You are on page 1of 77

SPACECRAFT FORMATION CONTROL: ADAPTIVE PID-EXTENDED MEMORY

RECURRENT NEURAL NETWORK CONTROLLER

A THESIS

Presented to the Department of Mechanical and Aerospace Engineering

California State University, Long Beach

In Partial Fulfillment

of the Requirements for the Degree

Master of Science in Aerospace Engineering

Committee Members:

Praveen Shankar, Ph.D. (Chair)


Joseph Kalman, Ph.D.
Justin Bailey, Ph.D.

College Designee:

Hamid Rahai, Ph.D.

Juan Gonzalez

B.S., 2005, University of California, Irvine

December 2018




ProQuest Number: 10978237




All rights reserved

INFORMATION TO ALL USERS
The quality of this reproduction is dependent upon the quality of the copy submitted.

In the unlikely event that the author did not send a complete manuscript
and there are missing pages, these will be noted. Also, if material had to be removed,
a note will indicate the deletion.






ProQuest 10978237

Published by ProQuest LLC (2019 ). Copyright of the Dissertation is held by the Author.


All rights reserved.
This work is protected against unauthorized copying under Title 17, United States Code
Microform Edition © ProQuest LLC.


ProQuest LLC.
789 East Eisenhower Parkway
P.O. Box 1346
Ann Arbor, MI 48106 - 1346
ABSTRACT

SPACECRAFT FORMATION CONTROL: ADAPTIVE PID-EXTENDED MEMORY

RECURRENT NEURAL NETWORK CONTROLLER

By

Juan Gonzalez

December 2018

In today’s space industry, satellite formation flying has become a cost-efficient

alternative solution for science, on-orbit repair and military time-critical missions. While in orbit,

the satellites are expose to the space environment and unpredictable spacecraft on-board

disturbances that negatively affect the attitude control system’s ability to reduce relative position

and velocity error. Satellites utilizing a Proportional Integral Derivative (PID) or Adaptive

controller are typically tuned to reduce the error induced by space environment disturbances.

However, in the case of an unforeseen spacecraft disturbance, such as a fault in an Inertial

Measurement Unit (IMU), the PID-based attitude control system effectiveness will deteriorate

and will not be able to reduce the error to an acceptable magnitude.

In order to address the shortcomings a PID-Extended Memory Recurrent Neural Network

(EMRNN) adaptive controller is proposed. A PID-EMRNN with a short memory of multiple

time steps is capable of producing a control input that improves the translational position and

velocity error transient response compared to a PID. The results demonstrate the PID-EMRNN

controller ability to generate a faster settling and rise time for control signal curves. The PID-

EMRNN also produced similar results for an altitude range of 400 km to 1000 km and

inclination range of 40 to 65 degrees angles of inclination. The proposed PID-EMRNN adaptive

ii
controller has demonstrated the capability of yielding a faster position error and control signal

transient response in satellite formation flying scenario.

iii
ACKNOWLEDGEMENTS

This thesis would not have been possible without the wonderful encouragement and

support I have received during my Master’s program. I would like to thank Dr. Praveen Shankar

for introducing me to the concept of formation flying and the application of neural network

adaptive controls, and for his continued support through research and problem formulation.

Without Dr. Shankar’s guidance and support, the completion of this thesis would not have been

possible. I am also thankful to Dr. Joseph Kalman and Dr. Justin Bailey for accepting to be part

of my thesis committee and for their valuable reviewing time.

I would like to thank my friends and fellow graduate students, in no particular order,

Andres Rivera, Eddie Hidalgo, Kevin Anglim and Andrew Blackney for being willing to peer

review this thesis. Also, my gratitude is extended to Dan, Nina, Ub, Mayra, Luvy, Luis and

others who have been a constant source of encouragement throughout these years. Above all, I

would like to thank my amazing wife Linda, Polo and Lola for their unconditional and constant

support in this arduous journey. Thanks to my parents Anabella and Gustavo, siblings Gabby and

Gustavo Jr., whose love and support have gotten me to where I am today.

Most importantly, I would like to dedicate this thesis to the Lord Jesus Christ for giving

me the wisdom, determination and perseverance to complete this thesis.

iv
TABLE OF CONTENTS
ABSTRACT .................................................................................................................................... ii

ACKNOWLEDGEMENTS ........................................................................................................... iv

LIST OF TABLES ......................................................................................................................... vi

LIST OF FIGURES ...................................................................................................................... vii

LIST OF NOMENCLATURE ....................................................................................................... ix

1. INTRODUCTION ......................................................................................................... 1

2. BACKGROUND AND LITERATURE REVIEW ....................................................... 6

3. CONTROLLER DESIGN ............................................................................................ 25

4. SIMULATION AND RESULTS .................................................................................. 36

5. CONCLUSION ............................................................................................................. 56

APPENDIX: SIMULINK BLOCK DIAGRAMS FOR RNN, GRAVITATIONAL AND


ATMOSPHERIC DISTURBANCE MODELING ................................................................. 58

BIBLIOGRAPHY ......................................................................................................................... 62

v
LIST OF TABLES

1. Initial Parameters for Elliptical Orbits-Leader Satellites .......................................................... 27

2. Simulation Parameters-Initial Conditions ................................................................................. 29

3. PID Tuned Controller Parameters ............................................................................................. 30

4. RNN-1A and RNN-1B Simulation Results .............................................................................. 34

5. RNN-1B Performance to Up to Six Back Steps for 50 Seconds .............................................. 35

6. RNN-1B Performance to Up to Six Back Steps for 100 seconds ............................................. 35

7. PID, PID-RNN and PID-EMRNN Simulation Parameters....................................................... 36

8. PID, PID-RNN and PID-EMRNN at Various Simulation Parameters ..................................... 36

9. PID Position Error Without Space Environment Disturbances ................................................ 41

10. Position Error with Space Environment and IMU Sensor Noise ............................................ 43

11. r- Axis Control Signal Transient Response from Figure 23 ................................................... 47

12. φ- Axis Control Signal Transient Response from Figure 24 .................................................. 47

13. z- Axis Control Signal Transient Response from Figure 25 ................................................... 47

14. Space Environment Control Output Average RSME Results (25 sims, 1000 sec) ................ 47

15. r- Axis Control Signal Transient Response from Figure 26 ................................................... 52

16. φ- Axis Control Signal Transient Response from Figure 27 .................................................. 52

17. z- Axis Control Signal Transient Response from Figure 28 ................................................... 52

18. IMU Sensor Noise Control Signal Average RMSE Results (25 sims, 1000 sec) ................... 52

vi
LIST OF FIGURES

1. ECI coordinate system [9]. ......................................................................................................... 8

2. Typical relative position of two satellites flying in formation [12]. ........................................... 9

3. Gravitational 𝐽2 acceleration profile [15]. ................................................................................ 13

4. Atmospheric drag effect profile for a satellite elliptical orbit [14]. .......................................... 14

5. Gain Scheduler Control Architecture [17]. ............................................................................... 17

6. Model Reference Adaptive Control Architecture [17]. ............................................................ 18

7. Adaptive Pole Placement Control Architecture [4]. ................................................................. 18

8. Typical Recurrent Neural Network architecture. ...................................................................... 23

9. Extended Memory Recurrent Neural Network architecture. .................................................... 26

10. PID and RNN control architecture. ......................................................................................... 32

11. PID and EMRNN (t-6) control architecture............................................................................ 33

12. Unforced position of follower satellite and leader satellite in circular orbit. ......................... 37

13. Unforced position of leader and follower satellite orbit in ECI frame. .................................. 38

14. Unforced position of follower and leader satellite in elliptical orbit. ..................................... 39

15. Position error profile without space environment disturbances. ............................................. 40

16. Velocity error profile without space environment disturbances. ............................................ 40

17. r-axis error profile with space environment disturbances. ...................................................... 44

18. φ -axis error profile with space environment disturbance. ..................................................... 44

19. z-axis error profile with space environment disturbances. ..................................................... 45

20. r-axis control signal profile with space environment disturbances. ........................................ 45

21. ϕ-axis control signal profile with space environment disturbances. ....................................... 46

22. z-axis control signal profile with space environment disturbances. ....................................... 46

vii
23. Total position error profile with disturbances and IMU noise. ............................................... 49

24. r-axis position error profile with IMU noise. .......................................................................... 49

25. φ-axis error position profile with IMU noise. ......................................................................... 50

26. z-axis error position profile with IMU noise. ......................................................................... 50

27. r-axis control signal profile with IMU noise........................................................................... 51

28. φ -axis control signal profile with IMU noise......................................................................... 51

29. z-axis control signal profile with IMU noise. ......................................................................... 52

30. Rise time mean values for range of 30, 45 and 60 deg. inclination. ....................................... 55

31. Settling time mean values for range of 30, 45 and 60 deg. inclination................................... 55

32. RNN system Simulink block diagram for r, ϕ, z-axis............................................................. 59

33. RNN simulink block diagram for r, ϕ, z-axis 10 neurons. ...................................................... 59

34. RNN neuron simulink block diagram for r, ϕ, z-axis. ............................................................ 60

35. Atmospheric disturbance simulink block diagram. ................................................................ 60

36. Gravitational disturbance simulink block diagram. ................................................................ 60

37. Gravitational disturbance simulink subsystem block diagram. .............................................. 61

viii
LIST OF NOMENCLATURE

𝑟𝑙 Leader satellite position

𝑟𝑓 Follower satellite position

𝑚𝑓 Mass of follower satellite

𝑚𝑙 Mass of leader satellite

𝑢𝑓 Follower satellite control signal

𝑢𝑙 Leader satellite control signal

𝑓𝑑𝑙 Leader satellite disturbance forces

𝑓𝑑𝑓 Follower satellite disturbance forces

𝑒 Satellite position error

CW Clohessy-Wiltshire equations

𝜇 Gravitational parameter

𝑝 Satellite relative position

G Gravitational force

𝑅𝑒 Radius of the Earth

𝐽2 Earth’s oblateness acceleration effects

𝑓𝑎𝑡𝑚,𝑙 Atmospheric forces perturbing the leader satellite

𝑓𝑎𝑡𝑚,𝑓 Atmospheric forces perturbing the follower satellite

𝑓𝑔𝑟𝑎𝑣,𝑙 Gravitational forces perturbing the leader satellite

𝑓𝑔𝑟𝑎𝑣,𝑓 Gravitational forces perturbing the follower satellite

𝜑𝑗 Radian Basis Activation Function

𝑤𝑗 (𝑡) Updated weight

η Learning rate

ix
LMS Least Mean Square algorithm

𝑓𝐼𝑀𝑈−𝑛𝑜𝑖𝑠𝑒 Disturbance forces caused by IMU sensor noise

𝑢𝑙 Leader satellite control signal

𝑢𝑓 Follower satellite control signal

RAAN Right Angle of Ascending Node

𝑔𝑒𝑎𝑟𝑡ℎ Earth’s gravity constant

RNN Recurrent Neural Network

EMRNN Extended Memory Recurrent Neural Network

PID Proportional Integral Derivative Controller

SFF Satellite Formation Flying

x
CHAPTER 1

INTRODUCTION

In recent years, Satellite Formation Flying (SFF) has attracted the attention of

government agencies, military and private industry [1] for application in missions such as

observation, inspection, repair, weather observation, sensing networks, and a new generation of

space telescopes. As more missions are planned for the near future, the advancement of

innovative SFF technologies become necessary in industry today [2].

For nearly a decade, researchers have applied modern control systems to spacecraft

formation control, which has encountered the challenge of ensuring that multiple spacecraft

remain in precise formation, avoiding in-space collisions or interferences, while controlling time

varying relative positions, velocities and attitude control of participating spacecraft. In addition,

formation control is highly susceptible to disturbances, which introduces extremely challenging

tracking problems [3]. To this end, this thesis proposes a novel direct adaptive control

architecture: Proportional Integrator Derivative-Extended Memory Recurrent Neural Network

(PID-EMRNN) which will control formation flying of satellites in orbit considering a nonlinear

dynamic system that includes 𝐽2 gravitational effects, atmospheric effects and on-board

spacecraft Inertia Measurement Unit (IMU) sensor noise-induced disturbances. The PID-

EMRNN controller will be capable of reducing translational position and velocity error in

circular and elliptical orbits, with a faster settling and rise transient response than a PID and a

PID-RNN adaptive controller.

1
1.1 Problem Statement

The control of satellite formation flying is a complex problem. Satellites flying in Low

Earth Orbit (LEO) experience space environment disturbances that cause a conventional PID

controller effectiveness to deteriorate. This is due to the inability of the PID controller to adjust

its gains to nonlinear and uncertain systems. When exploring the application of adaptive

controllers’ gain scheduling to solve the control problem of SFF, one of the disadvantages is that

the adjustment mechanism of the controller gains is precomputed off-line and, therefore,

provides no feedback to compensate for incorrect schedules. Unpredictable changes in the plant

dynamics may lead to deterioration of performance, or even to complete failure. Another

possible drawback of gain scheduling is the high design and implementation costs that increase

with the number of operating points [4].

In order to address the nonlinear translational control problem of Satellite Formation

Flying, a novel control methodology is presented. A PID controller is augmented with a novel

neural network architecture named Extended Memory Recurrent Neural Network (EMRNN) that

has the capability to generate a control signal that can directly be added to the signal generated

by the PID with its set gains. The EMRNN consists of a RNN that stores the history of multiple

time steps of its output and utilizes it to more accurately predict a control signal that will further

improve the settling time and rise time of transient response of the controller. It also reduces the

position error of the satellites in formation more than a traditional RNN.

The leader-follower satellite orbital dynamics are analyzed with respect to the center of

the Earth (ECI) coordinate system. The orbital motion of the leader, including the disturbance

and control forces, is given by:

𝜇 𝒇 𝒖
𝒓𝒍̈ = 𝒓𝒍 + 𝑚𝒅𝒍 + 𝑚𝒍 (1)
𝑟𝑙3 𝑙 𝑙

2
Where 𝑟𝑙 is the position vector of the leader satellite from the center of the Earth, 𝑓𝑑𝑙 is the

disturbance force vector and 𝑢𝑙 is the leader’s control actuating force vector acting in the leader

satellite. ml is the mass of the leader satellite and µ is the Earth’s gravitational constant.

The follower orbital motion is given by:

𝜇 𝒇 𝒖 𝑢𝑓,𝐸𝑀𝑅𝑁𝑁
𝒓𝒇̈ = 𝒓𝒇 + 𝑚𝒅𝒇 + 𝑚𝒇 + (2)
𝑟𝑓3 𝑓 𝑓 𝑚𝑓

Where 𝑟𝑓 is the position vector of the follower satellite from the center of the Earth, 𝑓𝑑𝑓 is the

disturbance force vector and 𝑢𝑓 is the follower’s control actuating force vector acting in the

leader satellite. m𝑓 is the mass of the follower satellite and 𝑢𝑓,𝐸𝑀𝑅𝑁𝑁 is the EMRNN control

actuating force.

The relative acceleration of the leader to the follower satellite is described as

𝜇 𝒇 𝒖 𝜇 𝒇 𝒖 𝒖𝒇,𝑬𝑴𝑹𝑵𝑵
𝒑̈ = (𝑟 3 𝒓𝒍 + 𝑚𝒅𝒍 + 𝑚𝒍 ) − (𝑟 3 𝒓𝒇 + 𝑚𝒅𝒇 + 𝑚𝒇 + ) (3)
𝑙 𝑙 𝑙 𝑓 𝑓 𝑓 𝑚𝑓

where the leader and follower satellite disturbance forces are described as

𝒇𝒅𝒍 = 𝒇𝒂𝒕𝒎,𝒍 + 𝒇𝒈𝒓𝒂𝒗,𝒍 (4)

𝒇𝒅𝒇 = 𝒇𝒂𝒕𝒎,𝒇 + 𝒇𝒈𝒓𝒂𝒗,𝒇 + 𝒇𝑰𝑴𝑼−𝒏𝒐𝒊𝒔𝒆 (5)

and the respective control signals are described as

𝑼𝑷𝑰𝑫+𝑬𝑴𝑹𝑵𝑵 = 𝑼𝑷𝑰𝑫 + 𝑼𝑬𝑴𝑹𝑵𝑵 (6)

𝑼𝑷𝑰𝑫 = 𝒖𝒍 − 𝒖𝒇 (7)

𝑼𝑬𝑴𝑹𝑵𝑵 = 𝒖𝒇,𝑬𝑴𝑹𝑵𝑵 (8)

It can be observed in Equation (3) that the nonlinear disturbance affects and IMU noise term will

cause the controller performance to deteriorate over time. Meanwhile, the input of the EMRNN

control signal 𝑈𝐸𝑀𝑅𝑁𝑁 will directly assist the PID controller to adjust the gains with the objective

to reduce the relative error in position to zero.

3
Where given the desired relative position 𝒓𝒍 = [𝑟𝑙,𝑟 , 𝑟𝑙,𝜙 , 𝑟𝑙,𝑧 ], the PID- EMRNN

adaptive controller demonstrates that it reduces position error to zero of a leader-follower

relative translational position and velocity in an Earth Centered Inertial (ECI) frame:

𝒆 = [𝑒𝑟 , 𝑒𝜙 , 𝑒𝑧 ] = 𝒓𝒍 − 𝒓𝒇 (9)

The EMRNN tracking performance will be compared to the performance of a PID

method and a PID-RNN method. The comparative results will demonstrate that the PID-

EMRNN architecture achieves a transient control response 𝑈𝑃𝐼𝐷+𝐸𝑀𝑅𝑁𝑁 with shorter rise time

and settling time.

1.2 Thesis Organization

In Chapter 2, the literature review and background information of SFF are presented. The

Clohessy Wiltshire (CW) dynamic equations are shown, the radial-R, along-track- S and cross-

track-W (RSW) and ECI coordinate systems that are utilized throughout the thesis are defined.

The space environment perturbations that are considered in the simulation modeling such as

gravitational and atmospheric are defined accordingly. A simulated IMU noise proliferated due

to a faulted sensor that is introduced into the model is defined. Also, in Chapter 2, an overall

review of generally used Adaptive Controllers architectures such as gain schedulers, Model

Reference Adaptive Controller (MRAC) and Adaptive Pole Placement Control (APPC), is

presented. Neural Networks (NNs) are introduced, specifically RNNs with definitions of Least

Mean Square (LMS) gradient descent method and its Radial Basis Function activation function.

In Chapter 3, the design of controller is presented, including a short description of a PID

controller and its application in this thesis. The proposed novel EMRNN algorithm is defined

and the control architectures that will be modeled such as the PID, PID-RNN and PID-EMRNN

adaptive control architectures. In addition, Chapter 3 contains the methodology to tune the PID

4
controller and the rationale behind the selection of the RNN and EMRNN performance

parameters such as the learning rate and proposed RNN-extended memory. The analysis

demonstrates the rationale for the selection of parameters.

In Chapter 4, simulations and results are presented. Simulations were performed for the

CW dynamic equations (limited to circular orbits only), elliptical orbit dynamic equations, PID

controller without any kind of perturbations, PID-RNN control architecture with Atmospheric,

Gravitational and IMU-sensor noise perturbations and PID-EMRNN control architecture also

with Atmospheric, Gravitational and IMU-sensor noise perturbations. The resultant position

error and control signal transient response plots are analyzed and discussed.

In Chapter 5, the conclusion of the thesis results is presented, and further work needed in

line with this thesis topic is discussed.

5
CHAPTER 2

BACKGROUND AND LITERATURE REVIEW

In this chapter, the background mathematical formulation and assumptions of SFF

problem for Earth-orbiting satellites is presented. The Clohessy-Wilshire (CW) dynamic

equations [5] are shown together with their limitations in application and their practical

application as a linearized system. The ECI coordinate system is briefly introduced as it is

utilized as the frame of reference in the derivation of elliptical orbit dynamics presented in

Section 3.4.

The application of adaptive control architectures such as gain schedulers, model reference

adaptive controllers, adaptive pole placement controllers and NNs in addressing the SFF problem

are presented as part of the literature review. In the last section of this chapter, a RNN control

architecture is discussed, together with an explanation of the LMS Gradient Descent Method and

Radial Basis Function (RBF) Activation Function.

2.1 Formation Flying Dynamics and Control Approach

The Clohessy-Wiltshire (CW) equations represent the orbital dynamics of a leader and

follower satellite in space [6]. The derivation of the model is based on the laws of Newton and

Kepler, and it assumes that both satellites are in a circular orbit with no perturbations. The CW

dynamics assume that the distance between the satellites is small compared to the satellites’

distance from the center of the Earth. Kristensen et al. [5] present the use of CW equations and

develop a nonlinear mathematical model in six-degrees of freedom for two spacecraft in a leader-

follower flying formation. Kristensen et al.’s model includes the mathematical expressions for

orbital perturbations originating from gravitational variations, atmospheric drag, solar radiation

and perturbations to other celestial bodies.

6
Previous work by Kristensen and Niklasson, using the linear CW equations, translational

motion equations with no disturbances, showed that the application of a Proportional Derivative

(PD) controller in a leader-follower dynamics, obtained minimal overshoot in position and

velocity error results when tuning for settling time and power consumption [7]. Kristensen and

Niklasson simulated and compared power consumption and tracking control of relative position

results for a PD, Lyapunov-based, Velocity Error Sliding Surface, Position Error Sliding Surface

and Integrator Backstepping control methods, from which the results showed the PD controller

yielding lower power consumption and considerably superior tracking.

Furthermore, the position error settling time results obtained by Kristensen and Niklasson

[7] on the application of a PD controller to a SFF CW model and the results obtained by Wu [8]

in the application of a PID controller successfully addressing triangular formation flying

rejection of space environment perturbation, which in nature tends to minimize the effects of

external disturbances and parametric uncertainties. Based on shorter settling time results in the

application of a PD controller and the ability of the integrator (I) in reducing the error to zero by

adding a control effect due to the cumulative value of the error, the selection of a PID controller

as the base controller for this thesis was made.

2.2 Coordinate Systems

The modeling of the orbital dynamic system in which two satellites are flying in

formation is derived in two different coordinate reference frames: the ECI (Earth Centered

Inertial) frame and Satellite Coordinate System (RSW). In the ECI frame, the rotations are

dependent on the position of the leader spacecraft orbit and expressed by rotation about the z-

axis, which is the angle of right ascension of the ascending node angle Ώ. Rotation around the x-

axis, which is the inclination angle of the orbit i. And a rotation angle about the z-axis, which

7
includes the true anomaly ν and the argument of perigee ω. Obtaining the total rotation of vector

written as 𝑅𝑖𝑙 = 𝑅𝑧,𝜔+𝜃 𝑅𝑥,𝑖 𝑅𝑧,Ώ where 𝑅𝑖𝑙 is the resultant position vector of the leader satellite

in space is shown in Figure 1.

FIGURE 1. ECI coordinate system [9].

2.3 The Clohessy-Wiltshire (CW) Circular Orbit Dynamic Model

The CW dynamic model for relative translational motion is utilized extensively in

formation flying analysis and because of this reason is presented as a reference only in this

thesis. The CW equations are derived using the RSW coordinate system, where the R-axis is

collinear with the position vector, S-axis is in the direction of the velocity vector aligned with the

local horizontal and W-axis is normal to the orbit plane. The CW equations of motion are

defined using two-body motion [10], [11], based on Newton’s and Kepler’s laws. As previously

mentioned, the CW model also assumes a circular orbit with gravitational perturbations and

distance between the satellites are relatively small compared to the distance of the satellites to

the center of the earth.

8
The orbital equation of motion is given by,
𝜇
𝒓̈ + 𝑟 3 𝒓 = 0 (10)

where 𝒓 is the relative position of the mass and the gravitational parameter 𝜇 = 𝐺(𝑚𝑙 + 𝑚𝑓 ),

where G is the universal gravity constant. In this thesis, 𝑚𝑙 is defined to be the mass of the leader

satellite and 𝑚𝑓 the mass of the follower satellite.

The leader and follower satellites will have the similar equations, and since the objective

is for the satellites to rendezvous or fly in formation. Forces such as thrusting, drag and gravity

effects are accounted in equations (11) and (12) as shown in Figure 2.

Equations (11) and (12) are derived for the leader and the follower satellite as follows:

𝜇 𝒇 𝒖
𝒓𝒍̈ = 𝒓𝒍 + 𝑚𝒅𝒍 + 𝑚𝒍 (11)
𝑟𝑙3 𝑙 𝑙

𝜇 𝒇 𝒖
𝒓𝒇̈ = 𝒓𝒇 + 𝑚𝒅𝒇 + 𝑚𝒇 (12)
𝑟𝑓3 𝑓 𝑓

FIGURE 2. Typical relative position of two satellites flying in formation [12].

Where 𝒇𝒅𝒍 and 𝒇𝒅𝒇 are the disturbance force terms due to external perturbations effects

and 𝒖𝒍 and 𝒖𝒇 are thrusting or control forces acting on the leader and follower satellites defined

as:

9
𝒇𝒅𝒍 = 𝒇𝒂𝒕𝒎,𝒍 + 𝒇𝒈𝒓𝒂𝒗,𝒍 (13)

𝒇𝒅𝒇 = 𝒇𝒂𝒕𝒎,𝒇 + 𝒇𝒈𝒓𝒂𝒗,𝒇 (14)

𝑓𝑎𝑡𝑚,𝑙,𝑟 𝑓𝑎𝑡𝑚,𝑓,𝑟
𝒇𝒂𝒕𝒎,𝒍 = [𝑓𝑎𝑡𝑚,𝑙,∅ ], 𝒇𝒂𝒕𝒎,𝒇 = [𝑓𝑎𝑡𝑚,𝑓,∅ ] (15)
𝑓𝑎𝑡𝑚,𝑙,𝑧 𝑓𝑎𝑡𝑚,𝑓,𝑧

𝑓𝑔𝑟𝑎𝑣,𝑙,𝑟 𝑓𝑔𝑟𝑎𝑣,𝑓,𝑟
𝒇𝒈𝒓𝒂𝒗,𝒍 = [𝑓𝑔𝑟𝑎𝑣,𝑙,∅ ], 𝒇𝒈𝒓𝒂𝒗,𝒇 = [𝑓𝑔𝑟𝑎𝑣,𝑓,∅ ] (16)
𝑓𝑔𝑟𝑎𝑣,𝑙,𝑧 𝑓𝑔𝑟𝑎𝑣,𝑓,𝑧

The leader and follower satellite gravitational and atmospheric vector formed in all three axis are

presented above and further defined in Section 2.4.1 and 2.4.2.

The controller signals represent the thrust of the spacecraft for the leader and follower

satellites, and they are defined independently for each axis as

𝑢𝑙,𝑟 𝑢𝑓,𝑟
𝒖𝒍 = [𝑢𝑙,∅ ], 𝒖𝒇 = [𝑢𝑓,∅ ] (17)
𝑢𝑙,𝑧 𝑢𝑓,𝑧

Where 𝑟 is in the radial direction, 𝜙 is the angle of rotation about the satellite orbit plane z-axis

and 𝑧 is the direction of the vector normal to the satellite orbit plane.

The relative position vector as 𝒑 = 𝒓𝒇 − 𝒓𝒍 and its derivatives.

𝜇𝒓𝒇 𝜇𝒓
𝒑̈ = − 𝑟 3
+ 𝑭𝒕𝒉𝒓𝒖𝒔𝒕 + 𝑟 3𝒍 (18)
𝑓 𝑙

After mathematical manipulation [10] solving the relative-range vector equation for the

follower:

𝒓𝒇 = 𝒓𝒍 + 𝒑 (19)

and differentiating twice we obtain:

𝒓𝒇̈ = 𝒓𝒍̈ + 𝒑̈ (20)

10
After derivation, the resultant vector components with coordinate frame define in Figure

2 are written separately as the known Clohessy-Wiltshire equations for near circular orbits:

𝑥̈ − 2𝜔𝑦̇ − 3𝜔2 𝑥 = 𝑓𝑥
𝑦̈ + 2𝜔𝑥̇ = 𝑓𝑦 (21)
𝑧̈ + 𝜔2 𝑧 = 𝑓𝑧

Where the accelerations present in the 𝑓𝑥 equation from left to right are: total, Coriolis

and centripetal acceleration. The x-axis is along the radius vector of the leader satellite, the z-

axis is along the angular momentum vector of the leader satellite, and y-axis completes the right

handed system.

Despite the CW equation’s usefulness in modeling the dynamics of SFF, it is restricted to

near circular orbits. The CW equations are not valid for elliptical orbits, do not account for

nonlinearities and 𝐽2 effects [13]. Due to this limitation, the dynamics for SFF in an elliptical

orbit, which from this point forward will be utilized to model the controllers, are derived in the

next section. The CW equations are shown and simulated to present to the reader that they are

considered and utilized to validate the circular orbit results obtained from the elliptical orbit

dynamic simulation.

Perturbing forces for the linear CW case are assumed to be zero, but when they are

considered, the satellites flying will experience orbit-dependent disturbances such as

electromagnetic torque, gravitational perturbation, atmospheric drag and solar radiation pressure.

Gravitational perturbation and atmospheric drag terms will be considered in the elliptical orbit

dynamic model.

2.4 Space Environment Perturbations

The space environment perturbations that are taken in consideration in the dynamic

model are those produced by the gravitational and atmospheric effects. The solar and

11
electromagnetic perturbing forces are not be taken in consideration for analysis in the model due

to the fact that generally these forces usually affect more distant satellites, while Earth’s oblate

shape and atmospheric drag strongly influences the motion of a satellite near the Earth [14] . In

addition to space environment perturbations, perturbing effects of an IMU sensor failure is

introduced and considered in the dynamic analysis.

2.4.1 Gravitational Perturbing Force

Satellites flying in formation in a Low Earth Orbit (LEO) will experience Earth

gravitation perturbation caused by the nonhomogeneous distribution of the Earth’s mass. Since

the Earth is not a single point mass, but an oblate body, correction factors need to be added based

on the position of the satellite. In this analysis, only the 𝐽2 effects will be considered, since this is

the largest acceleration (430 times the value of J3 effects) when considering the oblateness of the

Earth [10]. In Figure 3, it is observed that the Solar Radiation Perturbations (SRP), Sun and
𝑚
Moon acceleration magnitudes are of small magnitude (10−6 𝑠2 ) and as such not implemented in

these thesis simulations.

The gravitational force G acting on the satellite is obtained from the gradient of the scalar

potential as follows [5]:

𝑟 1 𝑟𝑥 𝑟𝑧2 𝑟
− 𝑟𝑥3 + 2 𝐽2 𝑅𝑒 2 (15 − 3 𝑟𝑥5 )
𝑟7
𝑟 1 𝑟𝑦 𝑟𝑧2 𝑟𝑦
𝐺 = 𝜇 − 𝑦3 + 𝐽2 𝑅𝑒 2 (15 − 3 𝑟5) (22)
𝑟 2 𝑟7
𝑧 𝑟 1 2 𝑟𝑧3 𝑧 𝑟
[ − 𝑟 3 + 2 𝐽2 𝑅𝑒 (15 𝑟 7 − 3 𝑟 5 ) ]
𝑠
and the Earth’s 𝐽2 gravity perturbation force 𝒇𝑔𝑟𝑎𝑣 are the latter terms (the first term represents

the gravitational potential),

12
𝑟𝑥 𝑟𝑧2 𝑟
(5 − 𝑟𝑥5 )
𝑟7
3 𝑟𝑦 𝑟𝑧2 𝑟𝑦
𝑠
𝒇𝑔𝑟𝑎𝑣 = 2 𝜇𝐽2 𝑅𝑒2 𝑹𝑖𝑠 (5 − 𝑟5) (23)
𝑟7
𝑟𝑧3 𝑧 𝑟
[ (5 𝑟 7 − 𝑟 5 ) ]

where 𝑹𝑖𝑠 = 𝑅𝑧,𝜔+𝜐 𝑅𝑥,𝑖 𝑅𝑧,Ω and 𝑅𝑒 is the radius of the Earth.

FIGURE 3. Gravitational 𝐽2 acceleration profile [15].

2.4.2 Atmospheric Drag

The proposed satellite formation flying in LEO at an altitude of 747 km will experience

atmospheric drag forces, determined by the physical geometry of the leader and follower

satellites. Since the geometric configuration of the satellites is not the focus of this thesis, the

constant atmospheric drag value for the leader and follower satellites will be chosen to be 𝐶𝑑 =

2.2. The 𝐶𝑑 value assumes a flat plate based model [14], a nominal density 𝜌𝑎𝑖𝑟 = 7.248 x10 -11

𝑘𝑔
using an assumed surface area A of 2 𝑚2 and using the velocity vector 𝑽 in the direction of
𝑚3

13
the satellite motion are utilized to determine the atmospheric drag in the model. 𝐶𝑎𝑠 is the orbit

frame transformation matrix [11]. In Figure 4, the Exponential Atmospheric Model provided by

[14] shows the effects of the atmospheric effect in altitude decay for a satellite in an elliptical

orbit in LEO.

0
1
𝒇𝒂𝒕𝒎,𝒍 = 𝒇𝒂𝒕𝒎,𝒇 = 𝐶𝑎𝑠 [2 𝜌𝑽2 𝐶𝑑 𝐴] (24)
0

and

𝑓𝑎𝑡𝑚,𝑙,𝑟 𝑓𝑎𝑡𝑚,𝑓,𝑟
𝒇𝒂𝒕𝒎,𝒍 = [𝑓𝑎𝑡𝑚,𝑙,∅ ], 𝒇𝒂𝒕𝒎,𝒇 = [𝑓𝑎𝑡𝑚,𝑓,∅ ] (25)
𝑓𝑎𝑡𝑚,𝑙,𝑧 𝑓𝑎𝑡𝑚,𝑓,𝑧

FIGURE 4. Atmospheric drag effect profile for a satellite elliptical orbit [14].

14
2.4.3 Sensor Noise Due to Inertia Measurement Unit

In order to attain a successful satellite formation orbiting around the earth, the satellites

require accurate translation and rotational orientation. The two principal types of attitude

determination hardware utilized are attitude sensors and angular velocity sensors.

The most common attitude sensors in Earth-orbiting satellites is the Earth sensor, which

optically senses the globe contour and uses this information to calculate the attitude of the

satellite with respect to the Earth. The nominal expected accuracy is 0.5 degrees, but a drawback

of attitude sensors is that they are noisy, with levels of about 0.03 degrees or higher. Sun sensors

are also very common on Earth-orbiting satellites and they are more accurate than Earth sensors,

with accuracies in the order of 0.001 degrees (RMS) and Star sensors which are the most

accurate, but less reliable than Earth and Sun sensors [16].

In terms of angular velocity sensors, the rate integrating gyros (RIGs) are commonly used

and very accurate. It can precisely estimate a satellite attitude. The RIGs are not dependent on

the spacecraft orientation in space and are utilized in various control tasks such as rate control

and damping in the attitude control system [16], [15].

In order to examine the adaptive PID-EMRNN controller response to an unexpected mid-

flight IMU-induced error, a disturbance is introduced into the control system. The disturbance

error is attributed to corrupt data in the gyroscope measurement or a malfunctioning Earth sensor

inducing a significant error in translational position and velocity. The noise is introduced into the

Simulink dynamic control system in the form of a random non-repeatable Uniform signal with

minimum magnitude of -0.2 and maximum magnitude of 0.2 that simulates the noise introduced

by the corrupted attitude data or malfunctioning sensor.

15
2.5 Adaptive Controllers and Architectures

Adaptive controllers have been utilized to address the problem of satellite formation

control. Adaptive controllers are valuable in this application since they have the capability to

adapt to new or changed circumstances. With this approach, as the plant parameters vary they

can compensate by changing the controller gains as a function of the auxiliary measurements

causing the changes.

This thesis proposes the implementation of a Direct Adaptive Controller: where the plant

model directly utilizes as an input the estimated controller signal output. Some examples of

direct adaptive controllers are a) Gain Scheduler, b) Model Reference Adaptive Control and c)

Adaptive Pole Placement Control.

A Gain Scheduler controller architecture, as the one shown in Figure 5, consists of a

look-up table and the appropriate logic for detecting the operating points and choosing the

corresponding value of the controller gain. With this approach, parameter changes due to

nonlinear plant dynamics can be compensated by changing the controller gains [4].

Model Reference Adaptive Controller (MRAC) Architecture, as the one shown in Figure

6, is derived from the model reference control (MRC) problem. In MRC, a good understanding

of the plant and the performance requirements it has to meet allow the designer to create a

reference model, that describes the desired I/O properties of the closed-loop plant. The objective

of MRC is to find the feedback control law that changes the dynamics of the plant so that its I/O

properties are exactly the same as those of the reference model. The objective of MRC is to find

the feedback control law that changes the dynamics of the plant so that its I/O properties are

exactly the same as those of the reference model.

16
FIGURE 5. Gain Scheduler Control Architecture [17].

It is clear that the MRC in order to meet its objective, must have exact knowledge of the

plant’s output vector. Therefore, when an output parameter is unknown the MRC cannot be

implemented. One way of dealing with the unknown output parameter is to use an estimate of

that parameter in the control law. The resulting control architecture are known as MRAC. In

Figure 6, a direct MRAC architecture is shown [4].

Adaptive Pole Placement Control (APPC) Architecture or self-tuning regulator, as shown

in Figure 7, is derived from the pole placement control (PPC) and regulation problems used in

LTI plants with known parameters. In a PPC, the performance parameters defined the desired

locations of poles of the close loop plant. A feedback control law is developed, which places the

poles of the closed loop plant at the desired locations. The structure of the controller and

parameter vector are chosen in order to obtain desired input and output results, but when a

parameter vector is unknown, the structure of the controller is also unknown and the PPC cannot

17
be implemented. Therefore, as in the case of the MRAC, an estimated reference parameter is

utilized to replace the unknown parameter vector.

FIGURE 6. Model Reference Adaptive Control Architecture [17].

The resulting control architecture is referred to as APPC and if the structure of the

controller is updated directly using an on-line parameter estimated is referred to as direct APPC

[4].

FIGURE 7. Adaptive Pole Placement Control Architecture [4].

18
2.6 Adaptive Control in Satellite Formation Flying

As shown in the previous section, the adaptive controllers can be effective in addressing a

number of control problems including the satellite SFF problem. For example, Lim et al. [18],

developed an adaptive back stepping control law to contribute a thrust error model for a single

thruster with misalignment and using a Lyapunov-based control design approach to solve the

relative position tracking of satellite formation flying under the presence of this disturbance.

Another example is Dong et al. [19], who developed a robust adaptive controller for SFF based

on the Lyapunov method to demonstrate control efficiency in addressing the presence of

unknown disturbances, unknown reference orbit, unknown control of leader and mass of the

follower satellite.

Despite the positive response of adaptive controls in satellite SFF, the controllers assume

the dynamics are linear and time invariant at nominal conditions, and they have stability and

command augmentation systems to meet required parameters such as gain schedules. In extreme

flying conditions, where unexpected disturbances or noise is introduce, the performance of the

conventional adaptive controllers begin to deteriorate due to the un-modeled effects of strong

nonlinearities inherit in formation flight dynamics. Hence, the advantage of a conventional

controller augmented with a neural network is that it will be able to compensate for uncertainties

in the systems due to un-modeled dynamic phenomena and unforeseen space environment.

2.7 Neural Networks Application to Satellite Formation Flying

Neural networks have wide application as augmentation to a classical controller in

aerospace vehicles control in various ways. For example, for deep-space satellite formation

flying application [20] utilized a Linear Quadratic Regulator (LQR) as a baseline for an adaptive

nonlinear single hidden layer (SHL) neural network to demonstrate the reduction of position

19
error induced by solar, lunar, and earth gravitational disturbances over 600 days. Another

example of NN application in satellite formation presented in [21], where an adaptive sliding

mode controller with the adaptive gain and NN is propose in a leader-follower tracking problem.

The numerical results for the sliding –NN controller, demonstrates that it can effectively

compensate for the effects of modelling errors and unknown disturbances in multiple spacecraft

formation flying. One more example shown in [22], utilizes an augmented LQR-J2 stems.

Shankar [23] implemented a PID controller as a baseline to a Self-Organizing Radial Basis

Function Network (SORBFN) to a military F-15 aircraft. Shankar showed that utilizing a

parametrization structure that is adapted on-line reduces the effect of inverse dynamic error

between the design model and actual dynamics. The training algorithm to grow the network and

adapt the parameters were derive from Lyapunov theory. Similarly, implemented a “sigma-pi”

neural network adaptive controller that compensated for aerodynamic and control errors of a F-

15B military aircraft is implemented [24]. Utilizing a Proportional Integral (PI) controller

augmented with the “sigma-pi” named Gen2 NN demonstrated that it reduced the tracking error

in pitch, yaw and roll dynamics between 27 and 50 percent.

The application of neural networks as demonstrated in the examples presented were place

in the control architecture to achieve specific desired results. This thesis centers on the use of a

novel type of neural network to augment a conventional PID controller. The PID-NN

architecture has been attempted in the past, but the implementation of an EMRNN is novel.

2.8 Neural Networks and Architectures

The NN controllers in modern control systems in the last decade are increasingly being

applied to find solutions to nonlinear aerospace control problems. The justification for the

utilization of a NN controller in this thesis is that the proposed adaptive controller can estimate a

20
control signal in an improved “on-line” timely manner. Neural networks are model-free

controllers and have the ability to estimate output control signal based in provided training data

(off-line training) or “learn” estimate the data as its being simulated (on-line training) as in the

case of RNN.

A NN controller utilizes the current state of the dynamic system to train, either on-line or

off-line, and utilizes this data to estimate the output control signal that will reduce the error

between the desired state and the actual state to zero. Neural networks controllers offer several

advantages over classical controllers such as:

1. Superior handle of changes in the underlying system since it is not dependent on a prior

model.

2. Superior robustness in the case of widely varying reference signal.

Based on the aforementioned characteristics and analytical results presented by Aksakalli

et al. [25] and other researchers, an adaptive neural network will yield an improved dynamic

control result over classical controllers such as PID or LQR controllers [26].

In this thesis two types of neural networks are considered for analysis and comparison as

augmentation to a PID controller: a Recurrent Neural Network and an Extended Memory

Recurrent Neural Network. A single hidden Feedforward Neural Network is not considered in

this thesis due to its inferior settling and rise time performance when compared to a RNN while

addressing trajectory problems as shown in [17] .

2.8.1 Recurrent Neural Network (RNN)

Recurrent Neural Networks are universal approximations of dynamic systems used to

model the behavior of a wide range of systems, including nonlinear control systems. A feature

21
of a RNN is its ability to “learn” sequence data and predict the next data point in that sequence,

this is accomplished either by a provided training set of data or “on-line” learning [27].

The RNN consists of three layer of neurons: the input layer, hidden layer and an output

layer. The input layer contains the input vector to the neural network and the hidden layer is

composed of several nodes that contain a sigmoidal or Radial Basis Function (RBF) activation

function. Each node estimates a function result based on the RNN ‘s ability to use its output 𝑌𝑗 at

(𝑡 − 1) in order to estimate the weight 𝑤𝑗 (𝑡). The functions are initially activated by a randomly

generated weight.

Figure 8 shows graphical representation of a RNN process. Where the subscript 𝑗 defines

the number of neurons, 𝑋𝑗 is the input to the RNN and 𝑌𝑗 is the output of the RNN.

𝑤𝑗 (𝑡) = 𝑤𝑗 (𝑡 − 1) + η (𝑌𝑗 -𝑋𝑗 ) [𝜑𝑗 (𝑋𝑗 (t − 1)) + 𝜑𝑗 (Y𝑗 (t − 1))] (26)

𝑌 = ∑𝑛𝑗=1 𝑤𝑗 (𝑡)𝜑𝑗 (𝑡) (27)

The normalized function 𝜑𝑗 is defined in Section 2.8.3 as

2
𝜑𝑗 ∶= exp (∑𝑝𝑖=1 −(𝑐𝑗,𝑖 − 𝐼𝑖 ) /𝜎𝑗2 ) (28)

In a gradient descent algorithm, the parameters (weights) of the system are adjusted at

each step in the direction of steepest descent, or in the direction of the negative of the gradient

vector of the error function. For stochastic or on-line mode, the weights are thus updated

according to equation (29).

𝑤𝑗 (𝑡) = 𝑤𝑗 (𝑡 − 1) + η (𝑌𝑗 -𝑋𝑗 ) [𝜑𝑗 (𝑋𝑗 (t − 1)) + 𝜑𝑗 (Y𝑗 (t − 1))] (29)

Where 𝑤𝑗 (𝑡) is the updated weight, η is the learning rate and 𝜑𝑗 is the value of the

activation function, 𝑌𝑗 is desired output and 𝑋𝑗 is the RNN output. The RNN is trained to

minimize the error between 𝑌𝑗 and 𝑋𝑗 [28], [29].

22
FIGURE 8. Typical Recurrent Neural Network architecture.

2.8.2 Gradient Descent Algorithm

The presented incremental algorithm is termed stochastic gradient descent algorithm and

is also refer to as the Least Mean Square (LMS) algorithm [30] or Incremental Gradient Descent.

The LMS algorithm shows to have faster performance than Recursive Least Square (RLS)

algorithms for internally recurrent networks [31].

For instance, the training algorithm computational time in seconds per training cycle vs.

the square of the number of network weights (using 200 MHz Pentium CPU) showed a slope in

training computational time of 4x 10-5 seconds. Due to the LMS or Incremental Gradient

Descent performance over other training algorithms such as the Conjugate Gradient Descent,

global Extended Kalman Filter, node-decoupled Extended Kalman Filter or multiple extended

Kalman algorithm [29]. Due to its demonstrated faster performance, the LMS training algorithm

was selected to be utilized for the novel EMRNN algorithm proposed in this thesis.

23
2.8.3 Activation Function: Radial Basis Function Neural Network (RBFNN)

The activation function is represented in the learning algorithm as 𝜑𝑗 , a Gaussian Kernel

for a pattern recognition application. The normalized function 𝜑𝑗 is defined as

2
𝜑𝑗 ∶= exp (∑𝑝𝑖=1 −(𝑐𝑗,𝑖 − 𝐼𝑖 ) /𝜎𝑗2 ) (30)

Where 𝑐𝑗,𝑖 represents the center of 𝑗𝑡ℎ neuron corresponding to the 𝑖𝑡ℎ input 𝐼𝑖 and 𝜎𝑗 2 is

its variance. The learning strategy utilizes the Fixed Centers Selected at Random approach. In

this strategy the centers 𝑐𝑗,𝑖 of the activation function remain constant throughout the training

process and are chosen at random. For this thesis an isotropic Gaussian function whose standard

deviation is fixed according to the spread of the centers is employed.

The standard deviation 𝜎𝑗 is calculated using the following equation presented by Haykin

[30] to choose an adequate standard deviation as follows:

𝜎𝑗 = 𝑑𝑚𝑎𝑥 /√2 ∗ 𝑆 (31)

Where S is the number of centers and 𝑑𝑚𝑎𝑥 is the maximum distance between any two

centers. This formula ensures that the individual activation functions are not too peaked or too

flat. As a result of the utilizing Fixed Centers Selected at Random approach, the only parameters

that need to be learned are the weights in the output layer of the network since, as explained

above, the centers are fixed.

24
CHAPTER 3

CONTROLLER DESIGN

3.1 Proportional Integral Derivative (PID) Controller

A Proportional-Integral-Derivative (PID) controller is a linear controller widely used

presently in the aerospace industry. In the basic concept of a PID controller, 𝒖 is the control

signal and 𝒆 is the control error (𝒆 = 𝒓 − 𝒚). The reference value 𝒓, is also called the set point.

Therefore, the action of a Proportional-Integral- Derivative Controller is defined by

𝑡 𝒅𝒆(𝒕)
𝒖(𝒕) = 𝐾𝑝 𝒆(𝒕) + 𝐾𝑖 ∫0 𝒆𝝉 𝑑𝜏 + 𝐾𝑑 (32)
𝒅𝒕

and the transfer function is

𝑈(𝑠)
= 𝐾𝑝 + 𝐾𝑖 /𝑠 + 𝐾𝑑 𝑠 (33)
𝐸(𝑠)

The PID controller, like various other control systems, is frequently faced with the task of

adjusting the controller parameters to obtain the desired behavior. In order to obtain a desired

control behavior when encountering nonlinear effects in the system, this thesis proposes the

augmentation of a PID controller with a novel EMRRN.

3.2 Extended Memory Recurrent Neural Network (EMRRN)

This thesis proposes the application of EMRNN, a novel type of RNN which not only

utilizes the output of the RNN 𝑌𝑗 (𝑡 − 1), but also utilizes the output values of the RNN at time

back steps (𝑡 − 2, 𝑡 − 3, … . , 𝑡 − 𝑛) in order to calculate 𝑤𝑗 (𝑡), where 𝑛 is the defined number of

back steps. The EMRNN utilizes a similar architecture as the RNN, with the exception that it

contains more than one internal recurrence. It has an input layer, one hidden layer and one output

layer. Figure 9 shows a graphical representation of a EMRNN process and shows a graphical

representation of the weight update method of the EMRNN.

25
FIGURE 9. Extended Memory Recurrent Neural Network architecture.

The input layer contains the input vector, which can also be the output vector of another

system. The hidden layer is composed of an established amount of nodes or neurons that contain

a sigmoidal or Radial Basis Function (RBF) activation function 𝜑𝑗 that estimates the trajectory

based on the recurrent history. Each neuron estimates a trajectory result based on the RNN

ability to use the current state 𝑋𝑗 (𝑡 − 1) and NN output time history of 𝑌𝑗 at (𝑡 − 1, 𝑡 − 2, … , 𝑡 −

𝑛) in order to predict the 𝑤𝑗 (𝑡). The functions, which are activated by a randomly generated

weight, compute the NN output values 𝑌𝑗 (𝑡 − 1), 𝑌𝑗 (𝑡 − 2), … , 𝑌𝑗 (𝑡 − 𝑛) where 𝑗 is the number

of neurons in the hidden layer and 𝑛 is the number of time back steps.

Each node or neuron 𝜑𝑗 in the hidden layer outputs its value to the output layer where

they are multiplied by the weight values 𝑤𝑗 (𝑡) and summed to produce an estimated output

vector in the output layer. Refer to Figure 9 for the graphical representation of an EMRNN

process. The EMRNN has the ability to be on-line trained and utilizes the gradient descent

26
algorithm (BPTT) which integrates backward in time by adding neuron values evaluated over a

single step forward. In the case of the EMRNN it will integrate backwards up to (𝑡 − 𝑛) times.

The mathematical equations for the proposed Extended Memory RNN are as follows

𝒘𝒋 (𝒕) = 𝑤𝑗 (t-1)+η{𝜑𝑗 X(t-1)+ 𝜑𝑗 Y(t-1)+ 𝜑𝑗 Y(t-2)+…….+ 𝜑𝑗 Y(t-n)} (𝒆𝒋 ) (34)

𝑆𝑢𝑚(𝜑) = 𝜑𝑗 X(t-1)+ 𝜑𝑗 Y(t-1)+ 𝜑𝑗 Y(t-2)+…….+ 𝜑𝑗 Y(t-n) (35)

𝑌 = ∑𝑛𝑗=1 𝑤𝑗 (𝑡)𝑆𝑢𝑚(𝜑) (36)

The normalized function 𝜑𝑗 is defined in Section 2.8.3 as

2
𝜑𝑗 ∶= exp (∑𝑝𝑖=1 −(𝑐𝑗,𝑖 − 𝐼𝑖 ) /𝜎𝑗2 ) (37)

where 𝑐𝑗,𝑖 represents the center of 𝑗𝑡ℎ neuron corresponding to the input 𝐼𝑖 and 𝜎𝑗 2 is its variance.

3.3 Elliptical Orbit Dynamics

For this thesis the dynamics of the leader-and-follower elliptical satellite orbits are

analyzed independently from each other with respect to the center of the Earth (ECI).

As previously mentioned, because the Clohessy-Wiltshire method for formation flying is

limited to near-circular orbital dynamics, a dynamic model applicable to both circular and

elliptical orbits is derived for utilization in this thesis. The initial condition parameters for the

elliptical orbit dynamic simulation without any control application are shown in Table 1.

TABLE 1. Initial Parameters for Elliptical Orbits-Leader Satellites

𝐴𝑙𝑡𝑖𝑡𝑢𝑑𝑒, 𝐼𝑛𝑐𝑙𝑖𝑛𝑎𝑡𝑖𝑜𝑛 𝑚𝑎𝑠𝑠, 𝑒𝑐𝑐𝑒𝑛𝑡𝑟𝑖𝑐𝑖𝑡𝑦,


𝑟𝑝𝑒𝑟𝑖𝑔𝑒𝑒 𝑟𝑎𝑝𝑜𝑔𝑒𝑒
ℎ 0
RAAN
𝐴𝑛𝑔𝑙𝑒 𝑚 𝑙 𝑔𝑒𝑎𝑟𝑡ℎ 𝑒
𝑚
747 𝑘𝑚 25 𝑑𝑒𝑔𝑟𝑒𝑒𝑠 65 𝑑𝑒𝑔𝑟𝑒𝑒𝑠 7,125 𝑘𝑚 7,878 𝑘𝑚 100 𝑘𝑔 9.81 2 0.0502
𝑠

The equations that describe the orbital motion of the leader including the disturbance and

control forces are:

𝜇 𝒇 𝒖
𝒓𝒍̈ = 𝒓𝒍 + 𝑚𝒅𝒍 + 𝑚𝒍 (38)
𝑟𝑙3 𝑙 𝑙

27
and respectively the follower orbital motion is described as

𝜇 𝒇 𝒖
𝒓𝒇̈ = 𝒓𝒇 + 𝑚𝒅𝒇 + 𝑚𝒇 (39)
𝑟𝑓3 𝑓 𝑓

where disturbance forces for the leader and follower satellite are described as

𝒇𝒅𝒍 = 𝒇𝒂𝒕𝒎,𝒍 + 𝒇𝒈𝒓𝒂𝒗,𝒍 (40)

𝒇𝒅𝒇 = 𝒇𝒂𝒕𝒎,𝒇 + 𝒇𝒈𝒓𝒂𝒗,𝒇 + 𝒇𝑰𝑴𝑼−𝒏𝒐𝒊𝒔𝒆 (41)

and the IMU noise is expressed as

𝑓𝐼𝑀𝑈,𝑟
𝒇𝑰𝑴𝑼−𝒏𝒐𝒊𝒔𝒆 = [𝑓𝐼𝑀𝑈,∅ ] (42)
𝑓𝐼𝑀𝑈,𝑧

The IMU noise 𝒇𝑰𝑴𝑼−𝒏𝒐𝒊𝒔𝒆 is simulated with a random signal injected in the system,

utilizing the MATLAB random signal generator and it only affects the follower satellite as

defined in equation (42).

In order to determine the relative position and velocity of the follower relative to the

leader, the position and velocity of the leader is subtracted from the follower as shown in

equation (43),

𝒑 = 𝒓𝒍 − 𝒓𝒇 (43)

𝒑̈ = 𝒓𝒍̈ − 𝒓𝒇̈ (44)

𝜇 𝒇 𝒖 𝜇 𝒇 𝒖
𝒑̈ = (𝑟 3 𝒓𝒍 + 𝑚𝒅𝒍 + 𝑚𝒍 ) − (𝑟 3 𝒓𝒇 + 𝑚𝒅𝒇 + 𝑚𝒇 ) (45)
𝑙 𝑙 𝑙 𝑓 𝑓 𝑓

where relative control force U is defined as

𝑼 = 𝒖𝒇 − 𝒖𝒍 (46)

𝑢𝑙,𝑟 𝑢𝑓,𝑟
𝒖𝒍 = [𝑢𝑙,∅ ], 𝒖𝒇 = [𝑢𝑓,∅ ] (47)
𝑢𝑙,𝑧 𝑢𝑓,𝑧

28
It is assumed that the orbital position 𝒑 and velocity 𝒑̇ of the leader satellite are known at

all times. The follower satellite position and velocity initial conditions are assumed known. The

objective of the simulation is to control the follower satellite and have it fly in formation at a

desired position and velocity relative to the leader satellite.

Table 2 shows the elliptical orbit dynamics and PID controller initial condition

parameters that are utilized for simulations in this thesis.

TABLE 2. Simulation Parameters-Initial Conditions


Unforced Unforced
PID PID+
Parameter Circular Elliptical PID+ RNN
Controller EMRNN
Orbit Orbit
Simulation
6000 6000 1000 1000 1000
Time (sec)
Learning
N/A N/A N/A 0.5 0.5
Rate
Follower
Satellite
25 25 25 25 25
x-direction
(km)
Follower
Satellite IC
-25 -25 25 25 25
y-direction
(km)
Follower
Satellite
-25 -25 25 25 25
z-direction
(km)

3.4 PID Control Architecture

The PID control architecture is modeled in Simulink in order to simulate the position and

velocity error between the leader and follower satellites. There is one PID controller per axis,

one RNN per axis and one EMRNN per axis. Each controller is independent from each other (see

Appendix for reference). In this model, space environment disturbances and sensor noise due to a

faulted IMU models are not considered during the design of the PID controller.

The leader and follower elliptical orbital dynamics model are simulated using equations

(38) through (47) utilizing the parameters presented in Tables 1 and 2.


29
The PID controller tuning is performed using the MATLAB Simulink library. The space

environment disturbances are simulated by modeling the disturbance forces defined in equations

(23) and (24). The PID controller without disturbances is utilized as the baseline for the PID-

RNN and PID-EMRNN control architectures shown in Figure 8 and 9. The PID was tuned

utilizing the tuning tool in the PID Controller Block in the Control System Toolbox in the

MATLAB suite. The gain parameters for the each of the PID controller in the 𝑟, φ and z-axes are

shown in Table 3 as follows:

TABLE 3. PID Tuned Controller Parameters


Parameter P I D Filter Coefficient
r-axis 0.00042735512 9.23120535e-07 0.03968426 0.43227343
φ-axis 0.000721964 4.985562e-07 0.08639192 9.92087357
z-axis 2.151429257 1.184329e-08 0.00851984 0.09477603

The 𝑟-axis represent the radial direction, 𝜙-axis is the angle of rotation about the satellite

orbit plane z- axis and 𝑧- axis represent the direction of the vector normal to the satellite orbit

plane.

The PID controller presented throughout the analysis of this thesis, utilizes the MATLAB

built-in PID Tuner algorithm in order to meet the PID objectives: closed loop stability, balanced

performance and robustness [32]. The Mathworks™ proprietary built-in algorithm for tuning

PID controllers meets these objectives by tuning the gains to achieve a balance between

performance and robustness. The tuner achieves close loop stability by keeping the system

output bounded for bounded input. Performance is achieve by having the close loop system track

reference changes and suppress disturbances as rapidly as possible. The larger the bandwidth the

faster the controller responds to changes in the reference or disturbance loop. The tuner achieves

robustness by ensuring the loop design has enough gain margin and phase margin to allow for

30
modeling errors and variations in system dynamics. The algorithm chooses a crossover

frequency based on plant dynamics and designs for a target phase margin of 60 degrees [32].

3.5 PID-RNN Control Architecture

In the PID- RNN control architecture model, space environment disturbances and sensor

noise due to a faulted IMU are considered. The concept of augmenting the PID controller with a

RNN shows that with its backpropagation online learning algorithm, it is able to accurately

estimate a control signal input and achieve a more time efficient error transient response. In the

case of formation flying, the actual position of the follower with respect to the leader and the

error between the two are inputs to the RNN controller. The RNN using the RBFN activation

function generates a control output signal that augments the control signal generated by the PID

controller.

The RNN activation function is active when the error between the leader and follower

satellite position is non-zero. By inspection of equation (48), it can be observed that when the

difference (𝑌𝑗 -𝑋𝑗 ) is non-zero, the weights will begin updating and the RNN activation function

will be activated.

𝑤𝑗 (𝑡) = 𝑤𝑗 (𝑡 − 1) + η (𝑌𝑗 -𝑋𝑗 ) [𝜑𝑗 (𝑋𝑗 (t − 1)) + 𝜑𝑗 (Y𝑗 (t − 1))] (48)

Once the algorithm is activated the RNN will use Equation (48) to make a prediction of

the weight value at 𝑤𝑗 (𝑡). The predicted weight, given that an optimized learning rate η has been

chosen, will generate a RNN control signal that will improve the controller response and provide

an adaptive superior transient response.

The PID-RNN controller architecture model is shown in Figure 10.

31
FIGURE 10. PID and RNN control architecture.

3.6 PID-EMRNN Control Architecture

The PID-EMRNN and the PID-RNN have similar control architecture. For this thesis, the

major difference between the two is that the EMRNN has been designed for practical purposes to

evaluate up to six back step history (t-6), while the RNN evaluates only one back step (t-1)

history.

The EMRNN, similar to the RNN, is expected to become active, when the error

𝒆𝒌 between the leader and follower satellite position is non-zero.

𝒘𝒋 (𝒕) = 𝑤𝑗 (t-1)+η{𝜑𝑗 X(t-1)+ 𝜑𝑗 Y(t-1)+ 𝜑𝑗 Y(t-2)+…….+ 𝜑𝑗 Y(t-6)} (𝒆𝒋 ) (49)

Once the algorithm is activated the EMRNN will evaluate the NN output for all six back

step Y(𝑡 − 6) to make a prediction of the weight value of 𝑤𝑗 (𝑡). The predicted weight, given

that an optimized learning rate η has been chosen, will generate a EMRNN control signal that

will improve the PID-EMRNN control architecture results and provide improved settling and rise

time in the transient control response.

32
The PID-EMRNN controller architecture model is shown in Figure 11.

FIGURE 11. PID and EMRNN (t-6) control architecture.

3.6.1 RNN and EMRNN LMS Algorithm Analysis

During the formulation of the neural network LMS Gradient Descent, analysis work was

conducted between two different activation function designs, which in this thesis are referred as

RNN-1A and a variant referred to as RNN-1B. The purpose of the analysis is to determine

which activation function design algorithm yields less Root Mean Square Error (RSME) results.

The activation function algorithms were simulated for 100 seconds duration in both the RNN-1A

and RNN-1B. RSME values were calculated for a range of simulations to determine which of the

two learning algorithm architectures RNN-1A or RNN-1B yield a lower RSME value. The

learning rate or step size, which is multiplied by the LMS to estimate the next point, is utilized

for both algorithms is η=0.02 and the initial weights 𝑤𝑜 were randomly generated using the

MATLAB randn() function.

The description of the learning algorithms RNN-1A and RNN-1B are as follows:

33
RNN-1A activation function design consist of LMS using a prior weight 𝑤(𝑡 − 1) and as

shown below.

𝑤𝑗 (𝑡) = 𝑤𝑗 (t-1)+η𝜑𝑗 [X𝑗 (t − 1) + Y𝑗 (t − 1)](𝑌𝑗 − 𝑋𝑗 ) (50)

𝑤𝑗 (𝑡) = 𝑤𝑗 (t-1)+η𝜑𝑗 [X𝑗 (t − 1) + Y𝑗 (t − 1)](𝑒𝑗 ) (51)

RRN-1B at (t-1) activation function design consist of LMS using a prior weight 𝑤(𝑡 − 1)

as shown below.

The LMS activation function design with distributed nodes multiplied at each state and

NN output value:

𝑤𝑗 (𝑡) = 𝑤𝑗 (𝑡 − 1) +η{𝜑𝑗 X𝑗 (t − 1)+ 𝜑𝑗 Y𝑗 (t − 1)} (𝑌𝑗 -𝑋𝑗 ) (52)

𝑤𝑗 (𝑡) = 𝑤𝑗 (𝑡 − 1) +η{𝜑𝑗 X𝑗 (t − 1)+ 𝜑𝑗 Y𝑗 (t − 1)} (𝑒𝑗 ) (53)

Furthermore, two RNN models were constructed, one with the LMS Gradient Descent

(RNN-1A) and one with a Distributed (RNN-1B) as part of the activation function. Each RNN

has a hidden layer with 10 neurons. The input signal simulated consisted of a Simulink generated

Uniform random signal with a minimum of -1.0 and a maximum of 1.0 not repeatable signal.

The simulations duration are 100 seconds in length. See Table 4 for results.

TABLE 4. RNN-1A and RNN-1B Simulation Results


Average Results RNN (RNN-1A) RNN Distributed (RNN-1B)
RMSE- 50 Simulations 299.347 108.2674
RMSE- 100 Simulations 2.83E+20 3.50E+13
RMSE- 200 Simulations 5.72E+53 1.17E+14
RMSE- 300 Simulations 9.51E+14 2.96E+48

MAE - 50 Simulations 65.1692 1.29E-13


MAE- 100 Simulations 217.1815 4.42E-11
MAE- 200 Simulations 8.94E+13 7.65E-08
MAE- 300 Simulations 4.22E+12 3.25E-09

Per the RMSE and MAE results presented in Table 4, it can be determined that the RNN-

1B activation function design obtains better results than RNN-1A.

34
Based on simulation results, it is observed that the RNN-1B activation function design

offers lower RMSE and MAE results. The RNN-1B simulations are conducted over 50 and 100

second durations.

The input to the RNN, once again, is a random signal of type Uniform with a magnitude

of -1 to 1. The RNN-1B is simulated to observe its behavior as it evaluates a time history of the

NN output up to six back steps. Results are shown in Table 5.

TABLE 5. RNN-1B Performance to Up to Six Back Steps for 50 Seconds


Average RNN (t-1) RNN (t-2) RNN (t-3) RNN (t-4) RNN (t-5) RNN (t-6)
Results
RMSE- 50 1.75E-07 6.95E-08 2.14E-08 5.92E-08 4.46E-08 1.74E-08
Simulations
RMSE- 100 1.59E-04 1.73E-07 1.52E-07 6.60E-08 6.32E-08 8.39E-08
Simulations
RMSE- 200 8.82E-07 4.24E-07 2.14E-07 2.19E-07 1.63E-07 1.40E-07
Simulations
RMSE- 300 1.43E-05 3.81E-07 3.09E-07 3.16E-07 2.48E-07 1.69E-07
Simulations

TABLE 6. RNN-1B Performance to Up to Six Back Steps for 100 Seconds


RMSE- 50 4.90E-07 4.90E-03 1.22E-07 9.48E-08 2.97E-08 5.76E-08
Simulations
RMSE- 100 2.18E-04 2.34E-07 2.03E-07 1.63E-07 1.27E-07 1.19E-07
Simulations

The results shown in Table 5 and Table 6 indicate that a RNN with “distributed” nodes

activation function design RNN-1B evaluating a time history of the NN output up to six back

steps has superior performance than the RNN with a time history of one back step. The RNN

with a “distributed” activation function design will be define as Extended- Recurrent Neural

Network or EMRNN for the remainder of this thesis.

35
CHAPTER 4

SIMULATION AND RESULTS

The simulation results presented in this chapter are shown in various sections. The

simulation results shown in Section 4.1 and 4.2 are obtained using unforced dynamics for both

the CW dynamic model and the elliptical orbit dynamic model. The simulation results shown in

Sections 4.3, 4.4 and 4.5 are obtained in the range of values in shown in Table 7.

TABLE 7. PID, PID-RNN and PID-EMRNN Simulation Parameters


Inclination Altitudes (km) RAAN Simulation Control

(deg.) (deg.) Time (sec) Axes

PID 45 747 60 1,000 r, ϕ, z

PID-RNN 45 747 60 1,000 r, ϕ, z

PID-EMRNN 45 747 60 1,000 r, ϕ, z

The simulation results shown in Sections 4.6 are obtain from evaluating the controllers in

the range of values in shown in Table 8.

TABLE 8. PID, PID-RNN and PID-EMRNN at Various Simulation Parameters


Inclination Altitudes (km) RAAN Simulation Control

(deg.) (deg.) Time (sec.) Axes

PID 30, 45, 60 400, 747, 1,000 60 500 r, ϕ, z

PID-RNN 30, 45, 60 400, 747, 1,000 60 500 r, ϕ, z

PID-EMRNN 30, 45, 60 400, 747, 1,000 60 500 r, ϕ, z

36
4.1 Results for Unforced CW Frame Model Dynamics for a Circular Orbit

The results in Figures 12 and 13 show an oscillatory behavior of the relative motion and

velocity of the follower with respect to the leader satellite. As observed, the 𝑟-axis and 𝜑 axis

are independent from the 𝑧- axis because their motion does not couple with 𝑧- axis. For orbits

with very small eccentricities, the satellite in the higher orbit will appear to move backward in

relation to the satellite in the lower orbit because its period is longer and its velocity slower. The

conditions for this simulation are for an inclination of 20 degrees and a right ascension of the

ascending node of 60 degrees. The follower satellite is moving relative to the leader satellite,

which is orbiting the Earth at an altitude of 600 km at the velocity 𝑉 and the radius of the Earth

𝑅𝑒 .

𝜇
𝑉 = √𝑟 𝑤ℎ𝑒𝑟𝑒 𝑟𝑙 = 𝑅𝑒 + 600 𝑘𝑚 (54)
𝑙

FIGURE 12. Unforced position of follower satellite and leader satellite in circular orbit.

37
FIGURE 13. Unforced position of leader and follower satellite orbit in ECI frame.

Figures 12 shows the position vectors of the follower and leader satellites initially in the

RSW frame and subsequently transformed into ECI frame. The 3-axis figure shows how the

follower satellite has the oscillatory behavior predicted by the relative dynamics presented above

[10], [17], [15].

4.2 Results for Unforced Elliptical Orbit Dynamics

The results in this section show the unforced elliptical orbit dynamics derived in Section

3.4, with the initial orbital parameters shown in Table 1. As it can be observed the position

profile of the the leader and follower satellites are flying in the ECI coordinate frame system in

an orbital behavior validated by El-Shaboury et al. [33]. The dynamic simulation was conducted

unforced, with no environmental disturbances nor noise, and is utilized as the baseline for the

analysis of the PID, RNN and EMRNN controllers, including pertubations.

38
FIGURE 14. Unforced position of follower and leader satellite in elliptical orbit.

4.3 Results for PID Controller Without Environmental Disturbances

In this section, the simulation results of a PID controller without the effect of any

environmental disturbances or noise are presented. The simulation was performed in Simulink

for a duration of 1500 seconds. The initial conditions, orbital parameters and PID tuned

parameters are shown in Tables 1 through 3. As shown in Figure 15 and 16 the position and

velocity error are plotted against time, yielding position error results. The orbital dynamics in

each r, ϕ, and z -axes are simulated, in separate PID controllers and each controller tuned

independently.

39
FIGURE 15. Position error profile without space environment disturbances.

FIGURE 16. Velocity error profile without space environment disturbances.

40
The PID controller reduced the position error to nearly zero in approximately 1,400

seconds and the velocity error is reduced to zero in nearly 400 seconds. These results are the

baseline position responses magnitudes and durations that are utilized when comparing the

performance of the augmented PID-RNN and PID-EMRNN controllers in the following sections.

TABLE 9. PID Position Error Without Space Environment Disturbances


Maximum Position Error Error Magnitude
r- error max 2.207 km
φ-error min 2.555 km
z-error min 5.495 km

4.4 Results of PID, RNN, EMRNN with Space Environment Disturbances

In this section the PID, PID-RNN and PID-EMRNN control architectures are simulated

taking in consideration atmospheric and gravitational disturbances. The simulation results of

relative position error and control signal effort are generated for each 𝑟, 𝜙 𝑎𝑛𝑑 𝑧 axis. In fact, as

mentioned in the previous section, the dynamics in each axis are modeled separately. There is

one PID controller per axis, one RNN per axis and one EMRNN per axis. In addition, there exist

separate MATLAB S-functions written for the RNN and EMRNN on each axis. In short, there

are three controllers designed to work concurrently at any given time a simulation is performed.

This architecture will provide the ability to assume independent actuators (thrusters) in every

axis.

As shown in section 3.4, the PID was tuned utilizing the auto-tuning tool from MATLAB

Simulink and the gravitational, atmospheric and time varying potential disturbances are

introduced to the dynamics of both the leader and follower satellites. The space environment

disturbances are modeled to affect the control efforts of formation flying at an altitude of 747

km.

41
The atmospheric disturbances are simulated utilizing equation (24) in Section (2.4.2) and

are only simulated it tangential to the velocity vector 𝑽 in the direction of the satellite’s motion.

The modeled atmospheric disturbance is introduced to both the leader and follower dynamics.

The values of the assumed satellite surface area, atmospheric density, drag coefficient are

provided in Section 2.4.2.

The gravitational disturbances are simulated utilizing equation (25) in Section 2.4.1 for

all three axes of motion. The modeled disturbances are introduce to both the leader and follower

dynamics. The values of the assumed 𝐽2 coefficient, gravitational constant and Earth radius are

provided in Section 2.4.1.

The performance of the PID controller to compensate despite the introduction of

disturbance effects is demonstrated in simulation results shown in Figures 17 through 19.

Simulations are also performed with the proposed adaptive control architectures utilizing

the PID-RNN and PID-EMRNN. The performance of the PID-RNN with one back step memory

(t-1) and the PID-EMRNN with six back steps memory (t-6) are introduced in this simulation as

proposed in Figure 10 and 11. The learning rates, for the RNN and EMRNN, were optimized

through trial and error method, yielding the most desirable learning rate 𝜂 of 0.005 in the r-axis,

0.005 in the φ- axis and 0.01 in the z-axis. The learning rate was determined methodically in the

range of 0.002 through 1.0, by first testing in steps of 0.05 in order to identify the region with

optimal learning. Once the range from 0 to 0.1 was identified, the range was tested incrementally

by 0.01 until the optimal learning rate was found. This process was performed for all three axes

as shown in Figure 17 through 19. The RNN system, subsystem and neuron Simulink block

diagrams are shown in the Appendix. The actual design of both the PID-RNN and PID-EMRNN

design architecture are shown in Sections 3.4 and 3.6.

42
In the PID-RNN and PID-EMRNN adaptive controller, the error and current state

position are input vectors. These two vectors are inputs to each neuron located in the neural

network hidden layer. Each of the PID-RNN and PID-EMRNN have ten neurons in each r, φ and

z- axis. Inside each neuron, the simulation model for the LMS RBF activation function has been

constructed using Simulink and MATLAB S-function. In order to initialize the activation

function, the initial weights values are generated randomly using the MATLAB built-in function

rand () and the centers of the RBF remain fixed in order to simplify the computation of the

updated weights [30]. Inside the activation function, an S-function named RNN_weight and

EMRNN_weight respectively, are introduced to evaluate the updated weights based on one back

step time history in the case of the RNN and six back step time history in the case of the

Extended RNN.

After the weights are generated, they are each multiplied by the sum of the activation

functions φ per equation (28) and the output 𝑦𝑛𝑛 is obtained.

The PID, PID-RNN and PID-EMRNN parameters and simulation results are as follows:

TABLE 10. Position Error with Space Environment and IMU Sensor Noise
Space IMU Sensor
Environment Noise
Time (sec) Max Error(km) Time (sec) Max Error(km)
r-axis
PID 95.34 2.555 94.57 2.555
PID-RNN 94.96 2.534 93.94 2.513
PID-EMRNN 107.8 1.829 118.6 1.473
φ-axis
PID 60.22 2.207 61.21 2.207
PID-RNN 60.48 2.129 60.79 2.157
PID-EMRNN 62.22 1.896 62.53 1.956
z-axis
PID 428.1 5.494 422 5.49
PID-RNN 434.5 5.4 443.3 5.022
PID-EMRNN 465 4.536 454.3 4.77

43
FIGURE 17. r-axis error profile with space environment disturbances.

FIGURE 18. φ -axis error profile with space environment disturbance.

44
FIGURE 19. z-axis error profile with space environment disturbances.

FIGURE 20. r-axis control signal profile with space environment disturbances.

45
FIGURE 21. ϕ-axis control signal profile with space environment disturbances.

FIGURE 22. z-axis control signal profile with space environment disturbances.

46
Simulation results with learning rates 𝜂𝑟 = 0.005, 𝜂𝜑 = 0.005, 𝜂𝑧 = 0.01.

TABLE 11. r- Axis Control Signal Transient Response from Figure 23


Parameter PID RNN (t-1) EMRNN(t-6)
Rise Time (sec) 15.045 15.082 15.115
Settling Time (sec) 150.2 151.3 156.4

TABLE 12. φ- Axis Control Signal Transient Response from Figure 24


Parameter PID RNN (t-1) EMRNN(t-6)
Rise Time (sec) 3.4849 3.4849 3.4853
Settling Time (sec) 3.3306 3.3302 3.326

TABLE 13. z- Axis Control Signal Transient Response from Figure 25


Parameter PID RNN (t-1) EMRNN(t-6)
Rise Time (sec) 58.94 58.72 55.356
Settling Time (sec) 754.89 778.53 454.44

TABLE 14. Space Environment Control Output Average RSME Results (25 sims, 1000 sec)
Parameter r- Axis (sec) φ- Axis (sec) z - Axis (sec)
PID 0.018081 1.0639 0.0024095
PID + RNN(t-1) 0.018443 1.1422 0.0024203
PID + EMRNN (t-6) 0.017169 1.1414 0.0023039

In Figures 20, 21 and 22 it can be observed that the PID-EMRNN controller achieves an

improved position error trajectory over the PID and PID-RNN controllers. The r-axis and z-axis

have a significant improvement in error curve trajectory, while the error curve trajectory in the

φ-axis is marginally improved.

In terms of controller transient response, as shown in Table 9 through 13 in the r-axis the

PID controller achieves a faster Rise Time than the PID-RNN and PID-EMRNN controller, but

by approximately 0.04 and 0.07 seconds respectively. In the φ- axis, the PID and PID-RNN

controller are 0.001 seconds faster than the PID-EMRNN and in the z-axis the EMRNN is 3.5

seconds faster than the PID and PID-RNN controllers. In conclusion, in each axis, the

performance on the three control architectures is similar. There is only a 3.5 seconds

improvement from the addition of the EMRNN controller to the PID in the z-axis.

47
The results also show the Settling Time for the PID-EMRNN controller achieves a

significantly shorter Settling Time than the PID and PID-RNN controllers. In the r-axis, the

EMRNN is approximately 6 seconds slower, in the φ-axis is marginally faster by approximately

0.01 seconds and in the z-axis the settling time is a 300.45 seconds faster than the PID-RNN

controller.

The analysis above shows that the addition of an EMRNN to a PID controller in a

satellite formation flying application with space environment disturbances, as shown in Chapter

4, has a significant improvement in its trajectory. The EMRNN can yield an improved transient

response in the r-axis and z- axis, with a negligible impact to the φ- axis control performance.

These results translate into the possibility of reducing the time in which the thrusters are utilized

in order to accomplish formation flying successfully and by doing so, saving propellant

utilization.

4.5 Results of PID, RNN and EMRNN Including IMU Sensor Noise

Utilizing the same PID controller initial conditions and neural network (RNN and

EMRNN) initial parameters, an IMU sensor noise is introduced to the control system. For

implementation, an IMU sensor failure noise is added to the space environment activated in the

model at 200 seconds. The simulated noise is applied to all three axes utilizing the MATLAB

Simulink Random Signal Block “Uniform” to generate a random signal that simulates the

desired noise. The amplitude of the IMU noise is of +/-0.2 in magnitude.

The results show how the adaptive PID- RNN and PID-EMRNN control architectures

respond to the introduction of the IMU sensor noise.

48
FIGURE 23. Total position error profile with disturbances and IMU noise.

FIGURE 24. r-axis position error profile with IMU noise.

49
FIGURE 25. φ-axis error position profile with IMU noise.

FIGURE 26. z-axis error position profile with IMU noise.

50
FIGURE 27. r-axis control signal profile with IMU noise.

FIGURE 28. φ -axis control signal profile with IMU noise.

51
FIGURE 29. z-axis control signal profile with IMU noise.

Simulation results with learning rates 𝜂𝑟 = 0.005, 𝜂𝜑 = 0.005, 𝜂𝑧 = 0.01.

TABLE 15. r- Axis Control Signal Transient Response from Figure 26


Parameter PID RNN (t-1) EMRNN(t-6)
Rise Time (sec) 15.045 15.257 13.696
Settling Time (sec) 150.2 830.92 100.16

TABLE 16. φ- Axis Control Signal Transient Response from Figure 27


Parameter PID RNN (t-1) EMRNN(t-6)
Rise Time (sec) 3.4849 3.4512 3.448
Settling Time (sec) 3.3306 3.2476 3.2292

TABLE 17. z- Axis Control Signal Transient Response from Figure 28


Parameter PID RNN (t-1) EMRNN(t-6)
Rise Time (sec) 58.94 60.883 58.777
Settling Time (sec) 754.89 800.1 999.97

TABLE 18. IMU Sensor Noise Control Signal Average RMSE Results (25 sims, 1000 sec)
Parameter r- Axis (sec) φ- Axis (sec) z - Axis (sec)
PID 0.0096896 0.56912 0.0012739
PID+ RNN(t-1) 0.009862 0.60919 0.0012793
PID+EMRNN(t-6) 0.0093213 0.60823 0.0012075

52
In Figures 23 through 26 is observed that the PID-EMRNN controller achieves an

improved transient response over the PID and PID-RNN controllers despite the introduction of

the IMU sensor noise. The r-axis and z-axis have a significant improvement in position error

trajectory, while the position error trajectory in the φ-axis is marginally improved.

The control signals shown in Figures 27, 28, 29 and results shown Tables 13 through 15,

the r-axis the PID-EMRNN controller achieves a faster Rise Time than the PID and PID-RNN

controller by approximately 1.35 to 1.56 seconds. In the φ- axis, the PID-EMRNN controller is

0.04 seconds faster than the PID controller. In the z-axis the PID-EMRNN controller is 0.16

seconds faster than the PID controller and 2 seconds faster than the PID-RNN. In conclusion, all

control architectures Rise Time performances are similar. There is an improvement of the Rise

Time behavior from the application of the PID-EMRNN controllers over the PID and PI_RNN

controller.

The results also show the control signal Settling Time for the PID-EMRNN controller

achieves a significantly faster settling time than the PID and PID-RNN controllers. In the x-axis,

the EMRNN is approximately 50 seconds faster, in the φ-axis is faster than the PID controller by

approximately 100 seconds and 0.02 seconds marginally faster than the PID-RNN. In the z-axis,

the settling time is nearly 250 seconds slower than the PID-RNN controller.

The analysis above shows that the addition of an IMU sensor noise negatively impacts

the response of the PID-RNN when compared to the results obtained when only space

disturbances are considered. Based on the simulation results the PID-EMRNN is able to generate

the control signal that is able to yield faster rise and settling time, with the exception of the

slower settling time in the z-axis as shown in Tables 13 through 15.

53
The PID-EMRNN yields a faster transient response in the r-axis and z- axis, with a

negligible impact to the φ- axis control performance. These results show the ability of the

proposed adaptive algorithm to effectively modify the controller gains in each axis in a more

time efficient way than a PID controller for the given set of initial conditions and controller

parameters.

4.6 PID-EMRNN Adaptive Control at Various Conditions

The PID-EMRNN, PID-RNN adaptive control architectures and PID controller were

simulated for a length of time of 500 seconds with all three controllers experiencing space

disturbances and IMU sensor noise. The simulations were performed at altitudes of 400 km, 747

km, and 1,000 km, and at 30 degrees, 45 degrees, and 60 degrees inclination angles. During the

simulations the Right Angle of Ascension Node (RAAN) angles remained constant at 60

degrees. As seen in Figure 29, the PID-EMRNN is able to maintain the mean rise time shorter

than the PID-RNN and PID in the range described above for the φ and z- axes. The mean rise

time for the r-axis is marginally higher than both the PID and PID-RNN controllers.

In Figure 30, the mean settling time values for the PID-EMRNN controller are longer

than the PID and PID-RNN in every axis. Inspecting the numerical results it is evident that the

PID-EMRNN controller is highly susceptible to the change in inclination angle, specifically at 30

and 60 degrees. The longer settling time in the simulated range may be attributed to the lack of

controllability of the EMRNN at the given inclination angles and altitudes. The EMRNN

learning rates will need to be optimized to reduce the settling time over the specified range.

Since currently there is no algorithm that calculates an optimum learning rate, the learning rate is

calculated by trial and error. Methods such as adaptive learning rates [34] or newly proposed

Cyclical Learning Rates for Training Neural Networks [35] can be utilize in learning rates

54
optimization efforts. In the two proposed methods, the author postulates that a good learning rate

can be estimated by training the NN model with initially very low learning rate and linearly or

exponentially increase it at each iteration will lead to an optimized EMRNN learning rate.

FIGURE 30. Rise time mean values for range of 30, 45 and 60 deg. inclination.

FIGURE 31. Settling time mean values for range of 30, 45 and 60 deg. inclination.

55
CHAPTER 5

CONCLUSION

In this thesis, the application of an adaptive PID- EMRNN to a satellite formation flying

problem is proposed. In order to evaluate the performance of a PID-EMRNN control

architecture, the position error and control signal transient response is compared to the PID and

PID-RNN controller performance.

The Clohessy-Wiltshire (CW) dynamic model for spacecraft formation flying is

simulated. Though limited to circular orbits, the results are helpful in understanding formation

flying dynamic behavior and setting the baseline for the later derived elliptical orbit dynamic

model. The CW dynamic model results is utilized to validate the elliptical orbit dynamic model

for the circular orbit case. This elliptical orbit dynamic model is utilized as the plant in the

proposed control architectures.

The theory of RNN is introduced, algorithms defined and validated. Simulations are

performed for two different architectures of the LMS Gradient Descent algorithm. The LMS

algorithm RNN-1B is selected based on lower RMSE simulation results. The performance

results of the EMRNN of six time (𝑡 − 6) back steps show that it is adequate for

implementation. The PID control theory and proposed adaptive control architectures PID, PID-

RNN, and PID-EMRNN are simulated and results obtained.

Simulation results show that the adaptive PID-EMRNN controller with space

disturbances and IMU sensor noise have faster control signal transient response than the PID and

the PID-RNN controller in the r-axis, φ-axis and z-axis, with exception of the z–axis where it

shows slower settling response. This faster response can be attributed to the fact that the

EMRNN is able to use its history. Similar results are obtained in the case in which IMU sensor

56
noise is introduced into the system. The sensor noise negatively affects the performance of the

PID-EMRNN controller, since the EMRNN takes a considerably longer computational delay to

approximate the random sensor noise. The improvement margins achieved when considering

only space disturbances are reduced to half when compared to the case the performance achieved

when the IMU sensor noise is considered.

In conclusion, the results have shown that the PID-EMRNN adaptive controller yields a

more optimal results in transient position error and transient response of control signal effort for

the proposed satellite formation flying in at least two of the three axes in an ECI coordinate

system. If further researched and implemented the PID-EMRNN has the potential to optimize a

given formation flying trajectory.

5.1 Future Work

There are areas in which future work will enhance the solution to the SFF problem with

an adaptive PID-EMRNN controller. Future work should include exploring the adaptation of the

well-known Linear Quadratic Regulator (LQR) controller to an EMRNN and compare results

obtained here. Another area that could further enhance the results obtained is to add more

neurons to the EMRNN hidden layer and observe the settling and rise time results. The addition

of more neurons will provide the ability to estimate controller output with more precision, as the

number of activation functions would increase.

Lastly, this thesis will benefit from the addition of rotational dynamics to the problem

formulation. By adding attitude control for roll, pitch and yaw to the model, it will add fidelity to

the controller and provide an opportunity to address on-orbit difficulties in areas such as

precision pointing, laser communication and satellite robotics.

57
APPENDIX: SIMULINK BLOCK DIAGRAMS FOR RNN, GRAVITATIONAL AND

ATMOSPHERIC DISTURBANCE MODELING

58
FIGURE 32. RNN system Simulink block diagram for r, ϕ, z-axis.

FIGURE 33. RNN simulink block diagram for r, ϕ, z-axis 10 neurons.

59
FIGURE 34. RNN neuron simulink block diagram for r, ϕ, z-axis.

FIGURE 35. Atmospheric disturbance simulink block diagram.

FIGURE 36. Gravitational disturbance simulink block diagram.

60
FIGURE 37. Gravitational disturbance simulink subsystem block diagram.

61
BIBLIOGRAPHY

62
BIBLIOGRAPHY

[1] Bandyopadhyay, S., Subramanian, G.P., Foust, R., Morgan, D., and Hadaegh, F.Y., 2015,
"A Review of Impending Small Satellite Formation Flying Missions," Proceedings of
the 53rd AIAA Aerospace Sciences Meeting. Doi: 10.2514/MASM15.

[2] Sweeting, M.N., 2018, "Modern Small Satellites: Changing the Economics of Space,"
Proceedings of the 106th IEEE 106(3):323-361. Doi:10.1109/JPROC.2018.2806218.

[3] Hadaegh, F.Y., Singh, G., Acikmese B., Scharf D.P., and Mandic, M., 2008, "Guidance
and Control of Formation Flying Spacecraft", accessed October 31, 2018.
https://doi.org/ 10.1177/0278364907076709.

[4] Ioannou, P.A., and Sun, J., 2003, Robust Adaptive Control, Prentice Hall, Los Angeles.

[5] Kristiansen, R., Grotli, E.I., Niklasson, P., and Gravdahl, J.T., 2007, "A Model of Relative
Translation and Rotation Leader-Follower Spacecraft Formation," Modeling,
Identification and Control, 28, pp. 3-13. Doi: 10.4173/mic.2007.1.1.

[6] Vallado, D.A., 2013, "Clohessy Wiltshire Equations," Fundamentals of Astrodynamics and
Application, Microcosm Press, Hawthorne, CA, pp. 393–397.

[7] Kristiansen, R., and Niklasson, P., 2009, "Spacecraft Formation Flying: A Review and
New Results on State Feedback Control," Acta Astronautica, 65, pp. 1537-1552. Doi:
10.1016/j.actaastro.2009.04.014.

[8] Wu, A., 2013, "Daily-Repeat Stereo Monitoring from Formation Flying," Acta
Astronautica, 82, pp. 118-123. Doi: 10.1016/j.actaastro.2012.04.021.

[9] "ADCS for Beginners," accessed October 1,2018,


https://adcsforbeginners.wordpress.com/tag/ perifocal-frame/.

[10] Vallado, D.A., 2013, "Relative Motion," Fundamentals of Astrodynamics and Application,
Microcosm Press, Hawthorne, CA, pp. 388-419.

[11] Grotli, E.I., 2005, "Modeling and Control of Formation Flying Satellites in 6 DOF,"
master's thesis, Norwegian University of Science and Technology, Trondheim.

[12] Li, J., 2012, "Intelligent Control of Satellite Formation Control," Ph.D. dissertation,
Ryerson University, Toronto, Canada.

[13] Desai, U.P., 2012, "A Comparative Study of Estimation Models for Satellite Relative
Motion," master's thesis, Texas A & M University, College Station.

63
[14] Vallado, D.A., 2013, "Atmospheric Drag," Fundamentals of Astrodynamics and
Applications, Microcosm Press, Hawthorne, CA, pp. 551-567.

[15] Wertz, J.R., Everett, D.F., and Puschell, J.J., 2015, Space Mission Engineering: The New
SMAD, Microcosm Press, Hawthorne, CA.

[16] Sidi, M. J., 2006, "Effect of Noise and Disturbances on ACS Accuracy," Spacecraft
Dynamics and Control, Cambridge University Press, Cambridge, UK, pp. 172-175.

[17] Sclafani, R. J., 2014, "Variable Memory Recurrent Neural Network for Nano Sat Launch
Vehicle Attitude Control," master's thesis, California State University, Long Beach.

[18] Hyung-Shul Lim, H. B., 2009, "Adaptive Control for Satellite Formation Flying Under
Thrust Misalignment," Acta Astronautica, 65, pp. 112-122.

[19] Dong, X.G., Cao, X.B., Zhang, J.X., and Shi, L., 2013, "A Robust Adaptive Control Law
for Satellite Formation Flying" Acta Automatica Sinica, 39(2), pp. 128-137.

[20] Gurfil, M.P., Idan, M., and Kasdin N.J., 2003, "Adaptive Neural Control of Deep Space
Formation Flying," Journal of Guidance, Control and Dynamics, 26(3), pp. 491-501.

[21] Bae, Y.J., and Kim, Y., 2012, "Adaptive Controller Design for Spacecraft Formation
Flying Using Sliding Mode Controller and Neural Networks," Journal of the Franklin
Institute, 349(2), pp. 578-603.

[22] Joshi, G., and Padhi, R., 2014, "Robust Satellite Formation Flying Through Online
Trajectory Optimization Using LQR and Neural Networks," IFAC Proceedings
Volumes: Third International Conference on Advances in Control and Optimization of
Dynamical Systems, 47(1), pp. 135-141. Doi: 10.3182/20140313-3-IN-3024.00173.

[23] Shankar, P., 2007, "Self-Organizing Radial Basis Function Networks for Adaptive Flight
Control and Aircraft Engine State Estimation," Ph.D. dissertation, Ohio State
University, Ohio.

[24] Burken, J.J., Hanson, C.E., and Lee, J.A., 2009, "Flight Test Comparison of Different
Adaptive Augmentations of Fault Tolerant Control Laws for a Modified F-15 Aircraft,"
Proceedings of the AIAA Infotech Aerospace Conference, Seattle, WA, April 6-9,
2009, AIAA 2009-2056. Doi: https://doi.org/10.2514/6.2009-2056.

[25] Aksakalli, V., and Ursu, D., 2006, "Control of Nonlinear Stochastic Systems," Proceedings
of the 45th IEEE Conference on Decision and Control, San Diego, CA, pp.4145-4150.
Doi: http://dx.doi.org/10.1007/978-3-642-03479-4_10.

64
[26] Hunt, K.J., Irwin, G.R., and Warwick, K., 1995, "A Radial Basis Function Network Model
for the Adaptive Control of Drying Oven Temperature," Neural Network Engineering
in Dynamic Control Systems, Springer-Verlag, London, pp. 239-259.

[27] Donge, N., 2018, "Towards Data Science," Doi: https://towardsdatascience.com/recurrent-


neural-networks-and-lstm-4b601dd822a5.

[28] Haykin, S., 1999, "Least Mean Square Algorithm," Neural Networks: A Comprenhensive
Foundation, Prentice Hall, New Jersey, pp. 128-133.

[29] Medsker, L.R., and Jain, L.C., 2000, "Gradient Descent and Conjugate Gradient Descent,"
Recurrent Neural Networks Design and Applications, CRC Press, Washington, DC, pp.
253-259.

[30] Haykin, S., 1999, "Neural Networks: A Comprenhensive Foundation," Tom Robins, New
Jersey, pp. 299.

[31] Medsker, L.R., and Jain, L.C., 2000, "Comparison of Recurrent Networks for Trajectory
Generation," Recurrent Neural Networks Design and Applications, CRC Press,
Washington, DC, pp. 258.

[32] Help, M., 2018, "PID Tuning Algorithm," MATLAB.

[33] El-Shaboury, S.M., Ammar, M.K., and Yousef, W.M., 2017, "Analytical Solution of The
Relative Orbital Motion in Unperturbed Elliptical Orbits Using Laplace
Transformation," IRJET, 4(1), pp. 624-630.

[34] Duchi, J., Hazan, E., and Singer, Y., 2011, "Adaptive Subgradient Methods for Online
Learning and Stochastic Optimization," The Journal of Machine Learning Research, 12,
pp. 2121-2159.

[35] Smith, L. N., 2017, "Cyclical Learning Rates for Training Neural Networks," Proceedings
of the WACV 2017: IEEE Winter Conference on Applications of Computer Vision,
Santa Rosa, CA, March 24-31, 2017. Doi: 10.1109/WACV.2017.58.

[36] Ogata, K., 2010, Modern Control Engineering, Prentice Hall, New Jersey.

[37] Hunt, K.J., Irwin, G.R., and Warwick, K., 1995, "Recurrent Networks," Neural Network
Applications in Control, The Institution of Electrical Engineers, Herts, UK, pp. 46-59.

[38] Doreswamy, C. M. V., 2013, "Performance Analysis of Neural Network Models for
Oxazolines and Oxazoles Derivatives Descriptor Dataset," International Journal of
Information Siences and Techniques, 3(6), pp.1-15.

65
[39] Sharma, R., Kumar,V., Gaur, P., and Mittal, A.P., 2016, "An Adaptive PID Like Controller
Using Mix Locally RNN for Robotic Manipulator with Variable Payload," ISA
Transactions, 62, pp. 258-267. Doi: 10.1016/j.isatra.2016.01.016.

[40] McCamish, S.B., Romano, M., Nolet, S., Edwards, C.M., and Miller, D.W., 2009, "Ground
and Flight Testing of Multiple Spacecraft Control on SPHERES During Close
Proximity Operations," Journal of Spacecrafts and Rockets, 19. Doi:
10.1016/j.actaastro.2009.03.039.

66

You might also like