted analytically from mathematical
equations that describe the functionality of each part of
the system. However, building the rules for the knowledge-based
algorithms and extracting the mathematical
model for the model-based algorithms is difficult when
the attributes are many. Bu et al. [2] and Sugumaran and
Ramachandran [23] used decision trees to build rules for a
fault-diagnosis fuzzy classifier; however, their model
faced a limitation in performance due to the permanent
rules in the fuzzy logic model because it was not able to
detect unknown faults. Casas et al. [4] used decision trees
to detect anomalies in cellular network data, where they
used semisynthetic data. Their approach correctly recognized
more than 80% of the Abnormal instances with no
false positives. Cork and Walker [6] used a model-based
fault detection algorithm, where they estimated the state
of the UAV using a nonlinear dynamic model of the aircraft.
They used the divergence of the estimated state
from its actual value to detect system faults.
The data-driven algorithms depend on extracting statistical
knowledge to identify Abnormal values. The
advantage of the data-driven algorithms is that it does not
require experts nor domain knowledge, as the model automatically
learns and extracts knowledge and patterns from
data. Lin et al. [13] designed an online algorithm to detect
UAV sensor faults based on a statistical analysis of sensor
readings and navigation data. Khalastchi et al. [9] and
Pokrajac et al. [18] used statistical methods to produce an
anomaly score for each given point of the flight, where
each anomaly score depends on a sliding window of monitored
readings. Paffenroth et al. [17] developed a PCAbased
anomaly detector to predict cybernetwork attacks.
Camacho et al. [3], [31] used kernel PCA to detect anomalies
in large datasets and complicated correlations of UAV
sensor data. Their proposed method detected the injected
pulse and constant drift anomalies effectively. Lee and
IEEE A&E SYSTEMS MAGAZINE
7
IEEE - Aerospace and Electronic Systems - June 2022

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