This action might not be possible to undo. Are you sure you want to continue?

A.J.M. Raemaekers DCT 2007.141

Traineeship report Coach(es): Supervisor: dr. R. Hill (RMIT) prof. dr. L. Wang (RMIT) ass. prof. dr. C. Bil (RMIT) prof. dr. H. Nijmeijer (TU/e) Royal Melbourne Institute of Technology Aerospace Design Technology Sir Lawrence Wackett Centre

Technische Universiteit Eindhoven Department Mechanical Engineering Dynamics and Control Group Eindhoven, December, 2007

Abstract In this report a continuous time model predictive controller is applied to the control of an Unmanned Aerial Vehicle (UAV). The purpose of this work, is to see if it is possible to design a good tracking controller that is capable of simulating a highly nonlinear aircraft system. For the simulations a MATLAB Simulink model of the UAV 'Ariel' is used. The nonlinear Simulink model covers all the aircraft dynamics. The model consists of ve inputs, twelve states and twelve outputs. This model is linearized at several dierent operating conditions, for use with the model predictive controller. The model predictive controller is capable of simulating in real time, making it possible to use this in on-line applications. The Ariel model is linearized to a linear time invariant state space system. The state space model is then extended. This means the outputs of the original model are included in the state vector of the extended model. With this extended state space model, it can be shown that the selected outputs can be tracked by the controller. The controller uses an algorithm which reduces the required computational power compared to classic model predictive control. The possibility to apply constraints on the controller input, the derivative of the controller input, the states and the outputs of the system, make this model predictive controller a nonlinear system. Because in real life constraints are almost always present on any dynamical system, the presented model predictive controller has a broad application scope. The simulations show that the controller performs well, after tuning of the parameters. Applying the constraints puts some limits on the possible reference trajectories, but shows that the system is very capable of handling these controls. Also, simulation of failure of one of the controls is possible, simply by setting the maximum and minimum values to zero. The UAV controller is able to use the other three controls to compensate for the lack of the fourth, keeping the aircraft at steady ight. The application of model predictive control to aircraft control is relatively new. This report shows that using a model predictive controller works well for the simulations done in this report. This method of MPC can be used not only on aircraft systems, but in general on all types of dynamic systems, with the possibility of applying constraints. For future work, a switching controller can be designed to use the dierent linearized models at dierent operating conditions. Wrapping another controller around the model predictive controller presented here, should make it possible to let the aircraft follow a set of way-points, thus not requiring a pre-dened reference trajectory to be able to y fully autonomously.

Contents

1 2 Introduction UAV and ight background 2.1 UAV development . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Aircraft control history . . . . . . . . . . . . . . . . . . . . . . . . . . Flight dynamics 3.1 Coordinate system . . . . . . . . . . . . . . 3.1.1 Body-xed coordinate system . . . . 3.1.2 Earth-xed coordinate system . . . . 3.1.3 Atmosphere-xed coordinate system 3.2 Equations of motion . . . . . . . . . . . . . 3.2.1 Aircraft controls . . . . . . . . . . . 3.3 Aircraft trim . . . . . . . . . . . . . . . . . 3.4 Ariel model . . . . . . . . . . . . . . . . . . 3.5 Linearizing the equations of motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 5 5 5 7 7 8 8 8 8 10 10 11 12 14 14 14 15 16 16 17 19 21 24 24 25 27 29 29 31 35 37 37 38 39

3

4

Model Predictive Control 4.1 Model Predictive Control . . . . . . . . . . . . . . . . 4.1.1 Receding horizon concept . . . . . . . . . . . . 4.1.2 Optimal control formulation . . . . . . . . . . . 4.2 Continuous time model predictive control design using functions . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.1 Introduction . . . . . . . . . . . . . . . . . . . 4.2.2 Description of control trajectory . . . . . . . . 4.2.3 Predicted plant output . . . . . . . . . . . . . . 4.2.4 Optimal control strategy . . . . . . . . . . . . . Simulations and results 5.1 Model selection . . . . . . . . . . . . . 5.2 Output selection . . . . . . . . . . . . 5.3 Tuning . . . . . . . . . . . . . . . . . . 5.4 Simulations . . . . . . . . . . . . . . . 5.4.1 Simulation without constraints 5.4.2 Simulation with constraints . . 5.4.3 Simulation of failure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . orthonormal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

5

6

Conclusions and recommendations 6.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2 Recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . . .

A Time dependent A matrix

1

B LTI Ariel model C Controls and states scenario 1 and 2 D Trim conditions

40 42 45

2

Chapter 1

Introduction

In recent years, model predictive control (MPC) has been an upcoming control technique in on-line applications due to the ever increasing computer power of modern computers. MPC is mostly used to control large time scale factory processes, but in recent years the application has moved to dynamic processes with smaller time scales. Most existing MPC implementations use a discrete-time form; the continuous-time counterpart was developed much later. Continuous time modeling gives the advantage that future predictions can be calculated in an analytical form. This reduces the required computational eort because there are less parameters that need to be optimized. The response time of the control signal in the continuous time MPC case is a bit slower compared to the discrete time case [1]. This means the performance is slightly worse, because it takes more time before the system is stabilized. However, continuous time MPC requires less optimization steps to calculate the control signal. It also makes the tuning of the controller more intuitive, and the calculated control signal smoother. One of the latest developments in continuous time MPC is the use of orthonormal functions to describe the control signal [1]. In this report the application of this method to the control of a Unmanned Aerial Vehicle (UAV) will be discussed. In recent years, unmanned aircraft have become more and more popular. The military has great interest in these machines, for amongst other things, battleeld surveillance, reconnaissance and targeting. They can also be used for civilian applications such as reconnaissance, crop dusting, observations, police surveillance, TV broadcasting, photography etc. The fact that they are unmanned make UAVs particularly suitable for going into inhospitable or hazardous environments. In general, UAVs have the potential to complement traditional manned surveillance and monitoring operations or, for some missions, can be a cost eective alternative. In the beginning of the last century the research was mainly focused on auto pilot systems, but this knowledge was restricted for the public for military security. Controllers designed before the development of classical control theory, were simple altitude and heading holds. But with the up-come of classical control theory and the introduction of the jet engine more and more controllers were applied to the aircraft to assist the pilot in ying the aircraft. Therefore, control of aircraft has been studied for a long time, ever since the development of modern computers in the 1950's. 3

The fth chapter contains the simulations done to show how the implementation of the model predictive controller is able to control the Ariel UAV and also how it is able to deal with constraints on and failure of one or more of the control inputs. the required computation eort decreases. controls and outputs of the linear state space model. MPC has been applied to UAV control before [3]. by using orthonormal functions to approximate the control signal and calculate the UAV response. needed to track the selected outputs of the Ariel aircraft model. The third chapter gives the reader some insight in ight dynamics and the equations of motion describing the states of an aircraft. dierent methods have been studied over the years. The model predictive controller makes it possible to add hard constraints to the states. The fourth chapter will familiarize the reader with model predictive control. The UAV 'Ariel' used in this report was modeled by Crump [2] with the MATLAB Simulink software. because. in the second chapter the background of UAV and ight control is described. 4 .To control UAVs. The controller calculates the system controls. It is also possible to simulate failure of one or more of the controls by setting the hard constraints to zero. The linearized model is used in the model predictive controller. Finally in chapter six the conclusions and recommendations are dealt with. This report is organized as follows. with different results. but this method diers from the previous work done. It also describes the linearization of the (non linear) equations of motion and how the steady state operating conditions are determined. This Simulink model is linearized at dierent operating points. and deals with the use of orthonormal functions to describe the control trajectory in more detail. making it possible to use this version of MPC to control the Ariel model in an on-line environment. In this report a continuous time model predictive controller with orthonormal functions will be designed to control a UAV.

These 5 . in these jet aircraft. articial damping had to be provided to allow satisfactory handling qualities. used for meteorological surveying and the Yamaha AeroRobot. the emphasis has moved from Remotely Piloted Vehicles (RPVs) to UAVs. 2. Another review of UAV activity can be found in [5]. They conclude that Australia has both a very large potential market for UAV applications and also a mature capability for advanced aircraft design. The development of a multi-purpose Australian UAV is discussed in [8]. which could be applied very successfully to a `small' design project such as a UAV. addressing some of the issues that must be considered in design for the Australian environment. an unmanned helicopter used for crop-dusting. A study of the market in Australia for both civilian and military uses is given in [6. where several specic aircraft platforms are discussed in reasonable depth. developed by General Atomics and produced in 1997 and the Northrop Grumman Global Hawk. 7].Chapter 2 UAV and ight background 2. The so called short period longitudinal mode and the dutch roll lateral mode fall in this category. The majority of UAV development has been military based.1 UAV development In the last few decades the use of Unmanned Aerial Vehicles (UAVs). A brief review of the last 20 years of UAV activity is given in [4]. whilst development within the Middle East. which has been developed as a high altitude. The autonomous capabilities of UAVs have improved over the last years due to improvements in (control) technology and improvements of computational power. Most well known are The Predator (a medium sized tactical UAV). large area surveillance aircraft. which have a form of autonomy. the Soviet bloc and Asia is also well underway. has become more and more popular. These documents only discuss the developments in Western countries. All reports indicate the increasing prominence which UAVs are taking in present day armed forces. The two most successful civilian UAV are the Aerosonde aircraft. Thus. Underdamped high frequency modes which became increasingly dicult for human pilots to control became evident. But there have been several civilian forays into the eld as well.2 Aircraft control history The introduction of the jet aircraft resulted in more complicated aircraft dynamics. Due to the more advanced technology available nowadays.

commonly known today as the Kalman lter. attempts to automatically synthesize an appropriate Kalman lter given a regulator or viceversa. This method does not have the guaranteed properties of each individual system on its own. The main point of Kalman's work is that it is a time domain approach that uses linear algebra. whilst the second more important problem is that of robustness. the application of model predictive control (MPC). This technique will generally almost fully recover the desired properties of the full state feedback regulator. 6 . because MPC requires a lot of computational power. with computing power still increasing every day. e. However. These techniques all use the frequency domain as a design region and are widely accepted. in the control of process industry. These classical techniques proved extremely successful for the aircraft of the time. it was possible to solve his matrix equations. These new control techniques have two problems compared to classical techniques. Examples of the use of this technique are given in [9]. and more recently.articial damping systems were developed using the emerging classical design techniques which are mainly based on working with single input single output (SISO) systems and using tools such as Bode plots. These papers contained detailed development of the optimal control of systems. The rst problem is the lack of engineering intuition in control design. these techniques became increasingly dicult to apply. implying that one is concerned with the internal system behavior as well as the system input/output behavior. Kalman published three major papers that were the beginning of the modern control era. only because of the development of digital computer technology. These problems lead to the introduction of robust modern control theory. In 1960. with all problems simply solved through the solution of matrix equations.g. the well known phase and gain margin concepts can not be easily applied to the multivariable systems that modern control is so suited for. Also. This method utilizes both the linear quadratic regulator (LQR) which requires full state feedback and the Kalman lter to provide a state estimate based upon system outputs. they have some well known limitations. The main design techniques are all based upon linear time invariant (LTI) SISO systems and with increasingly complicated aircraft and control objectives. utilizing phase margins and gain margins to illustrate robustness. However in recent years the application of MPC to systems with small time constants has increased and new applications arise every day. so the enhancement. which provides a unique solution with guaranteed performance. One method that has found a large amount of acceptance in the control industry is an extension of both linear quadratic regulators and Kalman lters to a technique known as Linear Quadratic Gaussian with Loop Transfer Recovery (LQG/LTR). disturbances and measurement noise are introduced. so systems with multiple inputs and multiple outputs (MIMO) can be handled with no extra eort. He also introduced the concept of the internal state. Nyquist plots and Evans Root Locus techniques. which are very time intensive and dicult to solve by hand. Kalman formalized the concept of `control optimality' by minimizing a quadratic energy function. MPC has been known and used since the 1970's and has been successfully applied to systems with large time constants. through the use of a Linear Quadratic Regulator (LQR) and the development of an optimal ltering technique. however. Loop Transfer Recovery (LTR). The guaranteed performance of modern control designs could quite easily be lost when unmodeled dynamics.

7 . 1. Atmosphere-xed coordinate system The body-xed frame can be expressed in terms of the Earth-xed frame by the use of Euler angles (φ.1 Coordinate system The movement of an airplane is described by the equations of motion with respect to the body frame coordinate system. In general there are three coordinate systems that are important when describing the motion of an airplane [11]. The majority of data on this aircraft is available in [10].1. M. θ and ψ ). which is an unmanned aircraft developed by the UAV group in the Department of Aeronautical Engineering at the University of Sydney.1: Ariel UAV 3. Body-xed coordinate system 2. Figure 3. A picture of this aircraft is shown in gure 3.Chapter 3 Flight dynamics In this report a continuous time model predictive controller is designed with orthonormal functions to control a UAV model. Earth-xed coordinate system 3. Crump has developed an accurate MATLAB Simulink model of the UAV Ariel [2].

The y -axis is normal to the plane of symmetry pointing in the direction of the right wing. Rotation about the z -axis.3 Atmosphere-xed coordinate system The atmosphere-xed coordinate system is dened such that all three axes are always parallel to those of the Earth-xed coordinate system. The atmosphere coordinate system is only used when deriving the equations of motion. The z -axis points downward. The x-axis points forward along some axis of the fuselage in the aircraft's plane of symmetry. the aircraft is considered as a rigid body. The x. completing the right-handed Cartesian system. pitch (θ) and yaw (ψ ) respectively. East and Down (NED frame).1. (a) Body CS (b) Rotations Figure 3. a matrix multiplication is used.2(a). However. the curvature of the Earth can be neglected for simplication. It is also used in wind tunnel testing of aircraft models. For navigation on a small piece of the Earth. To switch between the ABC and NED coordinate frame. See also gure 3.2 Earth-xed coordinate system The Earth-xed coordinate system points in the directions North. consisting of three rotations: 1. The NED frame is then xed and does not rotate. the atmospheric coordinate system moves at a constant velocity relative to the Earth-xed coordinate system.1 Body-xed coordinate system This coordinate system's origin is located at the aircraft center of gravity. Rotations about the x. This simplication is also called the 'at Earth' approximation. y and z -axis are called roll (φ). Rotation about the new y -axis.3. See also gure 3. positive nose up (pitch θ) 8 . y -plane is normal to the local gravitational vector with the x-axis pointing north and the y -axis pointing east. The Body-xed coordinate system is also called the ABC frame. 3. 3.1. positive nose right (yaw ψ ) 2.2(b). The z -axis points down.2 Equations of motion When deriving the equations of motion.1.2: Coordinate system 3.

3. Rotation about the new x-axis. the equations of motion of the aircraft can be described by: ˙ U ˙ V ˙ W ˙ φ ˙ θ ˙ ψ ˙ P = RV − QW − g sin θ + Fx /m = −RU + P W + g sin φ cos θ + Fy /m = QU − P V + g cos φ cos θ + Fz /m = P + tan θ(Q sin θ − R cos θ) = Q cos θ − R sin φ = = (Q sin φ + R cos φ)/ cos θ (c1 R + c2 P )Q + c3 L + c4 N ˙ Q = c5 P R − c6 (P 2 − R2 ) + c7 M ˙ R p˙ N p˙ E p˙ D = (c8 P − c2 R)Q + c4 L + c9 N = U cos θ cos ψ + V (sin φ sin θ cos ψ − cos φ sin ψ) + W (sin φ sin θ cos ψ + sin φ sin ψ) = U cos θ sin ψ + V (sin φ sin θ sin ψ + cos φ cos ψ) + W (cos φ sin θ sin ψ − sin φ cos ψ) = −U sin θ + V sin φ cos θ + W cos φ cos θ (3.4) With the denition of the state vector as: X= U V W φ θ ψ 9 P Q R pN pE pD T (3. pitch and bank angles are known as the Euler angles. positive right wing down (bank φ) The yaw.1) z ABC z N ED with cos θ cos ψ B = − cos φ sin ψ + sin φ sin θ cos ψ sin φ sin ψ + cos φ sin θ cos ψ cos θ sin ψ cos φ cos ψ + sin φ sin θ sin ψ − sin φ cos ψ + cos φ sin θ sin ψ − sin θ sin φ cos θ cos φ cos θ The inertia properties of the rigid airplane body can be dened by: c1 = c5 = c9 = 2 (Jy −Jz )Jz −Jxz τ Jz −Jx Jy Jx τ c2 = x yτ z c6 = Jxz Jy 2 τ = Jx Jz − Jxz (J −J +J )Jxz c3 = c7 = Jz τ 1 Jy c4 = c8 = Jxz τ 2 Jx (Jx −Jy )+Jxz τ (3. The transformation matrix B is given as: x x y =B y (3.5) .3) TB = M Fz N With these denitions.2) And the aerodynamic and propulsive forces and torques will be split into the body axes: Fx L FB = Fy (3.

3: Aircraft controls 3. and this can be used when navigation is only required over a small area of the Earth. pE and pD the positions with respect to the NED frame. which are shown in gure 3.1 Aircraft controls The Ariel UAV can be controlled by four deective surfaces and the throttle. 3. θ and ψ the three Euler angles • P . The controls and their eect on the aircraft are: 1. Elevator: Change the pitch (θ) of the aircraft 3. Rudder: Change the yaw (ψ ) of the aircraft. Flaps: Increase the lift of the wings. 5. The model is linearized using the so called 'at Earth' approximation. Q and R the angular rates • pN . Figure 3. The aircraft is 'tuned' in a way that it conducts a steady ight and no torque is applied to the 10 . Used for take-o and landing. Throttle: Accelerates/deccelerates the aircraft 2.with • U .2. V and W the velocity with respect to the body axes x.3 Aircraft trim Aircraft trim is the denition of the aircraft ying at steady state. Aileron: Change the banking angle (φ) of the aircraft. y and z respectively • φ.3. In this approximation it is assumed that the Earth is at. 4. This means that all the aircraft velocity variables are xed at a constant value.

1. elevators. rudder and aps. The Ariel model is linearized around dierent trim conditions. Q and R the angular rates • φ. throttle. W . ψ. Each trim condition is a steady state equilibrium. The values and rates of these inputs are limited. the trim condition also requires: ˙ ˙ ˙ φ.6) Steady wings-level ight and steady turning ight is allowed using the 'at Earth' simplication discussed in section 3.7) With these constraints an optimum trim condition can be found using an optimization algorithm. In the work of Crump [2] an optimization algorithm has been provided and is used to obtain trim conditions for the state. The model uses lookup tables to nd values for the aerodynamic and propulsive forces and aircraft parameters for dierent operating conditions. the state space system obtained for dierent operating conditions is stable.4 Ariel model The Ariel model is built in the MATLAB Simulink environment. ailerons. The 6-degree of freedom dynamic model is complemented by modeling the engine. More details about the Ariel Simulink model can be found in [2]. The model has ve inputs. The states of the equations of motion (3. Q. R. a good objective function to optimize is to minimize the sum of the square of the aircraft body accelerations. Ignoring the eects of changing density with altitude. therefore. R ≡ 0. U >0 (3.2. y and z direction with respect to the ABC-frame √ • VT the total velocity.airplane's center of gravity. steady wings-level climb and steady climbing turn are also trim conditions. 3. This implies that the aircraft acceleration components are all zero. Q. input and output variables for dierent operating conditions. P . for a steady wings level ight. V . In the simulations. sensor and actuator dynamics. the deviations around the steady states of the dierent state space models have to be added to the values of the trim conditions to provide the 'real' values that are used as input for the non-linear Ariel model. The measured outputs are: Y = with nx ny nz VT α β P Q R φ θ ψ T (3.8) • nx . acquired by windtunnel testing.5) are the internal states of the model. For instance. P. This denition leads to: ˙ ˙ ˙ ˙ ˙ ˙ U . VT = U 2 + V 2 + W 2 • α the angle of attack • β the sideslip angle • P . θ and ψ the Euler angles 11 . In general. For each of the dierent trim conditions extra constraints need to be applied. represented as δT δe δa δr δf respectively. ny and nz the accelerations in x. θ. φ ≡ 0 (3.

3) is applied. The matrix A(t) is specied in appendix A. t) u u g1 (x0 . t) ˙ ∂t δy(t) = C(t)δx(t) + D(t)δu(t) + ∂g (x0 . B(t). t) = ∂f (x0 . which are dened as follows: x f1 (x0 . . . . t0 ) = y0 . t) D(t) = . and D(t) of this linearized model represent the Jacobian matrices of the system.9) (3. Subtracting the operating point values from the states. t) u f1 (x0 . u0 . t) (3.12) The state-space matrices A(t). . t) ˙ y(t) = g (x(t). u0 . u0 .13) C(t) = . .. i. t) ∂t (3. . gp (x0 . u0 . the system matrix further 12 . . u0 . C(t). inputs. u(t) represents the m-inputs of the model. gp (x0 . A linearized model of this system is valid in a small region around the operating point t = t0 . u(t). t) . B(t). ∂f (x0 . C(t). t) where x f (x0 . and outputs denes a set of variables centered about the operating point: δx(t) = x(t) − x0 δy(t) = y(t) − y0 (3.t) ∂xn (t) (3. u0 . When the aircraft trim condition for steady wings-level ight (specied in section 3. t) u fn (x0 . and y(t) represents the p-outputs of the model. u(t). t) (3.u0 .t) ∂x1 (t) ∂f (x0 . x(t0 ) = x0 . u0 . u(t0 ) = u0 . A(t) = B(t) = .3. and D(t) are all time dependent. t) x x g1 (x0 .14) The obtained matrices A(t).5 Linearizing the equations of motion A general nonlinear system can be written in the following form: x(t) = f (x(t). u0 . x fn (x0 . x(t) represents the n-states of the model.u0 . u0 . when the departure from the operating point is small: δ x(t) = A(t)δx(t) + B(t)δu(t) + ∂f (x0 .. u0 .10) In these equations. u0 . Operating points can be found using the trim algorithm. u0 .e.t) ∂x2 (t) .11) δu(t) = u(t) − u0 The linearized model can be written in terms of these new variables and is usually valid when these variables are small. .u0 . and y(t0 ) = g(x0 .

01 −30 0 0 0 1 0.01 −30 0 0 −2.07c2 0 0 0 13 . In the steady wings-level ight case.07 0 0 0 A= 0 0 0 1 0 0 0 −0. The general equations of motion and the aircraft specic propulsive and aerodynamic force models are combined in the Simulink model of the Ariel UAV together with the sensor and actuator dynamics.81 0 0 9. C(t) and D(t) will also become time invariant.15) The other matrices B(t).07c8 0 0 −0. are used the state space system will remain time dependent. When the other trim conditions.05 0 30 0 0.3 there are four dierent trim conditions.02 0 −2. Because all the accelerations are zero and the external forces don't change over time. a numerical perturbation algorithm is used to linearize the system. ∂f ∂g ∂t (x0 .81 0 0 2. the state space system becomes a linear time invariant (LTI) state space system. The model only simplies to a LTI state space system when the trim condition for steady wings-level ight or steady wings-level climb is applied.07 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 1 0 −9. Using the time varying model in the model predictive controller will require more computations. only the LTI state space model is used. Both the perturbations and the responses are used to create the matrices of the linear state-space model. To obtain the Linear Time Invariant (LTI) model. t) ≡ 0 and these terms can thus be neglected.07c2 0 0 0 0 0 0 0 0.07c1 0 −0.05 0 0 0 −0.05 0 30 0 1 0 0 0 0 0 0 0 0.0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 (3. In appendix B an example of a state space model is given linearized around a trim condition. steady turning ight and steady climbing turn. and therefore makes it not (yet) possible to use this for real time control. simplies to: 0 0 0.01 0 0 0 1 0 0 0 0 0. In this research. As mentioned before in section 3. u0 .07 0 0 0 0 0 0 0. u0 . t) ≡ 0 and ∂t (x0 . This involves introducing small perturbations to the nonlinear model and measuring the responses to these perturbations.

general model predictive control will be discussed. every time p seconds ahead of the current time t. This method reduces the required computing power. Then the outputs are measured again (y + 1). and based on this output. a control eort is computed to reduce the system error. 14 . 4. constraints and control objectives. The online optimization problem takes account of system dynamics. Conventional model predictive control requires the solution of an open-loop optimal control problem. Based on the plants predicted output at time t + p. the use of orthonormal functions to design a continuous time model predictive controller will be discussed. the prediction horizon keeps shifting. This way. The main reason for the wide-scale adoption by industry of model predictive control is its ability to handle hard constraints on controls and states that arise in most applications.1 Receding horizon concept The receding horizon idea is the basic concept behind MPC [17]. model predictive control uses a discrete time setting. as long as they satisfy certain conditions (see 4. The rst step (t + 1) of the calculated control input is executed. in which the decision variable is a sequence of control actions. the new predicted plant states and outputs are calculated for the new prediction horizon t + 1 + p. This is done for the prediction horizon p. 4. at each sample time. Current control action is set equal to the rst term of the optimal control sequence.1 Model Predictive Control Model predictive control is an approach to controller design that involves on-line optimization calculations. Muske and Rawlings [15] and Rawlings et al. In this section a model predictive controller is presented which can be used to control all general types of state space models. including the papers of García et al. This principle is illustrated in gure 4. Several authors have published reviews of MPC theoretical issues. First. Ricker [13].3). Morari and Lee [14].1. In most cases. At time t the predicted plant output is calculated using the internal plant model. In chapter 5 the application and results of this design method to control the Ariel model will be described. thus giving it signicant advantages when used in an on-line environment. [16].2. Secondly. The big advantage of this method is that it reduces the required computing power.1.Chapter 4 Model Predictive Control In the work of Wang [1] it is shown how to create a continuous time model predictive controller using orthonormal functions. [12]. thus following the reference trajectory.

u. the following partial dierential equation needs to be solved: ∂ 0 V (x. t) ≥ 0. satisfying l(x(t).2) where l(x(t). F (x(p)) is the terminal state weighting (at the prediction horizon p) and the control input is constrained to be in some set u(t) U . This is a very general formulation. u. Figure 4. The control objective is to minimize a cost function. λ) = l(x. with the boundary condition V (x.1) The model outputs are calculated up to time p. t) is some function.3) Where H(x. u). which has the form: p V (x. u(t). u.In this research a continuous time MPC is used. u(t). and just about every sensible problem can be represented in this form by using suitable functions f . u(t). ∂ 0 V (x. F and l. This equation is called the Hamilton-Jacobian-Bellman equation and in 15 . u) + λf (x. p) = F (x(p)). t)) ∂x (4.2 Optimal control formulation The plant behavior can be described by: dx = f (x.1: Receding horizon control 4.1. t) = 0 l(x(t). t)dt + F (x(p)) (4. t) dt (4. To solve the minimization of the cost function. u. This means that a continuous time model is used to calculate the system behavior and the desired control trajectory. The execution of the algorithm does use a discrete time setting in the sense that for a predened time step the control eort is executed before a new optimization step is made. t) = minu ∂t U H(x.

11) 4. One of the crucial steps widely used in the derivation of discrete time MPC is to capture the projected control signal over a prediction horizon using a moving average structure. t) has the special form: x = A(t)x(t) + B(t)u(t) ˙ and that the functions l and F are quadratic functions: (4.7) Then.2 Continuous time model predictive control design using orthonormal functions 4. In general. A candidate function to solve the HJB equation is to maintain a quadratic form of the cost function at all time t. This technique can not be naturally adopted in continuous time MPC design where the corresponding components are the derivatives of the control signal over the prediction horizon. 20]. In that case the Hamilton-Jacobi-Bellman equation simplies to an ordinary dierential equation. continuous time MPC is more complicated than its discrete time counterpart. time-varying feedback law: u(t) = F (t)x(t) (4. R(t) and S are square. u(t). u. 19. positive-denite matrices. so that f (x. the Ricatti equation. This is because in the continuous time case. Assume that: V 0 (x. long term increments of the future control signal are assumed to go to zero.most cases it is virtually impossible to solve this partial dierential equation. P (t) = P T (t) (4.1 Introduction Most of the MPC techniques developed over the last years have focussed on discrete time systems.6) where Q(t).5) (4.9) This can be solved and leads to a linear. These can not be assumed to be zero without aecting closed-loop performance.10) where F (t) is the 'state feedback' matrix. It is necessary to make some assumptions to get solutions in practice. Based on a state space model. t) F (x(p)) = xT (t)Q(t)x(t) + uT (t)R(t)u(t) = x (p)Sx(p) T (4.2.4) l(x(t). F (t) = R−1 (t)B(t)T P (t) (4. the HJB equation is simplied to the following Riccati equation: ˙ −P (t) P (p) = = P (t)A(t) + A(t)T P (t) + Q(t) − P (t)B(t)R−1 (t)B(t)T P (t) (4. An emulator was proposed to obtain the predictions through a Taylor series expansion in [18. which depends on the solution of the Ricatti equation. which is computationally more demanding. t) = xT (t)P (t)x(t).8) S (4. One way is to assume the plant is linear. Gawthrop [21] proposed to use a set of exponential 16 . To make the structure nite dimensional. In recent years the interest in continuous time model predictive control has grown. the future plant behavior is calculated through convolution instead of direct iteration.

are the coecients and li (t). an arbitrary function f (t) is expressed in terms of a series expansion as ∞ f (t) = i=1 ξi li (t) (4. The novelty of this approach lies in the fact that the problem of nding the optimal control signal is converted into one of nding a set of coecients for the Laguerre model. . when using orthonormal functions. . This approach eectively reduces the number of parameters required in the operation. thus oering a substantial advantage when used in an on-line environment. The rst derivative of the control signal can be approximated within the prediction horizon using orthonormal basis functions [22]. when using a Laguerre network. The scaling factor is made a performance tuning parameter. . Future prediction is calculated in an analytical form. The accuracy of the approximation. is dependent on the choice of certain scaling factors. i = 1. . This technique is designed for continuous time. 2. The design technique presented here can be used as an alternative to classical receding horizon control without the need to solve the matrix dierential Ricatti equation. 4. The control trajectory is calculated using a pre-chosen set of orthonormal basis functions.15) In summary. . . there exists an integer N such that for all k ≥ N ∞ k (f (t) − 0 i=1 ξi li (t))2 dt ≤ (4. This parameter plays an important role in the response speed of the closed-loop system.14) then for any > 0 and 0 ≤ t ≤ ∞. but it can be readily extended to discrete time systems and it is expected to be applicable to non-linear systems as well [1].2.12) where ξi . a truncated expansion N i=1 ξi li (t) is used to closely approximate f (t).2 Description of control trajectory It is known that an arbitrary function f (t) has a formal expansion analogous to a Fourier expansion [23]. i = 1.functions to capture the control trajectory with the parameters related to the location of the closed-loop poles. Based on this technique. One set of orthonormal functions frequently used is the set of Laguerre functions. To simplify the performance specication procedure. are orthogonal functions satisfying ∞ ∞ li (t)2 dt = 1 0 0 li (t)lj (t)dt = 0 i=j (4.13) Further. assuming that f (t) is a piece-wise continuous function satisfying ∞ f (t)2 dt < ∞ 0 (4. For example. Laguerre orthonormal functions are used in this design. In his method the number of exponential functions matches the number of state variables. 2. 17 . the pole location is important.

. and within this window the control signal is denoted by u(t) and its derivative by u(t). the control signal in a stable closed-loop system does not converge to zero as t → ∞ if there exists a constant external signal such as a set-point signal. .. . . when the closed loop system is stable. Then the Laguerre functions satisfy the dierential equation: ˙ L(t) = Ap L(t) where (4.. consequently leading to the exponential convergence of the derivative.16) where p is positive and often called scaling factor. the projected control signal u(t) converges exponentially to a constant. −2p 0 −p . ˙ In general. Specically. . to zero. .. when it is assumed that the closed loop system is stable ˙ within each moving horizon window ti ≤ t ≤ ti + Tp . . lN (t)]T and L(0) = 2p[1 1 . the derivative of the control signal can be described by using a Laguerre function based series expansion as: N u(t) = ˙ i=1 ξi li (t) = L(t)T η (4. .. Suppose that for ti ≤ t ≤ ti + Tp . . .2 responds in a time invariant fashion for each moving horizon window. which is satised by the Laguerre functions.21) where η = [ξ1 ξ2 . the control signal for a set-point change or a load disturbance rejection exponentially converges to a constant after the transient response period.1. From (4.20) By applying the technique used by Wiener and Lee [22]. As a result the orthonormal functions ˙ 18 ... .16).19) For a linear time invariant system. . step input disturbance or step output disturbance. Thus the derivative of the control signal u(t) converges to zero. It is seen that: ti +Tp u(t)2 dt < ∞ ˙ ti (4. let √ L(t) = [l1 (t) l2 (t) . −2p 0 0 . ξN ]T is the vector of coecients of the orthogonal functions li (t) which express the series expansion of u(t). derived in section 4. u(t). −p (4. a dierential equation can be derived. 1]T . Using this in the receding horizon control design.Laguerre functions are appealing to engineers because the Laplace transform of li (t) has a particularly simple form: ∞ li (t)e−st dt = 0 2p (s − p)i−1 (s + p)i (4. .18) The solution of the dierential equation (4.17) leads to a representation of the Laguerre functions in terms of a matrix exponential function: L(t) = eAp t L(0) (4.. ˙ Remarks: • Assume that a moving time window is taken as ti ≤ t ≤ ti + Tp . the future control signal. .17) Ap = −p −2p .

24) where E{} denotes expectation and δ(. temperature change of the air. • Model predictive control is an open-loop optimisation strategy unless the state variable x(t) is measurable. 4.2. if there is an external constant input signal to the closed-loop control signal. Introduce the variable z(t) = xm (t) = ˙ ¨ d Am xm (t) + Bm xm (t) + ω(t) dt (4. the derivative ∞ of the control signal does preserves the property that 0 u(t)2 dt ≤ ∞.2. by introducing the controller. However. the open loop properties of the system are aected. This is a realistic situation as setpoint change and disturbance rejection are the primary roles in which a feedback control system plays. changes of the air density) and the measurements of the outputs are disturbed with µ(t): xm (t) ˙ y(t) = Am xm (t) + Bm u(t) + ω(t) = Cm xm (t) + Dm u(t) + µ(t) (4. not the control signal itself.can not be used to describe the control signal with the expectation of convergence of the expansion in a nite number of terms.23) E E dω(t) dω(τ )T dt dτ dµ(t) dµ(τ )T dt dτ = Wω δ(t − τ ) = Rµ δ(t − τ ) (4. ˙ In summary.22) where xm (t) is the state vector of dimension n.25) The state space system (4. Here it is assumed that ω(t) and µ(t) are continuous time white noise processes.26) X(t) = z(t) y(t) A= Am Cm 0 0 19 B= Bm Dm C = [0 I] . The actual feedback is introduced through the use of an observer to estimate the state variables (see section 4. it is assumed that the plant suers from external disturbances ω(t) (e. Also. changing wind.4).22) can now be written in an augmented form: ˙ X(t) y(t) where = AX(t) + B u(t) + ˙ = CX(t) ω(t) ˙ µ(t) ˙ (4.) is the Dirac function. Assuming that the closed loop system is stable is therefore questionable.3 Predicted plant output The plant to be controlled is a multivariable state space system with r inputs u(t) and q outputs y(t). It is assumed that E E and dω(t) dt dµ(t) dt = = 0 0 (4. the suitable candidate variable which can be modelled by a set of orthonormal functions is the derivative of the control signal. Furthermore.g.

Because of this. .26) takes the rst derivative of the control signal as its input and its output remains the same.29) = eAτ X(ti ) + 0 eA(τ −γ) B u(ti + γ)dγ ˙ where the expected eect of the random disturbances (ω(t). . .30) r i=1 T T T where the coecient vector η T = η1 η2 . and it is not necessary to calculate the derivative of ω(t) and µ(t) in equation (4. Then at the future time ti + τ . the state variable X(ti ) is available. which requires the solutions of (n + q) × i=1 Ni integral 20 . .28) (4.I is the q × q unit matrix.2. T Br Lr (γ)T ηdγ (4.26). Because the white noise has a zero average. . It can be shown that if the augmented state space system is observable and controllable. ur (t)]T ˙ ˙ ˙ ˙ and the input matrix B as: B = [B1 B2 . . It is assumed that at the current time ti . the expected eect of the random disturbances (ω(t). lNi (t) and ηi = ξ1 ξ2 . The control signal can be written as: u(t) = [u1 (t) u2 (t) . . . ξNi . ηr has dimension Accordingly.. . is T X(ti + τ ) = eAτ X(ti ) τ + 0 eA(τ −γ) B1 L1 (γ)T B2 L2 (γ)T . this only holds when there are no constraints active on the system. Simulations show that the presence of measurement noise and disturbances do not inuence the performance of the controller [1]. . X(ti + τ ). . . τ > 0. 2 . it is possible to track the outputs of the original system using a model predictive controller [26]. Note that the augmented state space description (4. y(ti + τ ) = CX(ti + τ ) (4. the plant output prediction can be represented by Ni .27) (4. µ(t)) in the future prediction is assumed to be zero. Of course. µ(t)) in the future pre˙ ˙ diction is zero. Br ] where Bi is the i-th column of the B matrix.2. the noise terms are neglected in this analytical approach. because an observer is used to estimate the state variables so that explicit dierentiation is avoided.30). The measurement noise ω(t) and the disturbances µ(t) are not amplied. pi and Ni are pre-chosen. Then the predicted future state at time ti + τ . the predicted state variable X(ti + τ ) is described by the following equation: ti +τ X(ti + τ ) = eAτ X(ti ) + ti τ eA(ti +τ −β) B u(β)dβ ˙ (4. r) is expressed as: ui (t) ∼ Li (t)T ηi ˙ = i i i i i i where Li (t)T = l1 (t) l2 (t) . The i-th control signal ui (t) (i = ˙ 1. The projected control trajectory u(t) will be parameterized using ˙ the orthonormal functions described in section 4.31) The major computation load in evaluating the prediction comes from the convolur tion operation in (4..

equations. Upon obtaining the matrices Iint (τ )i for i = 1. to the future trajectory of the setpoint r(ti + τ ).2.33) where L(τ )T .17). 4. The convolution integral corresponding to the i-th input is dened as: τ Iint (τ )i = 0 eA(τ −γ) Bi Li (γ)T dγ (4.30) shows that the prediction of the future trajectory can be expressed in terms of Iint (τ )i with 1 ≤ i ≤ r. The tuning of the MPC will be done by changing the parameters p (poles of the Laguerre functions) and N (number of orthonormal functions). . The common objective of model predictive control is to nd the control law that will drive the predicted plant output y(ti + τ ) as close as possible. . Substituting this in (4. . 2.34) ˙ ˙ where Q and R are symmetric matrices with Q > 0 and R ≥ 0. r. such a numerical approximation is unnecessary and analytical solutions can be found for the convolution integral. where Tp is the prediction horizon. (4. Q is set to the unit matrix I and R to zero. as outlined above.18) and (4. Ψ2 = 0 φ(τ )QCeAτ dτ and ¯ R = diag (λi INi ×Ni ) where λi are the eigenvalues of extended system matrix A and INi ×Ni is the Ni ×Ni unity matrix.32) where Iint (τ )i is a matrix with dimensionality of (n + q) × Ni . L(0)T and Ap are dened by equations (4. For a given τ . . Assuming further that the reference trajectory does not change within the prediction horizon Tp . Suppose that a set of future setpoints r(ti + τ ) = r1 (ti + τ ) r2 (ti + τ ) . the integral equations can be solved numerically using nite sum approximation. However. The cost function can be written as a function of η instead of y(ti + τ ). the quadratic cost function can be written as: Tp J = η T Πη − 2η T {Ψ1 r(ti ) − Ψ2 X(ti )} + 0 wT (ti + τ )Qw(ti + τ )dτ (4. rq (ti + τ ) .35) where Tp Π= 0 Tp ¯ φ(τ )Qφ(τ )T dτ + R Tp Ψ1 = 0 φ(τ )Qdτ .19). In [1] it is shown that the performance of the MPC is predominantly determined by tuning the values of p and N and tuning of Q and R is therefore not necessary. 0 ≤ τ ≤ Tp are available. the prediction X(ti +τ ) can be determined. . . 21 . With this the predicted plant output can be calculated using (4.4 Optimal control strategy To compute the optimal control a cost function has to be optimized. The equivalent cost function can be dened as: Tp Tp J= 0 [r(ti +τ )−y(ti +τ )]T Q[r(ti +τ )−y(ti +τ )]dτ + 0 u(τ )T Ru(τ )dτ (4. To simplify the tuning process. in a least squares sense.31). The matrix Iint (τ )i can be expressed as: AIint (τ ) − Iint (τ )AT = −BL(τ )T + eAτ BL(0)T p (4.

. . . K .37) 0 0 . .. .41) 22 .. with p and N suitably chosen.The minimum of (4. . . Because the cost function (4. . it is necessary to apply the constraints only to the initial ˙ stage of the prediction horizon. . 0 0 ulow (ti + τi ) ≤ η + u(ti − ∆t) . τi 0 Lk (γ)T dγ = A−1 (eAp τi − I)L(0) p T (4. The bounds on the derivative of the control can be dened as: L1 (τi )T 0 . τi T 0 0 . is then given by the least squares solution: η = Π−1 {Ψ1 r(ti ) − Ψ2 X(ti )} The derivative of the control input u(t) ˙ L1 (0)T 0 . . . can now be computed by: 0 0 −1 Π {Ψ1 r(ti ) − Ψ2 X(ti )} . T Lr (0) (4. . . . .. u(ti ) = ˙ . Constraints on the control signal. . ..and the rst derivative of the control variables. This feature could potentially reduce the number of constraints required. . .. 0 L2 (0)T . . 2.. to guarantee closed-loop stability for continuous time model predictive control [18]. which guarantees the exponential decay of u(ti + τ ). . . . r are the exponential functions. Integration gives the control law: t u(t) = 0 u (τ ) dτ ˙ (4. . .. .. .39) . . . the control. without hard constraints on the variables. . discretization of the trajectories is necessary. .36) (4.39) yields a set of linear inequality equations. it is relatively easy to put hard constraints on the output-.40) where u(ti − ∆t) is the previous control signal. . . 0 ulow (ti + τi ) ≤ ˙ ˙ η ≤ uhigh (ti + τi ) (4. denotes the set of future time instants at which we wish to impose the limits on u. 0 Lr (γ) dγ ≤ uhigh (ti + τi ) (4. .. . . . 0 0 . . . (4. Lr (τi )T where τi . Simple computation shows that. subject to a number of inequality constraints is a standard quadratic programming problem.. . . .38) The cost function can be expanded by adding weighting to the terminal state [1]. . .. k = 1. 0 0 τi 0 L2 (γ)T dγ . Since ˙ Lk (τ ). . i = 0. 1. . the output variables and the system states can be dened as follows: τi L1 (γ)T dγ 0 . .35) is a quadratic function. The optimization of the cost function. To form a set of inequality constraints for the output and control signals.35). 0 0 L2 (τi )T .

Here. decays exponentially at any desired rate. equation (4. in theory. D) is completely observable. However. Thus. B. Iint (τi )r η ≤ Xhigh (4..43) The whole procedure described above assumes that the states at time ti are known.40) yields a set of linear inequality constraints on the control signal. In general it is not possible to measure the state of a plant. the observer gain Job is often limited by the existence of measurement noise. 23 . Xlow ≤ eAτi X(ti ) + and Iint (τi )1 Iint (τi )2 . for a given set of time instants τi . X(t) = X(t) − X(t). The observer can be designed using a continuous time steady-state Kalman lter or using other techniques such as pole assignment. The iterative solution of the Riccati equation is not required in real-time. Provided that the system model (A. The observer gain is calculated o-line.42) ylow ≤ CeAτi X(ti ) + C Iint (τi )1 Iint (τi )2 . u(t) is deter˙ mined from the optimal solution of the model predictive control strategy.. The observer equation that is needed for the implementation of the continuous time MPC is given by: ˙ ˆ ˆ ˆ X(t) = AX(t) + B u(t) + Job (y(t) − C X(t)) ˙ (4. Job is calculated recursively using a Kalman lter.18). C. The standard approach is the use of an observer to estimate te state variables X(ti ) [24. Job can be chosen such that the ˜ ˆ error.. either using the analytical solution or a quadratic programming approach.. Iint (τi )r η ≤ yhigh (4. in practice.44) ˆ Here X(t) is the estimate of X(t) and Job is the observer gain. 25].where Ap is dened by equation (4.

Altitude [m] 100 100 1000 1000 5000 5000 Velocity [m/s] 30 55 30 55 30 55 Rank A 10 10 10 10 10 10 obsv 13 13 14 13 14 13 ctrb 12 5 12 5 14 8 Table 5. Several linearized models are obtained by linearizing at dierent trim conditions for steady wings level ight.1 Model selection From the original twelve outputs: T Y = nx ny nz VT α β P Q R φ θ ψ of the Ariel model. In table 5. This results in a bad controllability condition.1: Observability and controllability of dierent linearized models Simply put. but simulations show that it is still possible to operate at this condition. Given the fact that there are four control inputs. up to 4 outputs can be selected for tracking. 1000 and 5000 m) and velocities (30 and 55 m/s) are selected. The rank of the controllability matrix is smaller than the rank of matrix A for the higher velocity. If the extended system is observable and controllable. This means that the selected outputs can be controlled. This is because the constraints on the aircraft are active in the nonlinear Ariel model. the observ24 . to satisfy the condition that the extended state space system must be both observable and controllable to be able to track the selected outputs [26]. At this trim condition the throttle is at its maximum. Dierent altitudes (100. If more then 4 outputs are selected for the extended state space system. it seems logical that it is possible to control a maximum of four outputs. as long as no increase in throttle is required.Chapter 5 Simulations and results 5.1 a list of the dierent operating points is shown. the selected tracking outputs become part of the state space system. implying that tracking of the selected outputs is not possible. then all the (extended) states can be controlled. by introducing the extended state space model. The exact trim conditions can be found in appendix D. thus being unable to increase the velocity.

Other systems may be able to include more or less outputs.07 0. depending on the specic system dynamics.068 0 0 0 0 0 0. both times a step is put on the yaw angle. in gure 5. the maximum number of outputs that can be selected for this particular system is equal to four. To compare the performance. However. This way it is possible to compare the response of the system and the resulting 25 . it is not possible to keep the aircraft from drifting o course. As can be seen in section 5. a switching controller can be designed. All the simulations are done by using one linearized time invariant model. (5. Therefore. nz and ψ have been chosen.1) 5. physical values that can be fed into the nonlinear system. This is a property of this particular system and is needed. These 'real' values are the values shown in all the simulation plots. After the simulations. φ.068 0 0. For better performance and more reliable results. which switches between dierent linearized models when the operating conditions of the aircraft change. to show that this control method works. because otherwise the system needs to be reduced by another state. ψ .ability and controllability conditions are no longer satised and the outputs and states of the original system can no longer be controlled and/or tracked.1. and care should be taken when a switching controller is applied that the overall stability of the system is maintained. only the magnitudes of the state space system dier for the dierent operating conditions.2 the simulation results are shown when the outputs VT . In gure 5.998 30. it can be seen that not all output variables give the desired system performance.4. The LTI model obtained by linearizing the Ariel model at 1000 m altitude and 30 m/s velocity is used in all the simulations. θ and ψ are chosen. outputs and controls to obtain the 'real'.1 the outputs nx . this does not necessarily needs to be implemented.012 2. Without being able to control the yaw angle of the aircraft. the general dynamic behavior remains the same. the values of the steady state operating conditions are added to the system states.2 Output selection One of the four outputs that needs to be selected is the yaw angle ψ . Compare two dierent selections.052 0 0 0 0 0. ny . For the four outputs that are tracked. Switching between dierent models can cause instability. By doing dierent simulations with dierent selected outputs for tracking.061 1 3 7 7 7 7 5 The reference signal is the change from the steady state operating conditions. dierent reference signals are dened.238 2.068 0 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 2 6 6 u0 = 6 6 4 δT δe δa δr δf 3 2 7 6 7 6 7=6 7 6 5 4 0.328 0.790 0.068 0 0 0 −1000 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 2 6 6 6 6 6 6 6 6 6 y0 = 6 6 6 6 6 6 6 6 6 4 nx ny nz VT α β P Q R φ θ ψ 3 2 7 6 7 6 7 6 7 6 7 6 7 6 7 6 7 6 7 6 7=6 7 6 7 6 7 6 7 6 7 6 7 6 7 6 7 6 5 4 0. This model is specied in appendix B and operates at the following equilibrium point: 2 6 6 6 6 6 6 6 6 6 x0 = 6 6 6 6 6 6 6 6 6 4 U V W φ θ ψ P Q R pN pE pD 3 2 7 6 7 6 7 6 7 6 7 6 7 6 7 6 7 6 7 6 7=6 7 6 7 6 7 6 7 6 7 6 7 6 7 6 7 6 5 4 30 0.

In appendix C the dierence between the control eorts and the states of the system can be seen. nz and ψ (solid: reference. This seems a much less natural way of controlling an aircraft. nx 0.5 1 1.5 0 0.015 0. dashed: control signal) 26 . The step op the yaw angle is mainly taken care of by the rudder. By selecting the accelerations as output. Although the tracking in both scenarios is good.01 0.06 0. which seems a logical control eort.005 −0.08 0. The throttle and elevator also take hold of a big part of the total control eort.g. the response is much faster when the angles and velocity are selected as output.5 0 20 40 t [s] 60 80 100 Figure 5.1: Tracking of the outputs nx . the control eort and the eect on the internal states dier. the step on the yaw angle results in a much smaller eort of the rudder.05 [m/s ] 2 1 [°] 0 20 40 t [s] 60 80 100 0.07 0.015 0 20 40 60 80 100 ! 1.05 0 −0. the outputs all follow the given setpoint (note that ny will also converge over time).95 −0. ny .actions of the four system inputs for the two selections.04 0 20 40 nz 60 80 100 −0. The control eort when choosing as reference output the three angles and the velocity for tracking.005 [m/s ] 2 ny 0. Also.01 0. e. seems to give a more 'natural' aircraft control than the other selection of outputs.

05 [°] 0.03 0 0. the weighting matrices Q and R and the weighting matrix for the system inputs: • parameter Tp : Prediction horizon. It is important to keep track of the velocity VT to keep the aircraft ying.07 0.01 0 −0. 27 .05 [m/s] 30 [°] 29.5 0 5 0. • parameter p: Pole location of the Laguerre model for the projected control signal.02 0.1 30.06 1 0. the prediction horizon Tp .04 0.02 0 5 ! 10 # 15 20 10 t [s] 15 20 Figure 5.02 0. φ. θ and ψ (solid: reference.9 29.04 0.2: Tracking of the outputs VT . the following congurations came out best: Conguration 1 2 3 Outputs VT φ θ ψ ny V T θ ψ VT P θ ψ Table 5. dashed: control signal) After comparing several possible output combinations.VT 30. the order of the control trajectory N .5 1.3 Tuning The behavior of the MPC can be changed by adjusting the location of the poles p.2: Selected outputs The choice of the outputs seems logical.01 0 5 10 " 0.03 0.95 29.01 0 5 10 t [s] 15 20 −0.85 0. In the following simulations.5 15 20 −0. the outputs of conguration 1 have been chosen as controlled output variables. The yaw angle ψ or the acceleration in y -direction ny seems a good choice to keep the aircraft on the right course. also the key parameter for control horizon. And the pitch angle θ is a good choice to control the altitude of the aircraft. 5.

approximately equal to 10/p.min 0 0 δe.3. The upper performance limit is reached when increasing p no longer changes the closed-loop response speed. Tp : For a stable system.e.2. The control signal tends to be more aggressive when N increases. For unstable and integrating processes.min ) · δt. • weighting matrices: Q and R.min 0 0 0 0 0 0 δa.min 0 −1 0 0 0 δr. i. Increasing N further still gives a little bit better results. Value δthrottle [% of max] δelevator [◦ ] δaileron [◦ ] δrudder [◦ ] min 1 -32 -16 -16 max 100 16 16 16 rate min -200 -250 -250 -250 rate max 200 250 250 250 Table 5.max − δa.• parameter N : Number of terms used in the Laguerre model to capture the projected control signal. • weighting matrix W : When constraints are applied the eorts of the controls need to be scaled.1. Q = I and R = 0 to simplify the tuning procedures. Prediction horizon p: p is a tuning parameter for the closed-loop response speed.2) 28 . The most eective tuning procedure for p is to start near the location of the plant dominant pole. The constraints on the Ariel model are shown in table 5. i. and thus faster closed-loop response speed. the prediction horizon is recommended to be chosen approximately equal to the open loop process settling time.max − δr. and to then subsequently speed up (or slow down) the closed-loop response by increasing (or decreasing) p to achieve the required performance. Parameter Weighting matrix W : When there are constraints active on the control signal. It was found that an upper limit was reached for |p| > 0.e. As N increases. these need to be scaled in such a was that the relative action of the controls is equally weighted. Tp is recommended to be chosen depending on the value of p.max − δt.3: Input constraints These are weighted and scaled as follows: W = (δe. the degrees of freedom in describing the control trajectory increases.4. Pole location N : The parameter N stands for the number of terms that will be used in capturing the projected control signal.min (5. the time that the response reaches 99 percent of its nal value. It was found that up to N = 9 the speed of the response still increases quite signicantly.max − δe.max − δe. but it is not really worth the extra computational eort for the small increase in response. larger than the time that the projected control signal eectively reaches a constant [1]. Larger p results in larger changes in the projected control signal in the initial period. As mentioned in section 4.

04 50 [m/s] 45 40 0 35 30 0 20 " 0. Without decreasing the pitch angle the aircraft starts to climb.15 0. The data in these tables was obtained by doing wind tunnel testing of the Ariel aircraft by Crump [2].05 0. Linearized model 30 m/s 55 m/s Operating velocity 55 m/s 55 m/s δt 114 100 θ 0. What can be seen.02 0 20 [°] 0.2 0 −0. is that the value of input δt goes to 114%.4 −0. the reference signal is very well tracked.3.00 δe 3. pitch angle and elevator input between the two linearized models both operating at 55 m/s can be seen in table 5. which is higher then the value at trim condition with a velocity of 55 m/s.2 −0. As can be seen in gure 5.3: System outputs for simulation of steps on the velocity (solid: reference. This dierence originates from the dierence of the atmospheric forces at dierent velocities. With increased velocity.9 3.05 0 −0.02 0.1 Simulation without constraints Several scenarios are simulated to show that the controller will always stabilize the system.4. The dierence in throttle input. Here simulations are shown when 5 m/s steps are put on the velocity until it reaches 55 m/s. if no constraints are applied.07 0.6 40 60 −0.4: Dierence in operating values for dierent models VT 60 55 0. which seems a logical result.4 Simulations 5. which are implemented in the nonlinear Simulink model by using lookup tables.4.5. The linearized model does not take this into account and this is where the dierence originates from.3 Table 5.06 ! 40 # 60 0 20 t [s] 40 60 0 20 t [s] 40 60 Figure 5. the pitch angle θ should be reduced to keep the aircraft at steady ight.1 [°] 0. dashed: control signal) 29 .

!throttle 200 150 [% of max] 3.4: System inputs for simulation of steps on the velocity pD 1120 1100 1080 1060 [m] 1040 1020 1000 980 0 10 20 30 t [s] 40 50 60 Figure 5.5 [°] [°] 0 −0.5 [°] 4 !elevator 3 0 20 ! 40 60 0 20 ! 40 60 aileron rudder 1 20 15 0.5: Aircraft altitude for simulation of steps on the velocity 30 .5 0 20 t [s] 40 60 10 5 0 −5 −10 0 20 t [s] 40 60 Figure 5.5 100 50 0 2.

01 0 −0. There are constraints on the maximum and minimum values of the controls and on the control rates. dashed: control signal) 31 . tracked outputs (solid: reference. VT 50 45 [m/s] 40 [°] 35 30 25 0. where the altitude of the aircraft is increased.5. as mentioned in section 5.3 0.02 −0. A second simulation is done decreasing the altitude of the aircraft.04 −0.04 0 100 ! 200 # 300 0 100 t [s] 200 300 0 100 t [s] 200 300 Figure 5.2 [°] 0.02 0 −0. 5.5. The reference and output trajectory.6: Aircraft climb.6.4 0. the controls and the altitude of the simulation are shown in gures 5.02 0.02 −0. for reconnaissance of the surroundings.5: Input constraints To see if the controller is capable of following a mission prole. two simulations are made.7 and 5. The reference trajectory is. To make the eort of all the controls the same.1. one increasing the altitude of the aircraft. ying 25 m/s and continues a steady ight at 2500 m altitude at 45 m/s. the controls are weighted so that the domain of each control input is valued equally.04 0. the aircraft controls are constrained.g.1 0. The aircraft starts at 1000 m altitude at 30 m/s.8 respectively. The constraints are listed in table 5.1 0 −0. typically representing the climb to cruise altitude.4.2 Simulation with constraints In reality. superimposed to the steady state operating point of the LTI state space model.06 200 300 −0. Value δthrottle [%of max] δelevator [◦ ] δaileron [◦ ] δrudder [◦ ] min 1 -32 -16 -16 max 100 16 16 16 rate min -200 -250 -250 -250 rate max 200 250 250 250 Table 5. begins to climb with a 8◦ climbing angle.03 0 100 " 0. e.01 −0. for the rst simulation.

4 −0.8: Aircraft climb. constrained system inputs pD 2600 2400 2200 2000 1800 [m] 1600 1400 1200 1000 800 0 50 100 150 t [s] 200 250 300 Figure 5.2 0 100 t [s] 200 300 0.6 0.2 1 0.2 −0. aircraft altitude 32 .7: Aircraft climb.2 0 −0.!throttle 100 80 [% of max] 60 [°] 2 1 0 40 20 0 4 3 !elevator 0 100 ! 200 300 0 100 ! 200 300 aileron rudder 1.8 [°] [°] 0.6 0 100 t [s] 200 300 Figure 5.4 0.

9.1 −0. the altitude is decreased from 1000 m to 200 m.2 −0.02 0 100 t [s] 200 300 0 100 t [s] 200 300 Figure 5.01 0 −0. dashed: control signal) 33 .10 and 5.3 0.9: Aircraft descent.11 respectively.02 0.1 0 [°] −0. The reference and output trajectory. VT 50 45 5 [m/s] 40 35 30 −5 [°] 10 x 10 −3 ! 0 0 100 " 200 300 0 100 # 200 300 0.01 −0.In the second simulation. the controls and the altitude are shown in gures 5. The velocity increases during the descent to 45 m/s. tracked outputs (solid: reference. 5. and goes back to 30 m/s at the steady levelled ight.

10: Aircraft descent. In the second simulation the throttle is completely closed (1%) when the aircraft starts the descent. In the rst simulation.1 0 0 100 t [s] 200 300 0 100 t [s] 200 300 Figure 5.5 20 10 3 0 0 100 ! 0. aircraft altitude Note that the controls hit the constraints in both situations. and goes down to zero several times after that to prevent the velocity from increasing too much.5 0 100 [°] 4 3.5 5.4 [°] 200 300 2.2 0. the throttle is fully open (100%).4 0. Dening a good trajectory by hand is quite dicult.5 0. However. constrained system inputs pD 1100 1000 900 800 700 [m] 600 500 400 300 200 100 0 50 100 150 t [s] 200 250 300 Figure 5.2 [°] 0 −0. An extra controller can be designed to translate for instance a list of way points intro a trajectory for the outputs that the MPC tracks.11: Aircraft descent.5 5 !elevator 200 ! 300 aileron rudder 0.4 0. this does not lie within the scope of this 34 . to reach the desired velocity as soon as possible.2 −0.3 0.!throttle 40 30 [% of max] 4. while the elevator is eecting a change to steady ight.

which takes into account the fact that one of the controls has a dierent numerical value. 35 . To simulate one of the controls becoming 'stuck' for instance.13. e. and this means the controller cannot change the banking angle by means of adjusting the ailerons any more. when a controller cable has failed. it is still able to y in a stable. steady ight. The simulation shows that the airplane is still capable of ying. This can simulate one of the controls being 'free' to move. within certain limits.3 Simulation of failure Another interesting feature of the model predictive controller with constraints applied is the ability to simulate some kind of failure of one of the controls.g. the minimum and maximum on the input constraint is set to zero. u = uf ixed = u0 . The dierence in the tracking of the reference signal and the control inputs can be seen in gures 5. The control signal of the throttle and the rudder changes to compensate for the failed ailerons. In this simulation the aileron is assumed to be stuck at its initial value. 5.4.research. e. allowing the UAV to return home for instance. This capability makes the model predictive control scheme useful for fault-tolerant control applications. The limitation is that the control can only be 'stuck' at it's equilibrium.g. The movement of the control surface will not be completely random however. Although the response is not as good without the aileron working. The simulation has been made with and without the failure of the ailerons. Using a dierent value for uf ixed requires a state space model which is linearized at a dierent operating point. Although the maneuverability of the airplane is limited. and more research needs to be done to investigate the 'free' movement of one of the control surfaces. the aircraft is still able to follow the reference trajectory. It will also be interesting to investigate how a random value for one of the controls aects the aircraft model.12 and 5.

05 0.45 0 [°] 0. dash-dotted: with failure) !throttle 26 25 [% of max] 2.07 30.01 0. dashed: without failure.02 0.35 −3 0 20 t [s] 40 60 −4 0 20 [°] −1 −2 2 1 40 60 2.05 0.VT 30.72 22 0 20 !aileron 0.8 !elevator 40 !rudder 60 40 t [s] 60 Figure 5.12: Tracked outputs (solid:reference.4 0.03 [m/s] 30.09 30.78 24 23 2.065 0.01 0 20 [°] 0.085 0.1 40 60 −0.06 0 −0.82 2.76 2.5 0.08 0.05 0.075 [°] 0.04 ! 40 # 60 0 20 t [s] 40 60 0 20 t [s] 40 60 Figure 5.07 0.13: System inputs (dashed: without failure.08 0.06 0 30.15 0. dash-dotted: with failure) 36 .7 0 20 [°] 2.05 0 20 " 0.74 2.

inputs and outputs of the aircraft model is easy. 37 . it is possible to track a given reference signal if no constraints are active. For example. The work presented in this report shows that the continuous time MPC using orthonormal functions to describe the control signal. but in general this behavior is not desired. A multi-input-multi-output linear time invariant state space system is derived from a nonlinear MATLAB Simulink model of an unmanned aerial vehicle. by setting a big step on the velocity. which uses orthonormal functions to describe the control signal. There are limits to what the aircraft model is still capable of. gives good performance when applied to the linear time invariant state space system derived from operating at the steady wings-level ight trim condition. It is shown that by using the MPC discussed in this report. the aircraft model is no longer able to follow any given reference signal. It should be noted that when the constraints are active. Simulations show that the application of 'hard' constraints works very well. Adding constraints on the controls. the controller will not just raise the throttle to its maximum and wait patiently for the aircraft model to reach the desired velocity.1 Conclusions In this report a continuous time model predictive controller is built. Control of a UAV by means of a model predictive controller works well. Constraints are applied to the controls and it is shown that the model predictive controller is also capable of fault tolerant control within certain limits. the controller tries to reach its objective by changing the other controls. The controller will never override the applied constraints. a MPC capable of handling time-varying state space systems needs to be designed to handle the trim conditions which result in time-varying state space systems. but it will pitch the aircraft in a downwards motion to speed up the acceleration process. The continuous time MPC which uses orthonormal functions to describe the control signal has not been used before to control a complex 4-input-4-output system like the aircraft model presented here. This is a logical result from the optimization algorithm. To use this controller for the entire ight scope of the aircraft.Chapter 6 Conclusions and recommendations 6. If the constraints are hit for a long period of time.

Another possible renement is to make a switching controller that switches between dierent linearized models for dierent operating conditions. It should be noted that a switching controller can cause instability of the whole system. Therefore care should be taken when implementing such a controller to assure stability of the aircraft model. The here presented MPC will assure that the aircraft model will smoothly follow the given trajectory and will also make the UAV robust to failure of any of the aircraft controls.6. a second (MPC) controller can be wrapped around the one presented here. to make trajectory following easier to implement. A switching controller can be implemented within the MATLAB Simulink environment. This trajectory must then be translated into the tracked output variables of the system. Finally. within certain limits. into an optimized North-East-Down and velocity vector. It is also possible to link this to the FlightGear ight simulation software to make a visual presentation of the simulation. the Simulink Aerospace toolbox can be used to translate between a reference trajectory of North-East-Down coordinates and the Euler angles to the selected reference model outputs. varying over time. The linearized models are typically not valid for the whole mission scope. and switching between state space models is necessary to assure good and reliable performance.2 Recommendations To obtain a good trajectory following solution. This second controller can translate for instance a given set of way-points. 38 .

81sφsθ −9.81cφsθ p1 −Qsθ Qsφ+Rcφ cθ 2 sθ 0 0 0 p2 p5 p8 0 0 0 p3 p6 p9 0 0 0 0 0 0 0 0 0 p4 p7 0 0 W −V 1 0 0 c2 Q c5 R − 2c6 P c8 Q 0 0 0 −W 0 U tnθsθ cθ sφ cθ V −U 0 −tnθcθ −sφ cφ cθ c 1 R + c2 P 0 c8 P − c2 R 0 0 0 c1 Q c5 P + 2c6 R −c2 Q 0 0 0 (A.81sφcθ 0 −Rcφ Qcφ−Rsφ cθ −9. tn = tan) 0 6 −R 6 6 Q 6 0 6 6 0 6 6 0 6 A(t) = 6 6 0 6 0 6 6 0 6 6 cθcψ 6 4 cθsψ −sθ 2 R 0 −P 0 0 0 0 0 0 p10 p12 sφcθ −Q P 0 0 0 0 0 0 0 p11 p13 cφcθ 0 9.1) 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 with: p1 p2 p3 p4 p5 p6 p7 p8 p9 p10 p11 p12 p13 = = = = = = = = = = = = = (1 + tnθ2 )Qsθ − Rcθ + tθQcθ + Rsθ V (cφsθcψ + sφsψ) + W (cφsθcψ + cφsψ) −U sθcψ + V sφcθcψ) + W sφcθcψ −U cθsψ + V (−sφsθsψ − cφcψ) + W (−sφsθsψ + sφcψ) V cφsθsψ + W (−sφsθsψ − cφcψ) −U sθsψ + V (sφcθsψ − sθcψ) + W cφcθsψ U cθcψ + V (sφsθcψ − cθsψ) + W (cφsθcψ + sφsψ) V cφcθ − W sφcθ −U cθ − V sφsθ − W cφsθ sφsθcψ − cφsψ sφsθcψ + sφsψ sφsθsψ + cθcψ cφsθsψ − sφcψ 39 .81cφcθ −9.81cθ −9.Appendix A Time dependent A matrix (c = cos. s = sin.

243 0.068 0 0 0 0 0 0.637 −0.615 −0.068 0 0.001 0.009 0 −2.669 0 0 0 0 0 0 0 0 −30.032 6 6 0.014 −7.021 −0.018 −0.001 −0.001 −2.006 0 0 0 0.126 0 0 0 0 0 0 0 2.146 6 6 −1.003 0 −1.000 −0.525 −1.012 2.004 0 0 0 0.052 0.011 6 6 1.001 −2.328 7 5 4 0.183 0 −0.001 0 0 0 0 0 0 0 0 0 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 Bm 3.052 0 0 0 0 0.787 0 0 0 0 0 0 0 0.061 5 1 3 2 6 6 6 6 6 6 6 6 6 =6 6 6 6 6 6 6 6 4 Am −0.967 −0.Appendix B LTI Ariel model The time invariant Ariel model linearized at trim condition 2 6 6 6 6 6 6 6 6 6 x0 = 6 6 6 6 6 6 6 6 4 U V W φ θ ψ P Q R pN pE pD 3 2 7 6 7 6 7 6 7 6 7 6 7 6 7 6 7 6 7 6 7=6 7 6 7 6 7 6 7 6 7 6 7 6 7 6 5 4 30 0.890 0 0.012 −9.068 −0.030 0.002 −0.998 30.002 0 0 0 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 0 9.156 0.161 0.012 −29.998 0 −0.089 −2.068 0 1.651 0 6 6 6 0.033 0 −0.473 0 0 0 0 0 0 0 0.681 0.238 7 6 2.012 0 −0.504 0 0 0 0 1 0 0.299 0 0 0 0 0 0 −2.012 30.702 0.075 0.068 0 0.07 0 0 0 0 0 0 0 0 0 −0.07 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 −0.408 0 −3.349 0 1 0 0 0 0 0.208 −0.068 0 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 δT 6 δe 6 u0 = 6 δa 4 δ r δf 2 3 2 0.037 −0.019 0.640 0 0 0 0 0 0 0 0.07 0.787 0 −0.215 0.790 7 7 7 6 7 = 6 0.128 =6 0 6 6 0 6 6 0 6 6 0 6 4 0 0 2 40 .478 −0.431 −0.068 0 0 0 −1000 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 2 6 6 6 6 6 6 6 6 6 y0 = 6 6 6 6 6 6 6 6 4 nx ny nz VT α β P Q R φ θ ψ 3 2 7 6 7 6 7 6 7 6 7 6 7 6 7 6 7 6 7 6 7=6 7 6 7 6 7 6 7 6 7 6 7 6 7 6 5 4 0.056 0 28.001 1 0 0 0 0 0 0 0.998 0 −0.119 0 0.529 −0.

012 0 0 0 0 0 0 0 0 0 0 0.002 6 0 6 =6 0 6 6 0 6 6 0 6 6 0 6 4 0 0 2 0.0223 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.372 0 6 6 6 −0.033 0 0 0 0 0 0 0 0 0.295 0 0 0 0 0 0 0 0 0 0 0.001 0 0.068 0.042 0 0.081 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 Dm 41 .044 6 6 0.Cm −0.0080 0 6 6 6 0.001 0 0.002 0 0 0.054 0.331 0.033 0 0 0 0 0 0 0.139 0 0 0 0 1 0 0 0 0 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 0 0.001 6 0 6 6 0 6 6 0 =6 6 0 6 6 0 6 6 0 6 4 0 0 2 0 −0.043 0 0 0 0 1 0 0 0 0 0 0.997 6 6 −0.

8 2.2 −0.1: Control eorts for outputs nx .5 [°] [°] −0.Appendix C Controls and states scenario 1 and 2 !throttle 34 32 [% of max] 30 28 26 24 22 0 20 40 60 80 100 [°] 2.6 !elevator 0 20 40 60 80 100 !aileron 1 0.85 2.5 −0.75 2.7 2.1 0 0.1 −0. ny . nz and ψ 42 .3 !rudder 0 0 20 40 t [s] 60 80 100 0 20 40 t [s] 60 80 100 Figure C.65 2.

2 0 50 t [s] 100 0 0 50 t [s] 100 0.1 0.1 0 50 ! 0.5 0 100 −0.2 [°/s] 0.6 [% of max] 20 [°] 2. nz and ψ 43 .8 2.3: States for outputs nx .8 !elevator 15 0 5 10 !aileron 15 20 0 5 10 !rudder 15 20 2 1.5 0. φ.4 2.1 0 −0.02 0 50 " 1. θ and ψ U 33 [m/s] 32 0 31 30 −0.05 0 50 # 100 0 50 t [s] 100 Figure C.05 0 −0.5 [°] 1 0.1 0.02 0.2 0.!throttle 25 2.2 W 2 0 50 P 100 0 50 Q 100 1.5 0 [°] 10 0 −10 −20 −30 −40 0 5 10 t [s] 15 20 0 5 10 t [s] 15 20 Figure C.2: Control eorts for outputs VT .2 2 10 1.4 [°] 0.2 0 −0. ny .5 1 100 0 0.8 0 50 R 100 0.5 V 2.

θ and ψ pD 2000 pD 2000 1800 1600 1400 1200 [m] 1000 800 600 400 200 0 1800 1600 1400 1200 [m] 1000 800 600 400 200 0 0 10 20 30 40 50 t [s] 60 70 80 90 100 0 2 4 6 8 10 t [s] 12 14 16 18 20 (a) scenario 1 (b) scenario 2 Figure C.04 0.95 29.08 0.9 29.1 0.02 0.5 0 −0.5 W 2 0 10 Q 20 1.04 0.5 0 10 # 20 0.1 20 −10 −20 10 0 V 2.1 0 10 " 20 −0.05 30 [m/s] 29.4: States for outputs VT .5 1 [°/s] 0 0 0.85 0 10 P 0.06 0.5 0 10 R 20 1.5 0 0 10 t [s] 20 Figure C.02 0 10 t [s] 20 0 0 10 t [s] 20 1 0.5: State pD 44 .U 30.1 0 10 ! 20 −0. φ.02 [°] 0 −0.

1: Operating conditions 45 .13 0 0 0 0 0 0.02 Table D.51 54.79 0.05 30.36 0.01 0.Appendix D Trim conditions Operating point alt [m] vel [m/s] 100 30 1000 30 5000 30 100 55 1000 55 5000 55 VT [m/s] 30.90 0.05 0.01 ψ [rad] 0 0 0 0 0 0 δt [%] 23 24 32 100 100 100 controls δe [rad] δa [rad] 2.07 0 0.55 54.84 0.24 54.24 3.47 0 0 3.68 outputs φ [rad] θ [rad] 0 0.33 1.11 0 0.06 0.07 30.33 0.06 0 0.24 δr [rad] 0.31 2.

The University of Sydney. Royal Melbourne Institute of Technology. and C. Bil A multipurpose autonomous ight vehicle system Technical report. [3] D. H. [2] M.a prospectus Aerospace America. Abanteriba. Howard and I. L. Kim and S. [7] K. 1588-1600. 1993. E. Nelson. Wong Unmanned aerial vehicles (UAVs) . 417425. [6] K. December 2003. Vol. Bil Uav's over Australia. 2002. Sastry Decentralized Nonlinear Model Predictive Control of Multiple Flying Robots in Dynamic Environments IEEE Conference on Decision and Control. Kaminer Survey of unmanned vehicles In Proceedings of the American Control Conference Vol. Wong Six Degree of Freedom Flight Dynamic and Performance Simulation of a Remotely-Piloted Vehicle Technical report. McDermott. 74(16). Newman and K. 1997. 1995. February. 1994. RMIT Wackett Centre. [4] R. R. pp. 1989. L. Wang Continuous time model predictive control design using orthonormal functions International Journal of Control. Wong and C. Barnett.are they ready this time? are we? Technical report. 46 . Vol. Thompson. Juul.market and capabilities In Bristol RPV/UAV Systems Conference. [5] R. M.Bibliography [1] L. 5: 2950-2953. H. RAES Sydney Branch. M. and G. N. 1995. 27: 26-28. Williams Submersible control using the linear quadratic gaussian with loop transfer recovery method In IEEE Sympsium on Autonomous Underwater Vehicle Technology. [10] D. D. 1998. J. 2001. [9] D. Shim. Crump The dynamics and control of catapult launching unmanned air vehicles from moving platforms Technical report. Moore Unmanned air vehicles . [8] L. S.

1992. Demircioglu. Fourth International Conference on Chemical Process Control. Fourth International Conference on Chemical Process Control. H. J. Chemical Process ControlCPC IV. Chemical Process ControlCPC IV. Ray (eds). 139. 1993. the bad. B. Amsterdam. Ricker Model predictive control: State of the art In Y. E. England: Harlow. Muske. R. 2001. 1960. P. Amsterdam. New York: Wiley. 2000. Lee Model predictive control: The good. 55-74. [23] C. L. J. [15] K. W. García. M. Vol. Sydney. 27. 28. 25(3): 335-348. Pt. Australia. [19] H. [13] N. D. W. [17] M. Vol. Maciejowski Predictive Control: with constraints Pearson Education Limited. D. Elsevier. H. 1994. R. J. 1989. Demircioglu. 1960. 419-444. Japan. E. Lee Statistical Theory of Communication New York: Wiley. pp. Morari Model predictive control: Theory and practicea survey Automatica. Morari. 271-296. R. 1992 [21] P. Gawthrop Multivariable continuous time generalized predictive control Automatica. Vol. [14] M. B. Vol. Clarke CGPC with guaranteed stability properties IEEE Proceedings. Phillips Mechanics of ight Hoboken. H. J. 1991. J. Arkun W. Arkun W. D. [20] H. pp. and the ugly In Y. K. 39(2): 262-287. 1991. Rawlings Model predictive control with linear models AIChE J. Wylie Advanced Engineering Mathematics New York: McGraw-Hill. pp. Gawthrop Linear predictive pole-placement control: practical issues 39th IEEE Cenference on Decision and Control. 697-713. J. Meadows. Elsevier. 1991. M.[11] Warren F. Demircioglu. S. Kyoto. [12] C. 2004. [16] J. Ray (eds). Gawthrop Continuous time generalized predictive control Automatica. 160-165. P. [22] Y. 47 . Muske Nonlinear model predictive control: A tutorial and survey ADCHEM '94 Proceedings. Prett. 371-380. Vol. Rawlings. [18] H.

Kailath Linear systems Englewood Clis. 2000.C. [25] G. Graebe. New Jersey: Prentice-Hall. Salgado Control System Design Englewood Clis. S. 48 . M. 2007. Goodwin. Wang Model Predictive Control System Design and Implementation using MATLAB In preparation. 1980. [26] L. New Jersey: Prentice-Hall.[24] T. Publisher: Springer-Verlag.

Sign up to vote on this title

UsefulNot useful- Aircraft Design
- Design of Autopilot for UAV
- Pole Placement Control Technique for Tracking Path Control of UAV by Longitudinal and Lateral Parameters Control
- Modelling of a Small Unmanned Aerial Vehicle 2168 9695 1000126
- 10.1.1.119
- Untitled
- A Combined MBPC 2 DOF H∞ Controller
- Documento de Regras do UAV Outback Challenge 2014
- Air Space Platforms[1]
- Hood Tech Vision Releases New Pod Featuring VIDAR
- Proposal of Certification of UAV
- JFCOM-UAS-PocketGuide
- UAV Thesis
- Coast Watch UAV
- DronesStartupLandscape Apr 2015
- Unmanned Aircraft Systems_ Part I
- 97-0349
- 10.1.1.126.6802.pdf
- Literature Review
- Co-Development And Co-Production Of Unmanned Aerial Systems [UAS] - Hindustan Aeronautics Limited [HAL]
- 9933817.pdf
- Unmanned Aerial Vehicle
- UAV Course Syllabus
- Commercial Drones Poised to Become Prodigious Worldwide Business by 2021
- Drone 280 User Manual V1.1.1.0
- Flight.international 24.February.2015
- uvahs
- Worldwide Commercial Drones Market Analysis Report
- 2014 4th_year Fall Project Memo
- Dept. of Transportation - Unmanned Aircraft System (UAS) Service Demand 2015 – 2035
- Uav Predictive Control