Recurrent Neural Network (GRU RNN), and Deep Gated Recurrent

Unit Recurrent Neural Networks (DGRU RNNs).

I. INTRODUCTION
Machine Learning Methods for
The tremendous number of spacecraft launched in the
Spacecraft Telemetry Mining past decades, enabled national dependence on space based
services; such as earth observation, communication, and
satellite navigation. The recent advances in technology and
corresponding spin off enabled satellite developers to use
state of the art electronics and software to either operate
SARA K. IBRAHIM the spacecraft or control its onboard equipment. Recent
Zagazig University, Zagazig, Egypt trend announced in the field of building satellites, is the
AYMAN AHMED dependency on low cost commercially available compo-
National Authority for Remote Sensing and Space Science Cairo, Cairo, nents. However, despite of using fault mitigation techniques
Egypt and additional shielding onboard spacecraft, may reduce
M. AMAL ELDIN ZEIDAN space environment effects on satellite components and sub-
IBRAHIM E. ZIEDAN systems, this harsh environment has many effects starting
Zagazig University, Zagazig, Egypt from thermal and vacuum conditions ending with radiation
effects. Spacecraft failure has many reasons, such as the
significant effects of radiation environment on one of the
critical components of satellite, for example the onboard
Spacecrafts are critical systems that have to survive space envi-
ronment effects. Due to its complexity, these types of systems are de-
computer, communication system, or power supply [1].
signed in a way to mitigate errors and maneuver the critical situations. Nowadays machine learning(ML)/Data Mining(DM)
Spacecraft delivers to the ground operator an abundance data related techniques are used widely in various fields such as space-
to system status telemetry; the telemetry parameters are monitored to craft operations support – the MARS express power chal-
indicate spacecraft performance. Recently, researchers proposed us- lenge [2] spacecraft ground systems [3], failure prognos-
ing Machine Learning (ML)/Telemetry Mining (TM) techniques for
telemetry parameters forecasting. Telemetry processing facilitates the
tic of avionics [4], and communication networks control
data visualization to enable operators understanding the behavior of which is an important aspect for both the service provider
the satellite in order to reduce failure risks. and end user. Data mining methods have been successfully
In this paper, we introduce a comparison between the different ma- used to address and optimized solutions to this issue where
chine learning techniques that can be applied for low earth orbit learning algorithms for data mining allow following and un-
satellite telemetry mining. The techniques are evaluated on the bases
of calculating the prediction accuracy using mean error and corre-
derstanding the network behavior so that control functions
lation estimation. We used telemetry data received from Egyptsat-1 and parameters can be updated during network operation
satellite including parameters such as battery temperature, power to achieve optimal performance in real condition [5]. For
bus voltage and load current. The research summarizes the perfor- the next-generation wireless networks, Machine learning is
mance of processing telemetry data using autoregressive integrated able to overcome the challenge of assisting the radio in in-
moving average (ARIMA), Multilayer Perceptron (MLP), Recur-
rent Neural Network (RNN), Long Short-Term Memory Recurrent
telligent adaptive learning and decision making, so that the
Neural Network (LSTM RNN), Deep Long Short-Term Memory Re- diverse requirements of next-generation wireless networks
current Neural Networks (DLSTM RNNs), Gated Recurrent Unit can be satisfied [6]. The Age of Digital Astronomy is such
an extremely data-rich environment beyond the capabili-
ties of traditional methodologies and approaches for ana-
Manuscript received May 17, 2017; revised April 21, 2018 and September lyzing and extracting new knowledge from the data. Way
25, 2018; released for publication October 3, 2018. Date of publication
et al. [7] have applied some state-of-the-art machine learn-
October 18, 2018; date of current version August 7, 2019.
ing and data mining techniques in astronomy; where the
DOI. No. 10.1109/TAES.2018.2876586 scientific discovery process is increasingly dependent on
Refereeing of this contribution was handled by T. Rossi. the ability to analyze massive amounts of complex data
generated by scientific instruments and simulations. JIAO
This work was supported by the National Authority for Remote Sensing et al. [8] presented a machine learning algorithm to detect
and Space Science-Egypt. automatic equatorial GPS amplitude ionospheric scintilla-
Authors’ addresses: S. K. Ibrahim, M. A. E. Ziedan, and I. E. Ziedan are tion and classify scintillation events based on training data
with the Department of Computer & Systems, Faculty of Engineering, in the frequency domain. ML/DM techniques are also ap-
Zagazig University, Zagazig 44519, Egypt, E-mail: (sara.khalil@zu. plied on real-time system traces [9], Cyclic Time Series
edu.eg; [email protected]; [email protected]);
classification [10], stock price forecasting [11], and fall de-
A. Ahmed is with the National Authority for Remote Sensing and
Space Science Cairo, Cairo 1564, Egypt, E-mail: (ayman.mahmoud@ tection ML approach for Range-Doppler Radars [12].
narss.sci.eg). (Corresponding author: Sara K. Ibrahim.) In this research, we investigate the ML/DM techniques
that can be utilized to analyze the performance of the space-
craft. We used Data mining to explore the performance pre-
0018-9251 
C 2018 IEEE
sented by telemetry parameter(s) that reflect the health of

1816 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 55, NO. 4 AUGUST 2019

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on March 30,2024 at 08:46:16 UTC from IEEE Xplore. Restrictions apply.
 certain onboard unit. This enables satellite operator to mon- dynamics, onboard subsystems status and mal-functioning,
itor the overall satellite health to reduce the risk of failure control program and commands execution, revealing of
with accurate and automated manner. One of the possi- malfunctions in the onboard subsystems operation, and
ble ways to monitor the satellite health is to use its online monitoring of satellite instruments and structural elements
telemetry to allow assessment of its status. The prediction temperature [14].
of telemetry parameters helps the operator to determine po- In our case, to guarantee the full confidence of inter-
tential/ upcoming satellite operating mode, which can help preting the input telemetry data set, and to compare the real
support decision making for urgent situations. This is an scenario of operation with the corresponding telemetry data
important issue, as an urgent situation may cause satellite set, authors used the telemetry data of EGYPTSAT-1, the
complete loss. The numerical nature of satellite teleme- first Egyptian remote sensing satellite, with support from
try parameter is usually formatted/presented as a time se- satellite operation and design team members [15].
ries due to nature of satellite operation. The time series
regression of the satellite telemetry parameters can point III. SATELLITE PERFORMANCE ANALYSIS USING MA-
to trend in telemetry parameter value change, which may CHINE LEARNING—LITERATURE REVIEW
cause satellite subsystem failure. Monitoring of such trends
Recently, researchers suggested using approaches to
will alarm for possible failure. One simple method is to
monitor and evaluate the performance of satellite subsys-
predict the next value(s) of one parameter and apply limit
tems; furthermore, advanced techniques may be used to
check, so that potential failure can be foreseen. When the
predict the performance of satellite devices and prepare for
predicted value probability exceeds the percentage of error
early stage decision-making. In this context, many algo-
probability defined by satellite operators/designer, it indi-
rithms have been developed to forecast the failure before it
cates that the related subsystem may go into a faulty behav-
happens based on telemetry data received from satellite.
ior and thus the satellite system will be affected; operator,
Yairi et al. [16] proposed a data-driven health mon-
then, should take precautions to avoid this situation.
itoring method based on probabilistic clustering and di-
This paper is organized as follow; we first give an in-
mensionality reduction for artificial satellites housekeeping
troduction about monitoring the performance of spacecraft.
data. Nassar and Hussein [17] presented a novel supervised
Then, we described the satellite subsystems. Section III
learning algorithm based on projection to latent structure
introduces a literature review about machine learning tech-
discriminant analysis technique (PLS-DA) applied to space-
niques used for diverse application; followed by Section
craft telemetry data in order to manage the nominal and off-
IV, we described the detailed algorithms to be evaluated.
nominal status of the spacecraft operations and overcome
Section V explains the format of the telemetry data received
faulty states in the space mission operation. Yang et al. [18]
from the EGYPTSAT-1 satellite and its associated correla-
proposed data mining methods for in-orbit satellite fault de-
tion theme. The evaluation methodology is then introduced
tection and prediction which is one of the key technologies
followed by the results of applying the selected algorithms
for health monitoring of in-orbit satellites.
on telemetry data. Finally, we conclude the research and
In this research, we carried out a comparative study
illustrate our future work.
between some state-of-the-art data mining techniques ap-
plied on the Egyptsat-1 telemetry. These algorithms are,
II. SATELLITE SUBSYSTEMS autoregressive integrated moving average (ARIMA), Multi-
layer Perceptron (MLP), Recurrent Neural Network (RNN),
Spacecraft has a set of subsystems, such as Attitude
Long Short-Term Memory Recurrent Neural Network
Determination and Control Subsystem (ADCS), Teleme-
(LSTM RNN), Deep Long Short-Term Memory Recurrent
try, Tracking, and Command (TT&C), Command and
Neural Networks (DLSTM RNNs), Gated Recurrent Unit
Data Handling (CDH), Electrical Power Subsystem (EPS),
Recurrent Neural Network (GRU RNN), and Deep Gated
Structures and Mechanisms, Guidance and Navigation, and
Recurrent Unit Recurrent Neural Networks (DGRU RNNs).
Thermal Control Subsystem (TCS). The ADCS stabilizes
We selected these techniques as surveyed by many previous
the vehicle and orients it in a desired direction during the
researchers [19]–[25]. Despite many researchers have built
mission despite the external disturbance torques acting on it.
their outcomes on satellite telemetry available through in-
The structure and mechanism subsystem mechanically sup-
ternet, with low level of confidence [26], [27]. Our research
port all other spacecraft subsystems, attaches the spacecraft
used telemetry data with very high-level of confidence due
to the launch vehicle, and provides for ordnance-activated
to the availability of both design documentations of each
separation [13].
satellite modules and telemetry format corresponding to
The telemetry measurements onboard the spacecraft,
data ranges of each sensor.
ensure obtaining of adequate information about the on-
board subsystems functioning during the flight operation
as well as ensure controlled counteracting the off-nominal A. Limit Checking
situations onboard the satellite. The satellite telemetry data Limit checking is the simplest algorithm, which is ap-
enables the operators in the ground station to monitor the plied earlier and widely used. The technique is based on
satellite in different situations such as separation after or- setting a proper range for the applied parameter such as
bital injection from the launcher, satellite orientation and its (temperature, voltage, and current). By monitoring the vari-

IBRAHIM ET AL.: MACHINE LEARNING METHODS FOR SPACECRAFT TELEMETRY MINING 1817

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on March 30,2024 at 08:46:16 UTC from IEEE Xplore. Restrictions apply.
 ance of each parameter, out of that range events can be easily Gaussian distribution or non- Elliptical distribution, KNN
detected. cannot solve these two kinds of problem effectively [35].
The only advantage of this algorithm is its simplicity
where limits can be set and modified to monitor spacecraft
operation. Limit checking can be applied for one sensor D. Other Approaches
value. Practically, there is a set of sensors need to be si- One of the well-known statistical algorithms is the au-
multaneously monitored to assess spacecraft performance. toregressive integrated moving average (ARIMA), which is
Hence, Limit checking is, still, not proper methodology for used for time series forecasting such as prediction of traf-
telemetry deep analysis [28], [29]. fic noise time series [36]. ARIMA usually produces low
forecasting accuracy in case of nonlinear long-term time
B. Expert System (ES) series. Therefore, it is usually combined with support vec-
tor machine (SVM), which is also a statistical algorithm,
Recently, the Artificial Intelligence is becoming an in- or with artificial neural network (ANN). The SVM and
terested field of application for automated systems; one of ANN support the ARIMA to produce better results. Zhang
its important developed algorithms is ES. ES can be applied [19] presented a hybrid of ANN with ARIMA to predict
by establishing knowledge database and knowledge-based the Canadian lynx time series. Pai and Lin [20] developed
reasoning engine; using the reasoning engine, the ES can a hybrid forecasting model consists of SVM and ARIMA
predict faults according to the telemetry data. Its disadvan- for stock price forecasting. The SVM is widely used to
tage is that a predefined knowledge rules should be set first minimize the generalized error bound in order to enhance
which requires an accurate knowledge of the system overall the performance for not only time series forecasting but
possible cases since it does not implement the self-learning also classification tasks. Yu et al. [21] made a real-time
concept. Consequently, the ES cannot produce new knowl- flood stooge forecasting using SVM. Tay and Cao [37] pro-
edge [28], [29]. posed a financial time series forecasting using SVM. The
ANN is a popular technique as well, that is used in recog-
C. Clustering Techniques nition or regression problems. Park et al. [38] presented
electric load forecasting using ANN by learning the rela-
There are several data driven software tools, such as
tionship between past, current, and future temperature read-
Orca and the Inductive Monitoring System (IMS), that have
ings and the corresponding loads. Khashei and Bijari [39]
been successfully applied to mission operations for both the
proved that ANN gives better performance than ARIMA in
Space Shuttle and the International Space Station. The IMS
time series forecasting. Recently long-short term memory
tool [30] uses K-Means clustering data mining technique
(LSTM) has been introduced as RNN architecture applied
to analyze archived spacecraft data and characterize nom-
to various real-world problems, such as protein secondary
inal interactions between selected parameters. The Orca
structure prediction [40], [41], reinforcement learning [42],
tool [31] uses a nearest neighbor approach to search for
speech recognition [43], and handwriting recognition [25].
outliers data points in multivariate data sets by calculating
It has solved several artificial problems that remain im-
the distance of each data point from neighboring points.
possible with any other RNN architecture [44]. Ballas
Iverson [32] described how such data driven techniques
et al. [45] presented video representations using Gated
have been applied to NASA mission control operations
recurrent units (GRU) RNN and stacked layers of GRU
where these “data driven” applications are able to character-
RNNs. Another technique known as Gaussian Process tech-
ize and monitor interactions between multiple parameters
nique is used for time series evolution prediction of com-
and can complement existing practice to provide valuable
plex systems across various engineering and business do-
decision support for mission controllers.
mains, such as the prediction of exchange rate in finance,
K-Means clustering is an approach of machine learning
weather, and demand for energy using mixture of experts.
techniques. The algorithm is based on partitioning of an
Chandorkar et al. [46] presented a methodology for gen-
(n) observation into k clusters in which, each observation
erating probabilistic predictions for the Disturbance Storm
belongs to the cluster with the nearest mean; it depends on
Time geomagnetic activity index based on Gaussian Pro-
minimizing the sum of within-cluster distances. The clus-
cess Regression models. Mattos et al. [47] performed non-
tering algorithm may converge to different final solutions
linear system identification in the presence of outliers using
based on the start point of search. These solutions may be
Deep recurrent Gaussian processes which comprise a pow-
local minima if the initial partition is not properly chosen;
erful kernel-based machine learning paradigm that has re-
that is why the convergence to local minima is the main
cently attracted the attention of the nonlinear system identi-
disadvantage of the K-means clustering algorithm [33].
fication community, especially due to its inherent Bayesian-
K-nearest neighbors (KNN) clustering technique de-
style treatment of the uncertainty. A Recently discussed
pends on the distance to neighboring members of a class;
technique used for machine learning is State-Space Model,
KNN splits the dataset into clusters based on a simple ma-
which is used for Identification of the nonlinear dynamical
jority vote of the nearest neighbors of each point [34]. KNN
systems. Noël et al. [48] selected state-space models with
is simple, straight, and effective, however, it cannot identify
polynomial nonlinear terms to identify hysteresis in dy-
the effect of attributes in dataset; For some cases like non-
namic systems. The researchers fitted the data to the model

1818 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 55, NO. 4 AUGUST 2019

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on March 30,2024 at 08:46:16 UTC from IEEE Xplore. Restrictions apply.
 using a rigorous two-step methodology involving weighted output is an input to the next one, and output layer. MLP
least-squares minimization. Jacob et al. [49] proposed a is free of cycling so it is called feedforward neural network
Bayesian approach to identify the battery parameters of (FNN) where the output is derived from current input and
generic fractional-order systems using state-space models do not depend on input history [22], [54].
where the latent process is not Markovian. Stathopoulos and
Karlaftis [50] proved that the multivariate state space mod- C. Recurrent Neural Network
eling of urban areas parameters is complex and tedious, RNN is an advanced form of MLP, its output depends
compared with ARIMA model, which gives high accu- not only on current input but also on previous neurons,
racy in the field of relatively short-term prediction of traffic because of cycling between neurons of hidden and output
characteristics. layers, which give better results than MLP network [55].
As for applying machine learning data mining tech- The hidden layer extracts a set of features from the input
nique in the field of spacecraft performance analysis (named vectors then they are translated into the target context by
telemetry mining), Quan Li et al. [28] and Yairi et al. [29] the output layer. The hidden and output layer outputs de-
introduced a comparison between most common machine pend on the nature of the presented problem (regression,
learning techniques applied in spacecraft telemetry mining. classification) and the applied cost function such as cross
Yairi et al. [29] have surveyed wide range of algorithms and entropy, least square errors.
techniques used in space systems data mining; the research Due to its recurrent nature, the network can maintain
concluded that “a significant issue is how we can guaran- a value inside, which enables the RNN to be used as a
tee the reliability and generality of the acquired informa- memory. However, it cannot keep a value for more than 5 to
tion from data. Effective and intuitive ways of presenting 10 time steps; this is known as vanishing gradient descent
outputs from the detection/diagnosis systems must be also [56]. RNN cannot control the timing of reset, forget and
considered, because a ML/DM technique is often used as a store of the data. Gating RNN algorithms such as Long
“black-box”. ” In this research, we overcome this issue by Short-Term Memory (LSTM) and Gated Recurrent Unit
using high reliability information from confident telemetry (GRU) are used to overcome the problems stated above by
data source. using additional gates dedicated for these purposes.

IV. MACHINE LEARNING TECHNIQUES D. Long Short-Term Memory Recurrent Neural Network

This section introduces the usage of telemetry data Hochreiter and Schmidhuber [57] have developed the
from the satellite in form of Time Series. We represent the basic Long Short-Term Memory (LSTM) architecture with
data vector: X = {x(1) ; x(2) ; . . . ; x(n) }, where each element less gates (without forget gate) and connections. Gers et al.
x(t) Rm pertaining to X is an array of m values such that [56] proposed the first modification for the LSTM architec-
(t) (t)
{x1 ; x2 ; . . . ; x(t) ture by adding the forget gate that allow LSTM to reset its
m }. Each value of m corresponds to input
variables measured in the time series telemetry data. memory cell. LSTM is one of the most effective way to carry
out learning process for the RNN, such that it can remember
A. Auto-Regressive Integrated Moving-Average values for longer time. LSTM is developed from RNN by re-
placing the RNN hidden layer neurons with LSTM blocks.
The autoregressive integrated moving average Each block has a memory cell that help to overcome the
(ARIMA) model is a generalized form of an autoregressive RNN vanishing gradient problem. LSTM block consists of
moving average (ARMA) model [51]. Both of them are memory cell to store information for longer time periods;
used to forecast time series data. ARIMA algorithm con- and three multiplication units called as gates, where each
sists of three parts; the first one is the Auto Regressive part gate use the sigmoid activation function to act as a switch
where the model uses a dependent relationship between with values 0 (gate off) and 1(gate on) [44], [55]. Srivastava
an observation and number of lagged observations. The and Lessmann [58] demonstrated that a properly configured
second part is the integration, where it uses difference LSTM model outperforms other techniques used in global
between raw observations to make a stationary time series. horizontal irradiance with satellite data.
The last part is the Moving Average where the model uses
the dependency between an observation and a residual error E. Deep Long Short-Term Memory Recurrent Neural
from a moving average model applied to lag observations. Network
The ARIMA model standard uses notation (p,d,q), where Deep Long Short-Term Memory Recurrent Neural Net-
p is The number of lag observations, d is the degree of works (DLSTM RNNs) are consisted of stacked multiple
differencing, and q is The size of the moving average layers of LSTM blocks, where each block output is an input
window [52], [53]. to the next block in next layer. It is used to maximize the
memory size in case of forecasting next values or classi-
B. Multilayer Perceptron fication problems [26]. Using more stacked layers usually
MLP is the simplest form of ANN. It consists of input enhances the prediction accuracy. This technique achieves
layer, one or more hidden layers which is used to transform higher learning capacity but needs large dataset for model
the input vectors into something that the output layer can training [59]. The key aspect of deep learning is that these
use or if there are more than one hidden layer; the one’s layers of features are not designed by human engineers;

IBRAHIM ET AL.: MACHINE LEARNING METHODS FOR SPACECRAFT TELEMETRY MINING 1819

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on March 30,2024 at 08:46:16 UTC from IEEE Xplore. Restrictions apply.
 they are learned from the dataset using a general-purpose
learning procedure [60]. Fischer and Krauss [61] applied
DLSTM to a large scale financial market prediction task
on the S&P 500, from December 1992 until October 2015;
they found that DLSTM is more suitable for the forecasting
domain rather than standard deep neural network and the
logistic regression by a very clear margin.

F. Gated Recurrent Unit Recurrent Neural Network

Gated Recurrent Unit (GRU) is an alternative form of
LSTM with more simplicity. The GRU combines the in-
put and forget gate together in one gate called update gate;
and the peephole connections is removed in-order to de-
crease the number of parameters used in calculations, this
helps to improve the training performance and accelerate
the algorithm speed [23]. The GRU does not have a sepa-
rate memory cell that is why it uses the gates to modulate
information flow in the unit. The update gate controls the Fig. 1. Correlation between the satellite telemetry parameters in
importance degree of the previous memory content, to ei- different subsystems.
ther keep it or update its content. The reset gate allows
the GRU to forget unnecessary memory content [62]. Zhao
et al. [63] applied GRU recurrent neural networks on ma- The correlation is estimated using the parameters read-
chine health monitoring systems for modern industries. ings from five successive telemetry files. These subsystems
are the most critical ones in the satellite. CDH failure means
G. Deep Gated recurrent unit Recurrent Neural Network complete loss of the satellite, EPS failure means no power
will be supplied to any subsystem, and communication sys-
Deep Gated Recurrent Unit Recurrent Neural Networks tem failure means satellite may be working by itself while
(DGRU RNNs) are built by stacking multi layers of GRU no communication with ground station (no intersatellite
units, such that each lower unit output feeds the next unit communication exists).
in next layer. It is used as storage for regression or classifi- The proposed parameters are Bt: represents battery tem-
cation tasks [24]. Deep learning neural network is making perature, Asolar: is the solar array current, Aload is the load
major advances in solving problems that have resisted the current, volt is the power bus voltage. VMP1X and VMP2X
best attempts of the artificial intelligence community for are the output power of main and reserve set of commu-
many years. It has turned out to be very good at discover- nication system transmitter. The tomb1 and tomb2 are the
ing complicated structures in high-dimensional data and is camera lens temperature. The t2gps parameter represents
therefore applicable to many domains of science, business the casing temperature of the GPS. The TMT parameter
and government [64]. Tan et al. [65] used a hierarchical is the telemetry secondary power subsystem temperature.
gated recurrent neural network to model the context infor- Thermal subsystem parameters are represented by the tem-
mation. They used gate mechanism at both word level and perature sensors reading t x mounted at external heat shield
sentence level to select words and sentences closely related of satellite body. In general, Satellite telemetry parameters
to the question. Mou et al. [66] proposed a novel deep RNN are high dimensional and highly correlated in the manner
model that can effectively analyze hyperspectral pixels as as shown in Fig. 1. Principle component analysis is usually
sequential data and then determine information categories used for dimensionality reduction; However, this method
via network reasoning. cause loss in physical representation of resulted dimen-
sions, and since domain expert knowledge will be more
V. FORMATION OF TELEMETRY PARAMETERS useful in our study, we decided to select system parameters
In a complicated system, such as satellite, telemetry pa- that reflects critical health of satellite. The correlation be-
rameters are the only indictors for satellite performance and tween these parameters allows us to select those who are
subsystems health. Total number of telemetry parameters highly correlated to each other’s, and consequently express
may be over few thousand. In this research, we select (based their behavior. The selected parameters can be treated sepa-
on system designers’ recommendation) the most essential rately (univariate with respect to other selected parameters,
parameters that indicate the satellite status/health, most of but covariate with respect to its correlated -not selected-
these parameters are much correlated to each other; as parameters); in the same time, they are more expressing
shown by performing pattern matching. This enabled us to overall satellite performance and possible failures.
find the most related parameters and subsystems behavior. In our case (Egyptsat-1), we based our study on the
Fig. 1 shows the correlation between parameters in follow- domain expert knowledge in conjunction with correlation
ing subsystems: communication subsystem, Command and matrix to select critical parameters. Similar study has been
Data Handling (CDH), Electrical Power Subsystem (EPS). carried out on x-11 Chinese satellite, where autoregressive

1820 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 55, NO. 4 AUGUST 2019

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on March 30,2024 at 08:46:16 UTC from IEEE Xplore. Restrictions apply.
 model, back propagation neural network and nonparametric
regression techniques have been compared and applied to
satellite power systems in order to predict the selected pa-
rameters with high accuracy (measured by mean percentage
error 1%) [67].

VI. EVALUATION METHODOLOGY

In this section, an evaluation of the above-mentioned al-
gorithms is introduced. We formatted the telemetry parame-
ters extracted from the raw telemetry data received from the
Egyptsat-1 satellite, so that it can be used as an input to each Fig. 2. Shewhart control chart for detecting abnormal values.
algorithm; based on this we selected the following teleme-
try parameters: power bus voltage, load current, and battery
temperature. These parameters are recommended by satel-
lite designers that can indicate potential failure of satellite
critical subsystems including power, communication and
onboard computer (these parameters have high correlation
with other parameters). The telemetry parameters are time
tagged counting for number of ticks (in seconds), so we
used the first 67% of the telemetry values for learning each
algorithm while the other 33% of the total period is used for
evaluation; as a result, about 7000–8000 value from each
parameter have been used in the evaluation process. We then
estimate the root mean square error (RMSE), mean abso-
lute error (MAE), Pearson coefficient, and r2 – correlation
[68] to assess prediction accuracy.
We started by using telemetry files from different peri-
ods of satellite lifetime; five successive raw data files have
been used from each period. The data are extracted and for- Fig. 3. ARIMA prediction result for 2008 telemetry data files.
matted to allow easy interpretation of values. As a result,
we constructed a comparison table to illustrate the accuracy LSTM was able to predict the satellite load current val-
of each algorithm. All techniques are implemented based ues (which is defined by expert as main indication of power
on Python3 programming language. The neural network subsystem health). When the LSTM predicted value ex-
techniques are implemented using Keras deep learning li- ceeds the upper control limit (UCL) or the lower control
brary on top of Google TensorFlow; figures are obtained limit (LCL), it alerts for possible failure.
using matplotlib library for Python 2D plotting; we run For each parameter, a time series is constructed and
algorithms on a computing platform with processor speed used as an input to the algorithm under evaluation. About
Intel core i5 - 2410M CPU@ 2.30 GHZ 2.30 GHZ, RAM 5000 values/readings have been used for training for each
6 GB, and 64-bit operating system windows 7. Unifying algorithm, followed by about 2700 reading used for test.
the platform of implementation allowed us to calculate the The predicted values are compared with the actual values
execution time of the tasks per telemetry file. We concluded to measure prediction accuracy. High prediction accuracy
the comparison between algorithms based on accuracy of means that the technique can inform about the future val-
prediction and time of execution. ues (either normal or abnormal state) with high probabil-
ity. Figs. 3–9 represent the results of evaluation for the
seven algorithms for “voltage sensor for power bus” for
VII. RESULTS
year 2008.
This section introduces the results of running the seven The blue lines, always behind the green and red data
selected algorithms (ARIMA, MLP, RNN, LSTM, GRU, lines- usually will not be in clear view, represent the orig-
DLSTM, and DGRU) for three telemetry datasets; each set inal/ real data values; while green lines represent the pre-
is composed of five successive telemetry files from years dicted values of data used for training; in this research we
2008, 2009, and 2010. repredicted the values of the data used for training to get
The time series regression of telemetry parameters has more accurate results for overall prediction performance of
been used to demonstrate the ability of this method to detect the algorithm under evaluation; the prediction accuracy of
possible failures; we used a simple approach to predict the the training data is not always 100%; the red lines repre-
next value(s) of one parameter and apply Shewhart control sent the prediction values of the test data. As for ARIMA
[69] to check the limits of the predicted values, as shown and LSTM, predicted values are almost identical to original
in Fig. 2. ones, so that blue lines are totally covered by both green

IBRAHIM ET AL.: MACHINE LEARNING METHODS FOR SPACECRAFT TELEMETRY MINING 1821

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on March 30,2024 at 08:46:16 UTC from IEEE Xplore. Restrictions apply.
 Fig. 4. MLP prediction result for 2008 data files. Fig. 7. GRU prediction result for 2008 data files.

Fig. 5. RNN prediction result for 2008 data files. Fig. 8. DLSTM prediction result for 2008 data files.

Fig. 6. LSTM prediction result for 2008 data files.

Fig. 9. DGRU prediction result for 2008 data files.

and red lines; on the other hand, for the other algorithms an
C. RNN Algorithm
observable difference occurs, so that visible parts of blue
D. LSTM Algorithm
lines can be distinguished.
E. GRU Algorithm
A. ARIMA Algorithm F. DLSTM Algorithm
B. MLP Algorithm G. DGRU Algorithm

1822 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 55, NO. 4 AUGUST 2019

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on March 30,2024 at 08:46:16 UTC from IEEE Xplore. Restrictions apply.
 TABLE I
Techniques Comparison

Fig. 10. Techniques Comparison according to RMSE.

Fig. 11. Techniques Comparison according to MAE.

VIII. DISCUSSION
A comparison between the proposed techniques is pre-
sented in Table I. The techniques (ARIMA, MLP, RNN,
LSTM, GRU, DLSTM, and DGRU) are applied on three
different datasets from 2008, 2009, and 2010 of Egyptsat-1
lifetime.
We calculated the root mean square error (RMSE),
Mean Absolute Error (MAE), Pearson coefficient, and r2 Fig. 12. Techniques Comparison according to Pearson.
correlation coefficient for each technique as an accuracy
measure. The average accuracy is then computed for each
year, as shown in Table I, and the overall accuracy of each consumed; multilayer perceptron (MLP) is the fastest neu-
technique is calculated in Figs. 10–13. The execution time ral network algorithm but with less accuracy. The dataset
is calculated in seconds for each process. for 2010 is relatively larger than dataset for 2008 and 2009
From the results Table I, we can see that Autoregres- which makes the accuracy of 2010 dataset relatively better
sive integrated moving average (ARIMA), has the best pre- than the others, as shown in Table I.
diction accuracy all over the whole datasets regarding to Figs. 10–13 present a comparison between different
RMSE and MAE. This statistical approach runs in a time techniques from point of view of each accuracy measure
series prediction behavior faster than all other neural net- RMSE, MAE, Pearson and r2 Coefficient. As for root mean
work techniques; moreover, ARIMA is relatively simple square error (RMSE) accuracy measure, shown in Fig. 10,
algorithm and cost-effective approach to carry on the re- ARIMA gives better accuracy followed by LSTM then
quired function. GRU; the same behavior is found in mean absolute er-
On the other hand, for the neural network algorithms, ror (MAE), as shown in Fig. 11. The correlation accuracy
long short-term memory recurrent neural network (LSTM- measure techniques: Pearson and r2 correlation, shows that
RNN) achieved the highest performance accuracy regarding the LSTM followed by GRU and DLSTM gives higher
Pearson coefficient and r2 correlation coefficient, but timely accuracy, as shown in Figs. 12 and 13.

IBRAHIM ET AL.: MACHINE LEARNING METHODS FOR SPACECRAFT TELEMETRY MINING 1823

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on March 30,2024 at 08:46:16 UTC from IEEE Xplore. Restrictions apply.
 the LSTM will be best candidate for this purpose, as it can
be used for prediction, fault diagnoses and classification.
Planned Future work is the implementation of LSTM
using wither Graphical Processing Unit or Field Pro-
grammable Gate Array for fast and real time data process-
ing, toward an integrated system for telemetry prediction,
fault diagnoses and classification. The system will be used
in operation of satellite Ground control station at Cairo,
Egypt for next satellite 2019.

ACKNOWLEDGMENT
Authors would like to acknowledge the full support from the
Fig. 13. Techniques Comparison according to r2 .
team members of Space Design Department in the National
As we can see from these figures; for artificial low Authority for Remote Sensing and Space Science-Egypt for
earth orbit satellites, that have relatively short life time (3– providing the information and required data to handle the
5 years) relative to communication and navigation satel- telemetry data set for the satellite Egyptsat-1. They would
lites (15–20 years), both regression techniques and neural also like to thank the handling editor and three anonymous
network technique (used for prediction) have very closed reviewers for their detailed and constructive comments and
accuracy measures values (especially ARIMA and LSTM). suggestions, which greatly helped in improving the quality
This can be explained due to the smaller datasets introduced of this paper.
to the neural network as learning period; terrestrial systems
and artificial communication satellites may last for longer REFERENCES
time (15–20 years life time) which means more data can be
provided for training. Fischer and Krauss [61] used LSTM [1] R. Leach
for large data set (1992–2015) with very good prediction Spacecraft system failures and anomalies attributed to the nat-
ural space environment
performance. In Proc. Space Programs Technol. Conf., 1995, paper 3564.
[2] L. Lucas and R. Boumghar
IX. CONCLUSION Machine learning for spacecraft operations support-the mars
This paper introduces comparison between machine- express power challenge
In Proc. 6th Int. Conf. Space Mission Challenges Inform. Tech-
learning algorithms (ARIMA, MLP, RNN, LSTM, DL- nol., 2017, pp. 82–87.
STM, GRU, and DGRU) used for prediction of spacecraft [3] Z. Li
telemetry data. Spacecraft parameters value are predicted Machine learning in spacecraft ground systems
using real telemetry data of Egyptsat-1 satellite. From the In Proc. 6th Int. Conf. Space Mission Challenges for Inform.
results, LSTM and GRU algorithms give a high predic- Technol., 2017, pp. 76–81.
[4] V. A. Skormin, V. I. Gorodetski, and L. J. Popyack
tion accuracy (from correlation accuracy measure point Data mining technology for failure prognostic of avionics
of view); while ARIMA and LSTM have highest predic- IEEE Trans. Aerosp. Electron. Syst., vol. 38, no. 2, pp. 388–403,
tion accuracy (from mean error accuracy measure point of Apr. 2002.
view). By applying these algorithms on presented param- [5] M. De Sanctis, I. Bisio, and G. Araniti
eters, we found that ARIMA and MLP models run with Data mining algorithms for communication networks control:
concepts, survey and guidelines
highest speed. While RNN takes relatively more time due IEEE Netw., vol. 30, no. 1, pp. 24–29, Jan./Feb. 2016.
to its recurrent nature. GRU is faster than LSTM due to its [6] C. Jiang, H. Zhang, Y. Ren, Z. Han, K.-C. Chen, and L. Hanzo
lower number of gates; however, LSTM gives better perfor- Machine learning paradigms for next-generation wireless net-
mance. DGRU and DLSTM take more time in processing works
between staked layers used in each algorithm; these algo- IEEE Wireless Commun., vol. 24, no. 2, pp. 98–105, Apr. 2017.
[7] M. J. Way, J. D. Scargle, K. M. Ali, and A. N. Srivastava
rithms give less accurate results because it requires large Advances in Machine Learning and Data Mining for Astron-
size of dataset for deep learning process. omy. Boca Raton, FL, USA: CRC Press, 2012.
The results show that, at least in Egyptsat-1 case, for [8] Y. Jiao, J. J. Hall, and Y. T. Morton
short lifetime satellites (3–5 years) it would be more effi- Automatic equatorial GPS amplitude scintillation detection us-
cient to use simple linear regression (such as ARIMA) for ing a machine learning algorithm
IEEE Trans. Aerosp. Electron. Syst., vol. 53, no. 1, pp. 405–418,
predicting critical parameters of satellite. Using neural net- Feb. 2017.
work may be more efficient in long-term prediction as the [9] S. Kauffman and S. Fischmeister
case of communication satellites (15–20 years). Mining temporal intervals from real-time system traces
We recommend simpler regression techniques such as In Proc. 6th Int. Workshop Softw. Mining, 2017, pp. 1–8.
ARIMA for implementation “for low earth orbit satellite [10] A. Gharehbaghi and M. Lindén
A deep machine learning method for classifying cyclic time
telemetry mining” that will give comparable results with series of biological signals using time-growing neural network
complex neural network. However, for building an inte- IEEE Trans. Neural Netw. Learn. Syst., vol. 29, no. 9, pp. 4102–
grated system for both telemetry mining and classification, 4115, Sep. 2017.

1824 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 55, NO. 4 AUGUST 2019

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on March 30,2024 at 08:46:16 UTC from IEEE Xplore. Restrictions apply.
 [11] J.-S. Chou and T.-K. Nguyen [27] Y. Gao, T. Yang, N. Xing, and M. Xu
Forward forecast of stock price using sliding-window Fault detection and diagnosis for spacecraft using principal
metaheuristic-optimized machine learning regression component analysis and support vector machines
IEEE Trans. Ind. Informat., vol. 14, no. 7, pp. 3132–3142, In Proc. 7th IEEE Conf. Ind. Electron. Appl., 2012,
Jul. 2018. pp. 1984–1988.
[12] B. Jokanović and M. Amin [28] Q. Li, X. Zhou, P. Lin, and S. Li
Fall detection using deep learning in range-doppler radars Anomaly detection and fault diagnosis technology of spacecraft
IEEE Trans. Aerosp. Electron. Syst., vol. 54, no. 1, pp. 180–189, based on telemetry-mining
Feb. 2018. In Proc. 3rd Int. Symp. Syst. Control Aeronautics Astronautics,
[13] W. J. Larson and J. R. Wertz 2010, pp. 233–236.
Space mission analysis and design [29] T. Yairi, Y. Kawahara, R. Fujimaki, Y. Sato, and K. Machida
Microcosm, Inc., Torrance, CA, USA, 1992. Telemetry-mining: A machine learning approach to anomaly
[14] M. Macdonald and V. Badescu detection and fault diagnosis for space systems
The International Handbook of Space Technology. New York, In Proc. 2nd IEEE Int. Conf. Space Mission Challenges Inform.
NY, USA: Springer, 2014. Technol., 2006, pp. 8 pp.-476.
[15] M. Mahmoud, A. Mahmoud, M. El-Sirafy, A. Hassan, A. Farrag, [30] D. L. Iverson
and A. Zaki Inductive system health monitoring
Micro satellites commissioning- Hands on experience In Proc. Int. Conf. Artif. Intell. (IC-AI04), CSREA, Las Vegas,
presented at the International Workshop on Small Satel- NV, Jun. 2004.
lites, New Missions, and New Technologies SSW, Turkey, [31] S. D. Bay and M. Schwabacher
Jun. 2008. Mining distance-based outliers in near linear time with random-
[16] T. Yairi, N. Takeishi, T. Oda, Y. Nakajima, N. Nishimura, and N. ization and a simple pruning rule
Takata In Proc. 9th ACM SIGKDD Int. Conf. Knowl. discovery Data
A data-driven health monitoring method for satellite house- Mining, 2003, pp. 29–38.
keeping data based on probabilistic clustering and dimension- [32] D. L. Iverson
ality reduction System health monitoring for space mission operations
IEEE Trans. Aerosp. Electron. Syst., vol. 53, no. 3, pp. 1384– In Proc. IEEE, Aerosp. Conf., 2008, pp. 1–8.
1401, Jun. 2017. [33] E. Vecchio, B. Lazzerini, S. Foley, and A. Donati
[17] B. Nassar and W. Hussein Spacecraft fault analysis using data mining techniques
State-of-health analysis applied to spacecraft telemetry based In Proc. 8th Int. Symp. Artif. Intell., Robot.Automat. Space,
on a new projection to latent structure discriminant analysis Munchen, Germany, 2005, pp. 5–8.
algorithm [34] G. Bonev
In Proc. IEEE Aerosp. Conf., 2015, pp. 1–11. Machine learning algorithms for automated satellite snow and
[18] T. Yang, B. Chen, Y. Gao, J. Feng, H. Zhang, and X. Wang sea ice detection
Data mining-based fault detection and prediction methods for ProQuest Dissertations and Theses; Ph.D. dissertation, City
in-orbit satellite Univ. New York, New York, NY, USA, 2017.
In Proc. Int. Conf. Meas., Inform. Control, 2013, pp. 805–808. [35] S. Sun and Y. Wang
[19] G. P. Zhang K-nearest neighbor clustering algorithm based on kernel
Time series forecasting using a hybrid ARIMA and neural net- methods
work model In Proc. 2nd WRI Global Congr. Intell. Syst., 2010,
Neurocomputing, vol. 50, pp. 159–175, 2003. pp. 335–338.
[20] P.-F. Pai and C.-S. Lin [36] K. Kumar and V. K. Jain
A hybrid ARIMA and support vector machines model in stock Autoregressive integrated moving averages (ARIMA) mod-
price forecasting elling of a traffic noise time series
Omega, vol. 33, pp. 497–505, 2005. Appl. Acoustics, vol. 58, pp. 283–294, 1999.
[21] P.-S. Yu, S.-T. Chen, and I.-F. Chang [37] F. E. Tay and L. Cao
Support vector regression for real-time flood stage forecasting Application of support vector machines in financial time series
J. Hydrology, vol. 328, pp. 704–716, 2006. forecasting
[22] G. Zhang, B. E. Patuwo, and M. Y. Hu Omega, vol. 29, pp. 309–317, 2001.
Forecasting with artificial neural networks: The state of the art [38] D. C. Park, M. El-Sharkawi, R. Marks, L. Atlas, and M. Damborg
Int. J. Forecasting, vol. 14, pp. 35–62, 1998. Electric load forecasting using an artificial neural
[23] J. Chung, C. Gulcehre, K. Cho, and Y. Bengio network
Empirical evaluation of gated recurrent neural networks on se- IEEE Trans. Power Syst., vol. 6, no. 2, pp. 442–449, May 1991.
quence modeling [39] M. Khashei and M. Bijari
2014, arXiv:1412.3555. An artificial neural network (p, d, q) model for timeseries fore-
[24] Y. Gao and D. Glowacka casting
Deep gate recurrent neural network Expert Syst. Appl., vol. 37, pp. 479–489, 2010.
In Proc. Asian Conf. Mach. Learn., 2016, pp. 350–365. [40] S. Hochreiter, M. Heusel, and K. Obermayer
[25] A. Graves, M. Liwicki, H. Bunke, J. Schmidhuber, and Fast model-based protein homology detection without align-
S. Fernández ment
Unconstrained on-line handwriting recognition with recurrent Bioinformatics, vol. 23, pp. 1728–1736, 2007.
neural networks [41] J. Chen and N. S. Chaudhari
In Proc. Adv. Neural Inf. Process. Syst., 2008, pp. 577–584. Protein secondary structure prediction with bidirectional lstm
[26] P. Malhotra, L. Vig, G. Shroff, and P. Agarwal networks
Long short term memory networks for anomaly detection in In Proc. Int. Joint Conf. Neural Netw.: Post-Conf. Workshop
time series Comput. Intell. Approaches Anal. Bio-Data, vol. 1, Aug. 2005,
In Proc. Conf. 23rd Eur. Symp. Artif. Neural Netw., 2015, pp. 538–541, doi: 10.1109/IJCNN.2005.1555888.
pp. 89–94.

IBRAHIM ET AL.: MACHINE LEARNING METHODS FOR SPACECRAFT TELEMETRY MINING 1825

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on March 30,2024 at 08:46:16 UTC from IEEE Xplore. Restrictions apply.
 [42] B. Bakker [56] F. A. Gers, J. Schmidhuber, and F. Cummins
Reinforcement learning with long short-term memory Learning to forget: Continual prediction with LSTM
In Proc. Adv. Neural Inf. Process. Syst., 2002, pp. 1475–1482. Neural Comput., 1999.
[43] A. Graves and J. Schmidhuber [57] S. Hochreiter and J. Schmidhuber
Framewise phoneme classification with bidirectional LSTM Long short-term memory
and other neural network architectures Neural Comput., vol. 9, no. 8, pp. 1735–1780, Nov. 15, 1997.
Neural Netw., vol. 18, pp. 602–610, 2005. [58] S. Srivastava and S. Lessmann
[44] A. Graves A comparative study of LSTM neural networks in forecasting
Supervised sequence labelling day-ahead global horizontal irradiance with satellite data
in Supervised Sequence Labelling With Recurrent Neural Net- Solar Energy, vol. 162, pp. 232–247, 2018.
works, New York, NY, USA: Springer, 2012, pp. 5–13. [59] H. Sak, A. Senior, and F. Beaufays
[45] N. Ballas, L. Yao, C. Pal, and A. Courville Long short-term memory recurrent neural network architec-
Delving deeper into convolutional networks for learning video tures for large scale acoustic modeling
representations In Proc. 15th Annu. Conf. Int. Speech Commun. Assoc., 2014,
2015, arXiv:1511.06432. pp. 338–342.
[46] M. Chandorkar, E. Camporeale, and S. Wing [60] Y. LeCun, Y. Bengio, and G. Hinton
Probabilistic forecasting of the disturbance storm time index: Deep learning
An autoregressive Gaussian process approach Nature, vol. 521, 2015, Art. no. 436.
Space Weather, vol. 15, pp. 1004–1019, 2017. [61] T. Fischer and C. Krauss
[47] C. L. C. Mattos, Z. Dai, A. Damianou, G. A. Barreto, and N. D. Deep learning with long short-term memory networks for fi-
Lawrence nancial market predictions
Deep recurrent Gaussian processes for outlier-robust system Eur. J. Oper. Res., vol. 270, pp. 654–669, 2017.
identification [62] J. Chung, C. Gulcehre, K. Cho, and Y. Bengio
J. Process Control, vol. 60, pp. 82–94, 2017. Gated feedback recurrent neural networks
[48] J.-P. Noël, A. F. Esfahani, G. Kerschen, and J. Schoukens In Proc. Int. Conf. Mach. Learn., 2015, pp. 2067–2075.
A nonlinear state-space approach to hysteresis identification [63] R. Zhao, D. Wang, R. Yan, K. Mao, F. Shen, and J. Wang
Mech. Syst. Signal Process., vol. 84, pp. 171–184, 2017. Machine health monitoring using local feature-based gated re-
[49] P. E. Jacob, S. M. M. Alavi, A. Mahdi, S. J. Payne, and D. A. current unit networks
Howey IEEE Trans. Ind. Electron., vol. 65, no. 2, pp. 1539–1548,
Bayesian inference in non-Markovian state-space models with Feb. 2018.
applications to battery fractional-order systems [64] K. Yao, T. Cohn, K. Vylomova, K. Duh, and C. Dyer
IEEE Trans. Control Syst. Technol., vol. 26, pp. 497–506, Depth-gated recurrent neural networks
Mar. 2018. arXiv:1508.03790v4 [cs.NE], Aug. 25, 2015.
[50] A. Stathopoulos and M. G. Karlaftis [65] C. Tan
A multivariate state space approach for urban traffic flow mod- Context-aware answer sentence selection with hierarchical
eling and prediction gated recurrent neural networks
Transp. Res. Part C, Emerg. Technol., vol. 11, pp. 121–135, IEEE/ACM Trans. Audio, Speech, Lang. Process., vol. 26, no. 3,
2003. pp. 540–549, Mar. 2018.
[51] G. E. Box and G. M. Jenkins [66] L. Mou, P. Ghamisi, and X. X. Zhu
Time Series Analysis: Forecasting and Control. Holden-Day, Deep recurrent neural networks for hyperspectral image classi-
1976. fication
[52] H. Akaike IEEE Trans. Geosci. Remote Sens., vol. 55, no. 7, pp. 3639–
A new look at the statistical model identification 3655, Jul. 2017.
IEEE Trans. Automat. Control, vol. 19, no. 6, pp. 716–723, [67] B. Chen, H.-Z. Fang, H.-D. Ma, and H.-Z. Fan
Dec. 1974. The study of the satellite telemetry parameters prediction
[53] D. C. Montgomery, C. L. Jennings, and M. Kulahci method based-on x-11 model
Introduction to Time Series Analysis and Forecasting, Hoboken, Recent Advances in Electrical & Electron. Eng. (Formerly Re-
NJ, USA: Wiley, 2015. cent Patents on Electrical & Electronic Engineering), vol. 8,
[54] C. W. Dawson and R. Wilby pp. 110–129, 2015.
An artificial neural network approach to rainfall-runoff [68] D. S. Moore, G. P. McCabe, and B. A. Craig
modelling Introduction to the Practice of Statistics, New York, NY, USA:
Hydrological Sci. J., vol. 43, pp. 47–66, 1998. Freeman, 2009.
[55] K. Greff, R. K. Srivastava, J. Koutnı́k, B. R. Steunebrink, and [69] C. W. Kang and P. H. Kvam
J. Schmidhuber Shewhart control charts
LSTM: A search space odyssey Basic Statist. Tools Improving Quality, New York, NY, USA:
IEEE Trans. Neural Netw. Learn. Syst., vol. 28, no. 10, John Wiley & Sons, pp. 97–124, 2011.
pp. 2222–2232, Oct. 2017.
Sara K. Ibrahim received the B.S. degree in computer and systems electrical engineering
from Zagazig University, Egypt, in 2014.
Since September 2014, she has been working as a Teaching Assistant in the Department
of Computer & Systems, Faculty of Engineering, Zagazig University. Her current research
involves artificial intelligence, machine learning, and satellite telemetry data mining.

1826 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 55, NO. 4 AUGUST 2019

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on March 30,2024 at 08:46:16 UTC from IEEE Xplore. Restrictions apply.
 Ayman Ahmed received the Ph.D. degree in engineering from Ain Shams university,
faculty of Engineering, Cairo, in 2013 and the MBA degree in management from L’École
Supérieure Libre des Sciences Commerciales Appliquées (ESLSCA), France/Egypt, in
2011.
Since 2002, he has been working as a System Engineer in Egyptian Space Program.
He is experienced in satellite technology, payload imaging camera, satellite testing and
verification, satellite system engineering. He participated in development of satellite for
space-based imaging missions, and he is currently the Principle Investigator in project for
development of space qualified camera computer.

M. Amal Eldin Zeidan received the M.Sc. and Ph.D. degrees, in 1995 and 1999, respec-
tively, in artificial intelligence techniques for computer networks from Zagazig University,
Egypt. He received the Post-Doctoral degree from Calgary University, Canada, in 2002.
He acted as Manager of specialized technical detailed studies Naivsat M.E.A (Com-
munication Satellites) from 2008–2012. He is currently acting as a Full-time Professor in
the Department of Computer & Systems, Zagazig University.

Ibrahim E. Ziedan received the M.Sc. and Ph.D. degrees, in 1964 and 1967, respectively,
in systems engineering from Aberdeen University, U.K.
He acted as the Head of the Department of Computer and System and as a Dean of
the College of Engineering from 1997. He is still acting as a Full-time Professor in the
Department of Computer and System, Zagazig University. His current fields of interest are
machine learning, computer architecture, control engineering, and image processing.

IBRAHIM ET AL.: MACHINE LEARNING METHODS FOR SPACECRAFT TELEMETRY MINING 1827

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on March 30,2024 at 08:46:16 UTC from IEEE Xplore. Restrictions apply.