Fault Detection and Identification on UAV System with CITFA Algorithm Based on Deep Learning

Full text

26
th
Iranian Conference on Electrical Engineering (ICEE2018)
978-1-5386-4916-9/18/$31.00 ©2018 IEEE
Fault detection and identification on UAV system with
CITFA Algorithm Based on Deep Learning
Mohammad Hasan Olyaei
Faculty of Electrical
Engineering, Sadjad
University of Technology,
Mashhad, Iran
mh.olyaei123@sadjad.ac.ir
Hasan Jalali
Faculty of Electrical
Engineering, Sadjad
University of Technology,
Mashhad, Iran
h.jalali144@sadjad.ac.ir
Amin Noori
Faculty of Electrical
Engineering, Sadjad
University of Technology,
Mashhad, Iran
amin.noori@sadjad.ac.ir
Najmeh Eghbal
Faculty of Electrical
Engineering, Sadjad
University of Technology,
Mashhad, Iran
najmeh.eghbal@sadjad.ac.ir
Abstract— in this paper, a new algorithm for detecting and
identifying faults in a UAV system is proposed, this algorithm uses
Color Images obtained from Time-Frequency-Amplitude
(CITFA) graphs for faults classification. The most important
innovations of CITFA algorithm are, image based processing and
classification using deep neural network. In most systems, faults
can cause irreparable costs. For this reason, detecting and
identifying faults is one of the most important issues, today. In this
paper, a variety of sensor and actuator faults are investigated. The
paper focus is mostly on deep learning and time-frequency graphs.
The selected system for fault detection and proposed algorithm
implementation is a UAV system. After designing the Linear
Quadratic Regulator (LQR) controller for the system, a variety of
faulty signals are made. Using the proposed algorithm, these
signals are converted to images. Finally, these images are classified
using the proposed algorithm, based on deep learning. The test
signals are classified into five types of faults with the accuracy of
98%.
Keywords-Deep Learning; Fault Detection; Classification; CITFA;
UAV; LQR
I. I
NTRODUCTION
Nowadays, fault detection and isolation (FDI), have become
a brand new worldwide subject. FDI can be employed to
improve systems performance. First, it is required to define what
the fault is and some of the basic concepts associated with FDI
will be introduced and discussed, as well [1].
Some papers have addressed the FDI in flight control
systems. In [3], modeling the guided missile and FDI is
presented, and fault detection is carried out for a Guided Missiles
Flight control system and compared with the normal weapons.
Proposed method is based on the linear parameter varying fault
detection filter [3].
In [4,5], the flight control system and sensor fault is
simulated in MATLAB/SIMULINK environment , and
presented Simulation of sensor faults in flight control system of
transport aircraft .FDI can used in nonlinear systems.
In[6,7] describe fault detection and identification in dynamic
systems when the system dynamics can be modeled by smooth
nonlinear differential equations including affine, bilinear or
linear parameter varying system [6, 7]. FDI method based on
Artificial Neural Networks (ANNs) is proposed in [8].
In [8], FDI based on deep learning is proposed to enhance
the fault detection, classification and prediction accuracy. It is
proved that the proposed method is pretty much efficient for
detecting faults, which cannot be detected by traditional
methods.
In [9], Deep Feature Learning Network is used as another
effective FDI methods. Fault Detection using Modelica
Language. Lee in [10], detected three kinds of abnormal states
in the air handling unit using a deep learning model. A FDI
system for Intelligent Vehicle Navigation System is designed in
[11]. Also, in [12], an automated design of FDI systems is
performed automotive Applications. It should be mentioned that
FDI can help Wind Turbine Generators and energy systems
optimization [13].
The CITFA method is superior to conventional methods for
two reasons. First, the involvement of frequency information in
CITFA images, second, use of deep learning for classification.
This paper is divided into the following sections. In section
II, the basic concepts are introduced and defined. In Section III,
the details of the CITFA algorithm are discussed and in Section
IV, the network parameters and the simulation results are
presented.
II.
BASIC CONCEPTS
A. Faults
Faults are likely to occur in any system. The fault can be due
to the undesired operation of the system sensor or actuator. Due
to the importance of the system, the significance of the fault is
varied. In other words, in sensitive systems, such as military
systems, security, airplanes, etc.
Faults can have irreparable costs, while in an insensitive
system, such as an irrigation system for green areas, the
988

26
th
Iranian Conference on Electrical Engineering (ICEE2018)
existence of fault has no effect on the goal. In general, the fault
is divided into two types of sensor and actuator faults.
Sensor fault can be due to the sensor's undesirable
performance, sensor failure, expiration, etc. which are divided
into four general categories [1]:
- Bias Fault: When the measured value in comparison with
the actual value is greater or less than a constant value.
- Freezing Fault: When the measured value is fixed on a
constant value.
- Drift Fault: When the measured value gradually deviates
from the actual value.
- Loss of Accuracy: When the measured value at any
moment is a bit deviated from the actual value of that moment.
In Fig. 1, sensor faults are shown for a signal.
The actuator fault can also be due to the actuator failure,
undesired operation and so on. These are also divided into three
general categories [1]:
- Lock in Place: Occurs when the actuator works such that
the output value reaches its maximum value.
- Float: Occurs when the actuator works such that the output
value remains constant.
- Hardover: Occurs when the output increases rapidly and
reaches its maximum value.
In Fig. 2, this actuator fault is also shown. In general, the
fault classification and their types can be shown in Fig. 3.
For each system, any of the above-mentioned faults (as in
Fig. 3) is likely to occur. It is required to design a system, which
detects and identifies the fault (that is, FDI block), and then
controls it (that is, FTC block). It should be noted that the
meanings of Fault detection (FD) and Fault Isolation (FI) are
different. FD means the presence/absence of the fault can be
detected and its occurrence can be determined. FI means that, in
addition to the presence of the fault, its location and the time of
the occurrence can be determined and isolated [14]. Once the
location and type of the fault are detected, the FTC process
becomes active and changes the controller to eliminate the fault
in the system or notifies that the sensor or actuator must be
repaired or in some cases, they should be changed. Fig. 4 shows
the block diagram of the FDI and FTC processes [15].
Figure 1. Sensor fault types.
Figure 2. Types of actuator faults.
Figure 3. Types of actuator and sensor faults.
So far, several methods have been represented and proposed
for the FDI process. For example, in [16], ANNs, Fuzzy and
Kalman filters is reviewed. In this work, a new algorithm,
CITFA, is represented, which gives great results in identifying
faults and their types. In the following sections, this method and
algorithm will be described in detail.
B. Deep Learning
Deep learning has various applications for solving different
problems. The goal of deep learning is to train a deep network
such that generates a desired output by applying the input to the
network. For example, deep learning can be used to recognize
and categorize the images [17]. The deep network eventually
learns what features belong to each category. After receiving an
image (input), the network declares the category of the image.
In Fig. 5, the results of this network are shown. In this paper, we
have used this great feature of deep learning for our specific
goal. We have used this great power of deep learning in
classification the images.
Figure 4. FDI and FTC blocks.
989

26
th
Iranian Conference on Electrical Engineering (ICEE2018)
Figure 5. The deep network results for image classification [17].
C. Time-Frequency Graph
To simplify the analysis, the signal can be analyzed in the
Time-Frequency-Amplitude (TFA) graph instead of the Time-
Amplitude graph. In TFA, the horizontal axis is time and the
vertical axis is frequency and the other axis represents the
amplitude. TFA graph shows the frequency amplitude/power at
each time. Applying a small change in the main signal, results in
a significant change in the TFA graph. An example of a TFA
graph for an audio signal is shown in Fig 6. As can be seen, the
color image obtained from the TFA graph (CITFA), contains
time and frequency information. Changing the main signal, the
CITFA image will change to a new one. The same idea is used
in this paper and the differences between the faulty signal and
the normal signal are recognized, using deep learning.
Figure 6. Signal and its TFA graph.
D. System Description for Fault Detection Purpose
To test the proposed algorithm, a system of the unmanned
aerial vehicle (UAV) is employed [18]. The UAV image is
shown in Fig. 7. For all UAVs, two types of longitudinal and
lateral motions are defined [19]. In [18], it is proved that after
the system linearization, the state space equations for
longitudinal and lateral motions are as (1) and (2), respectively
[18].
(1)
-5
-0.5944 0.8008 -9.791 -0.8747 5.077 * 10
-0.744 -7.56 -0.5294 15.72 -0.000939
A= ,
0001 0
-18
1.041 -7.406 0 -15.81 -7.284 *10
-0.05399 0.9985 -17 0 0
0.4669 0 0.9985 0.05399 0 0 0
-2.703 0 -0.
B= ,C=
00
-133.7 0
00
ªº
«»
«»
«»
«»
«»
¬¼
ªº
«»
«»
«»
«»
¬¼
003176 0.05874 0 0 0
,D= 0
00010
00100
0000-1
ªº
«»
«»
«»
«»
¬¼
The longitudinal and lateral motions equations shown in (1)
and (2), include 5 states, 5 outputs and 2 inputs. These
parameters are given in Table 1 and 2, respectively.
(2)
-0.8726 0.8789 -16.82 9.791 0
-2.823 -16.09 3.367 0 0
0.702 0.514 -2.775 0 0
A= ,
-24
0 1 0.05406 -4.088 * 10 0
-23
0 0 1.001 -7.573* 10 0
0 5.302 0.058820000
-156.5 -5.008 0 1 0 0 0
B= ,C=
11.5 -82.04 0 0 1 0 0
00 00010
0 0 0 000
ªº
«»
«»
«»
«»
«»
¬¼
ªº
«»
«»
«»
«»
¬¼
,D= 0
1
ªº
«»
«»
«»
«»
¬¼
Figure 7. UAV system [18].
990

26
th
Iranian Conference on Electrical Engineering (ICEE2018)
TABLE I.
PARAMETERS OF
L
ONGITUDINAL SYSTEM
Longitudinal
Outputs Inputs States
Va
(airspeed velocity)
įe
(elevator
deflection)
u
(forward velocity)
Į
(angle of attack)
įt
(thrust)
w
(downward
Velocity)
ࣄ
(pitch angle) - ࣄ
(pitch angle)
q
(pitch angular
Rate)
-
q
(pitch angular
Rate)
h
(height) - h
(height)
TABLE II.
PARAMETERS OF LATERAL SYSTEM
Lateral
Outputs Inputs
States
ȕ
(sideslip angle)
įa
(aileron
deflection)
v
(side velocity)
p
(roll angular rate)
įr
(rudder
deflection)
p
(roll angular rate)
r
(yaw
angular rate)
-
r
(yaw
angular rate)
ij
(Roll angle) - ij
(Roll angle)
Ȍ
(yaw angle) - Ȍ
(yaw angle)
In this paper, the focus is on the UAV's longitudinal motion
and height control. To control the height, a LQR controller is
designed. The coefficients of the controller (K) are obtained as
(3).
(3)
0.0031 0.0229 9.5982 0.6944 1.2247
00000
−− − −
=
ªº
«»
¬¼
K
The closed loop response of the system is shown in Fig. 8. It
can be seen that the flight is at height of 40 m (the desired
height).
In the following sections, identifying the system types of
faults are taken into account.
Figure 8. Closed loop response for flying at the height of 40 m.
III. T
HE
CITFA
ALGORITHM
B
ASED
ON
DEEP
LEARNING
The algorithm represented in this paper, consists of three
main steps: Building a database, training and testing. To train the
deep network, a good database is needed. At first step, a database
including both "without fault" (WOF) and "with fault" (WF)
data should be created. To this aim, different WOF and WF
signals are created and converted to CITFA. So, a collection of
images including WOF and WF information for deep network
learning is available. After training the deep network, a test
signal from the same system is applied to the network. The deep
network knows whether or not a fault has occurred in the signal.
If a fault exists, it will announce the alarms and detect this fault,
except the fault related category, sensor fault or actuator fault.
In Fig. 9, the block diagram for steps 1 and 2, are depicted. In
Fig. 10, the third step of the CITFA algorithm is showed. The
algorithm is named CITFA algorithm because CITFA images
are utilized for training the deep network. Simulation results
MATLAB and Python software are used to simulate the
work. At first step, building the database is performed using
MATLAB software. Data are categorized into 6 classes. Five
classes relate to WF data and one class relates to WOF. In this
paper, the database is made for Bias, Loss of accuracy, Float and
Drift faults. The database contains images of 200x200 pixels.
The second and third stages of the algorithm are performed using
Python software and Tensorflow backend. The deep network
parameters used in this paper are represented in Table 3.
Figure 9. First and second steps of the CITFA algorithm.
Figure 10. Third step of the CITFA algorithm.
991

26
th
Iranian Conference on Electrical Engineering (ICEE2018)
TABLE III. D
ETAILS OF DEEP NETWORK
Name Details
(or Number)
Input Image 200 x 200 pixle
Activation Function relu
Bath Size 24
Number of Class 6
Number of Convolution2D 2 (64 filter)
Number of MaxPooling2D 1
Number of Flatten 1
Dropout value 0.5
Number of Dense 2
Number of Epoch 1000
The training accuracy of the deep network is 98%. The cost
function graph for 1000 training steps is shown in Fig. 11. The
test results of the network are shown in Fig. 12. Fig. 12 shows
the probability of having each image in each class. In Fig. 13,
the results of the classification for the test images are
demonstrated. It is observed that the deep network is able to
properly classify the images in their original classes and
correctly identify them. The deep network is tested again, with
3 other signals, including Drift, Loss of Accuracy, and Float
faults. FDI results are shown in Fig. 14. According to Fig. 14, it
can be concluded that the CITFA algorithm can correctly detect
and categorize the faults.
Figure 11. Network Training Cost Function.
Figure 12. Test the network on 15 test images.
Figure 13. Test image classes detection.
Figure 14. Fault detection for new signals.
IV. C
ONCLUSION
Fault detection and identification in the UAV system was
discussed in this paper. Introducing the concept of deep learning,
a new CITFA algorithm was proposed to detect the UAV system
faults. Using the proposed algorithm, the WOF and WF signals
were converted into the CITFA images. The converted images
were used for deep network training. Finally, the results showed
that the CITFA algorithm was able to detect a signal class with
98% Accuracy.
992

26
th
Iranian Conference on Electrical Engineering (ICEE2018)
R
EFERENCES
[1] H. Alwi, C. Edwards, and C. P. Tan, Fault detection and fault-tolerant
control using sliding modes: Springer Science & Business Media,
2011.
[2] G. Feng, Z. Yang, and Y. Liu, "EKF-based fault detection for guided
missiles flight control system," in AIP Conferen ce Proceedings, 2017,
p. 060012.
[3] C. Yu, H. X. Han, and W. Min, "Missile fault detection based on
linear parameter varying fault detection filter," Information
Technology Journal, vol. 8, pp. 340-346, 2009.
[4] S. Li, H. Su, G. Zhang, and J. Wang, "Parameters modeling and fault
simulation for flight control system based on SIMULINK," in
Prognostics and System Health Management Conference (PHM-
Harbin), 2017, 2017, pp. 1-7.
[5] S. Seema and T. V. Murthy. "Simulation of sensor failure
accommodation in flight control system of transport aircraft: a
modular approach". World Journal of Modelling and Simulation, Vol.
11, No. 1, pp. 55-68, 2015
[6] J. z. Bokor and Z. n. Szabo, "Fault detection and isolation in nonlinear
systems," Annual Reviews in Control, vol. 33, pp. 113-123, 2009.
[7] T. F. Lootsma, Observer-based fault detection and isolation for
nonlinear systems: Institut for Elektroniske Systemer, Aalborg
Universitet, 2001.
[8] F. Lv, C. Wen, Z. Bao, and M. Liu, "Fault diagnosis based on deep
learning," in American Control Conference (ACC), 2016, 2016, pp.
6851-6856.
[9] G. Michau, T. Palme, and O. Fink, "Deep Feature Learning Network
for Fault Detection and Isolation."
[10] D. Lee, B. Lee, and J. W. Shin, "Fault detection and diagnosis with
modelica language using deep belief network," in Proceedings of the
11th International Modelica Conference, Versailles, France,
September 21-23, 2015, 2015, pp. 615-623.
[11] W. Huang and X. Su, "Design of a fault detect ion and isolation system
for intelligent vehicle navigation system," International Journal of
Navigation and Observation, vol. 2015, 2015.
[12] Svärd, C. (2012 ). Methods for automated design of fault detection and
isolation systems with automotive applications (Doctoral dissertation,
Linköping University Electronic Press).
[13] R. Saravanakumar, M. Manimozhi, D. Kothari, and M. Tejenosh,
"Simulation of sensor fault diagnosis for wind turbine generators
DFIG and PMSM using Kalman filter," Energy procedia, vol. 54, pp.
494-505, 2014.
[14] F. Gustafsson and F. Gustafsson, Adaptive filtering and change
detection vol. 1: Wiley New York, 2000.
[15] D. Rotondo, J.-C. Ponsart, D. Theilliol, F. Nejjari, and V. Puig, "A
virtual actuator approach for the fault tolerant control of unstable
linear systems subject to actuator saturation and fault isolation delay,"
Annual Reviews in Control, vol. 39, pp. 68-80, 2015.
[16] M. Thirumarimurugan, N. Bagyalakshmi, and P. Paarkavi,
"Comparison of fault detection and isolation methods: A review," in
Intelligent Systems and Control (ISCO), 2016 10th International
Conference on, 2016, pp. 1-6.
[17] A. Krizhevsky, I. Sutskever, and G. E. Hinton, "Imagenet
classification with deep convolutional neural networks," in Advances
in neural information processing systems, 2012, pp. 1097-1105.
[18] A. E. Ah med, A. Hafez, A. Ouda, H. E. H. Ahmed, and H. M. Abd-
Elkader, "Modeling of a small unmanned aerial vehicle," Adv Robot
Autom, vol. 4, p. 2, 2015.
[19] I. Elgayar, "Mathematical modelling, flight control system design and
air flow control investigation for low speed UAVs," City University
London, 2013.
993