You are on page 1of 190
Diss. ETH No. 14899 Modeling, Identification, and Control of Autonomous Helicopters A dissertation submitted to the SWISS FEDERAL INSTITUTE OF TECHNOLOGY ZURICH for the degree of Doctor of Technical Sciences presented by OLIVER TANNER Dipl. Masch.-Ing. ETH born February 11, 1971 citizen of Wolhusen, LU accepted on the recommendation of Prof. Dr. H. P. Geering, examiner Prof. Dr. W. Schaufelberger, co-examiner 2003 ISBN 3-906483-04-5 IMRT Press c/o Institut fiir Mess- und Regeltechnik ETH Zentrum 8092 Ziirich Schweiz 04/2003 IMRT yy Measurement and Conta Laboratory Abstract Modeling, identification, and control of unmanned aerial rotorcraft vehicles is the research area of this work. The thesis introduces a mathematical model for control design whose parameters are identified in closed loop. A new flight control structure for a small helicopter under hovering conditions is developed. Based on the mathematical helicopter model, the control design uses the H, method for improving robustness and safety during flight. The results of the tested controllers are presented and discussed. The helicopter model equations are derived by describing the physical char- acteristics mathematically. Particularly, the flapping of the main rotor and of the Bell-Hiller stabilizer bar are modeled as they represent important as- pects of the helicopter dynamics. The model developed provides a structure and details the parameters to be investigated. Closed-loop identification methods confirm the characteristics of the helicopter model. The design of model-based robust controllers with a two-degree-of-freedom structure is a main focus of this thesis. Two well-established methods are applied: In a first approach, the feedback controller and the prefilter are designed by shaping the closed-loop transfer functions and by using an Hq. optimization procedure. The second method uses the Hao loop shaping approach for the feedback design with an extended scheme for the feed-forward path. The controllers are tested on a helicopter test bench designed for indoor flights and modified to meet the requirements for an autonomous helicopter. The test stand is equipped with a wind generator allowing an extended evaluation of the flight controllers. In comparisons to a state feedback controller, both robust controllers demonstrate their superior performance of stability in numerous flight experiments. Seite Leer / Blank leaf Zusammenfassung Modellierung, Identifikation und Regelung eines unbemannten autonomen Helikopters ist das Forschungsgebiet, mit dem sich diese Arbeit auseinander- setzt. Ein mathematisches Modell wird entworfen, das fiir den Regleren- twurf verwendet wird, und um dessen Parameter im geschlossenen Re- glerkreis zu identifizieren. Eine neue Struktur des Flugreglers fiir Modellhe- likopter, die primér fiir den Schwebezustand ausgelegt ist, wird eingefiihrt. Das mathematische Modell und die H.-Theorie wird benutzt, um Flugre- gler zu entwerfen, die die Sicherheit wihrend des Fluges gewahrleisten. Die Resultate der getesteten Flugregler werden prasentiert und diskutiert. Die Gleichungen der Helikopterdynamik werden hergeleitet iiber die mathe- matische Beschreibung der physikalischen Eigenschaften des Helikopters. Speziell die Modellierungen der Schlagbewegung des Hauptrotors und der Paddel zeigen wichtige Aspekte der Helikopterdynamik. Spezifische Param- eter werden aus der Struktur des Modells herausgelesen und untersucht. Identifikationsmethoden werden im geschlossenen Regelkreis angewendet und bestitigen dabei die strukturelle Eigenschaften des hergeleiteten He- likoptermodells. Der Entwurf eines robusten, modellbasierten Reglers mit einer eingebauten Vorsteuerung ist der Schwerpunkt dieser Dissertation. Die folgenden zwei bekannten Methoden werden angewendet: a) Der Regler wird mit einer H..-Synthese berechnet, wobei die Spezifikationen des geschlossenen Regel- kreises miteinbezogen werden. b) Der Hoo-Regler wird mit den Spezifika- tionen der Kreisverstarkung erstellt. Beide Regler werden an einem He- likopterpriifstand getestet. Die Anlage besteht aus einem Modellhelikopter, der durch ein Gestiinge fest mit dem Boden verbunden ist, und einem Windgenerator, mit dem zusitzlich die Robustheit der Regler getestet wer- den kann. Bin Vergleich der beiden unterschiedlich entworfenen Regler mit einem Zustandsregler zeigt ihre Uberlegenheit in Stabilitat und Flugeigen- schaft aufgrund von zablreichen Flugexperimenten. iii Seite Leer / Blank leaf Acknowledgements This thesis was written at the Measurement and Control Laboratory from 1998 to 2002 integrated in the helicopter project. During these years, I have made a varied experience in my research field and have been supported by many people. © First of all, I would like to thank my supervisor Prof. Dr. Hans-Peter Geering for his support throughout the whole time, for encouraging me when my efforts failed, as well as for awarding me when I suc- ceeded. My thanks also go to Prof. Dr. Schaufelberger, who agreed to be my co-examiner. My special thanks go to the helicopter team that supported me during the work in the last years. Christoph Eck, who gave me many ideas and a comfortable working environment. Jaques Chapuis and Markus Kottmann provided me with useful hints on many subjects ranging from programming of real-time application to the discussion of control theory. Milica Vuscovic and Marco Gerig helped me build up the new test bench and risked their lives by being co-pilots at my flight tests. Oskar Brachs, who gave his prompt attention to the mechanical workings. Many thanks go to all colleagues and former colleagues at the IMRT, where I want to mention in particular Christian Roduner, Theo Auck- enthaler, Marzio Locatelli, Esther Baumann, Urs Christen, Hans Ul- rich Honegger, Brigitte Rohrbach, Daniel Matter, Christopher Onder, and Essi Shafai. My parents, who made it possible for me to study Mechanical Engi- neering at ETH, which was the first of several steps that finally led to this thesis. I want to thank them for giving me this opportunity. « I wish to thank the many people who have supported me during the doctorial studies in many valuable ways. Special thanks go to Beda Estermann, Rolf Arnet, Adrian Winiger, Ruedi Géldi, Patrick Hoffmann, Iwan Stécklin, Philip Tingeler, and Hasina Duelli. My final thanks goes to Hans-Peter Dictz, who showed me the path of the New Warrior. Ziirich, March 2003 Oliver Tanner vi Notation Symbols Physical variables Ais Au (longitudinal) cyclic angle for rolling angle of attack for collective angle of attack for tail rotor (lateral) cyclic angle for pitching moments and products of inertia about «, y, and z direction acceleration due to gravity external moment about the «’-axes external moment about the y’-axes torque of the electric motor complete load weight of the helicopter external moment about the z?-axes time derivative of @ in the body frame time derivative of 8 in the body frame time derivative of in the body frame main rotor thrust tail rotor thrust velocity in x? direction command signal for motor velocity in y? direction velocity in z? direction frame angles vii Dee eR Lock number pitch angle roll angle yaw angle rotor speed Navigation variables ce faye Ter Ya, 4a perbsn e,bjn ye ebm w direction cosine matrix (transformation matrix from the nav- igation to the body frame) unity vector in x, y, and z direction position, velocity, and acceleration vector in the point G x coordinate in the earth, body, or navigation frame y coordinate in the earth, body, or navigation frame z coordinate in the earth, body, or navigation frame angular velocity vector Control variables Tu Tew Ww Wi2 viii design plant feedback part of the controller feed-forward part of the controller loop gain GK» loop gain KyG, sensitivity at the plant output sensitivity at the plant input complementary sensitivity at the plant output complementary sensitivity at the plant input: closed-loop transfer function from signal w to signal z weighting function affecting signal x weighting functions for loop gain model uncertainty norm bound for Ho. synthesis Acronyms DOF GPS LQ MIMO NLCF PWM SISO UAV UARV degree(s) of freedom global positioning system linear quadratic multi-input multi-output normalized left coprime factorization pulse-width modulated single-input single-output unmanned aerial vehicle unmanned aerial rotorcraft vehicle Seite Leer / Blank leaf Contents 1. Introduction 1.1. Goal of this work... 6... ee ee 1.2. Project background and structure of the thesis ....... . Modeling, identification, and system analysis 2.1. Introduction to modeling 2.2. Motion of the aircraft... .. 2.2.1, Navigation ....... 2.2.2, Motion .........- 2.3. Physical nonlinear model 2.3.1. External forces and moments 2.3.2. Main and tail rotors 2.3.3. Engine model... ......-0020000 2.3.4. Coupling of all forces and torques... ... « 2.4. System description for control design... .. 2... 2.4.1. Linearization 2.4.2. Horizontal motion 2.4.3. Vertical motion . 2.5. Identification of deterministic models . 2.5.1. Identification methods . . L 2.5.2. Model structures and identification... ... 2.6. Helicopter models... ......-. . Control design 3.1. Controller for helicopters. ... 0.2.0.0 -00 005 3.1.1. Controller structure 3. Engine controller 3.2. Classical control design 3. Stability and robustness margi State feedback control design . . Attitude control Position control . xi Contents 5.1. Structure of the autopilot ..........- 5.1.1. Navigation 5.1.2. Control 5.2. Real-time system . . . 5.3. Results......... 5.3.1. Hovering 5.3.2. Hovering with wind generator . . . 5.3.3. Hovering with a disturbance input 5.3.4. Discussion... ......--..200-% . Summary and Outlook . Navigation A.1. Earth-to-navigation-plane transformation A.2. Euler angle derivatives... . . Equations B.1. Nonlinear plant. .... 2.000020 008 3.3. Hoo mixed sensitivity design 70 3.3.1. Introduction 70 3.3.2. Hq. theory 70 3.3.3. The GS/T-scheme ........ 17 3.3.4. Attitude control ........- 78 3.4. Heo loop shaping design .... 2... - 84 3.4.1. Introduction 84 3.4.2. Theory of the Ho. loop shaping method . . .. 84 3.4.3. Attitude control 89 3.5. Conclusions and results 92 3.5.1. LQ controller and position controller... .....- 92 3.5.2. Hao controllers .......... 93 3.5.3. Feed-forward controller 94 . Helicopter test bench 107 4.1. Overview of the helicopter test bench. . 107 4.2. Sensors and Actuators . 108 4.3. Helicopter modules 110 4.4, Drive aoe 4.5. Delta robot arm for position 112 4.6. Wind generator. .........--.-- 113 . Implementation and results 117 139 143 143 144 145 Contents B.1.1. Constant Values... 2... eee ee 145 . Variable values . . 146 B.1.3. Differential equations 149 C. Data of the helicopter and model 151 C.1. Helicopter... 2. ee 151 C.1.1. Parameters 151 C.2. Linear model . . . 154 C.2.1. ABCD matrices of the linear model... . 154 D. Control theory 157 D.1. Controller calculation 157 E. Simulation results 159 Curriculum Vitae 175 xiii 1. Introduction During the last years, a new market for unmanned aerial vehicles (UAV) has grown strongly in the civilian sector with a wide range of application. Depending on their task, there are different forms of UAVs. The best known UAVs are airplanes with turbo propeller or jet propulsion, which are mostly used in the military sector. Another area is the unmanned aerial rotorcraft vehicle (UARV) with the characteristics of vertical take-off and landing which includes also small helicopters of different structures. UAVs of this type are well suited for the film industry, which has a demand for hovering capabilities and for high precision flight. Some of the challenges when in the development UAVs of different types are very similar. Constraints like limited payload, small onboard energy resources, and aggressive atmospheric conditions are requiring light, small, and failsafe electronics components with low power consumption. Besides this trend in the field of electronics and embedded systems, recent devel- opments in sensor technology helped build various types of UAVs. The accuracy of sensors like GPS and gyros has increased, while their size and weight has decreased over the years. Nevertheless, the systems are getting very complex and safeness has to be guarantied in every situation. Moreover, the physical behavior of the air vehicle marks a challenge. It starts when the construction of the UAVs is finished and the first test runs are done. To distinguish between sensor noise and system dynamics is only possible if the quality of the signals is sufficient. Only then, the physical dynamics of the system can be investigated and used for the design of an autopilot. Small helicopters, which are often used as UARVs, certainly have an inter- esting physical background. There are mechanic and aerodynamic aspects which have to be modelled and included in a controller for stabilizing the helicopter. In contrast to the normal-sized helicopters, small helicopters of- ten have an additional servo rotor, called a Bell-Hiller stabilizer bar, which 1. Introduction is damping the helicopter rotor dynamics. All these components have to be considered for a control design. 1.1. Goal of this work The main focus of this work lies in the model based control design for a rotorcraft with the following physical characteristics: The helicopter has one main rotor with two blades, a servo, and a tail rotor. After the modeling phase and some general investigations of the rotor, a controller is developed based on this model. The procedure of the control design should be as flexible as possible so that it can be applied to every helicopter of the same type. There were three steps necessary for the realization of this goal. In each step, attention is focused on some specific problems. Modeling The aerodynamic aspects of a helicopter, and particularly of the main rotor, are described in a large number of books. ‘The main rotor blade makes different movements under the aerodynamic load. The flapping motion of the blades is a central phenomenon of the helicopter rotor and its dynamics are influenced by the airspeed of the helicopter and the rotor speed. The Bell-Hiller stabilizer bar, and its dynamics acting on the helicopter are decreasing the bandwidth of the vehicle. The whole formulation can be included in a mathematical model. This would go beyond the scope of this work. However, the necessary char- acteristics, such as the high-frequency dynamics and the instability of the plant, can be extracted from the theory and transformed into a useful linear system. Based on this system, further investigations are possible. Identi ion System identification is a broad field which offers several methods to deter- mine the structure and the parameters of an unknown plant. One method is to identify, the physical parameters of the plant. But sometimes it is not possible to express the unknown parameters in a suitable way for identifi- 1.2. Project background and structure of the thesis cation. In such a case, an other method is chosen were a system structure is taken and identifying the parameters identified which cannot always be interpreted as physical constants. The success of identification depends on the experimental data, the appro- priate system structure, and the applied method. For the helicopter, the closed-loop identification theory has been applied. Because the helicopter is an unstable system, a pilot or a controller is always in the loop and adul- terates the dynamics of the system. The theory tries to distinguish between the two parts and to find a suitable set of parameters which can be used for the control design. Control design Experience gained with classic compensators is the basis of the design of effective controllers. The classic compensators are also the first ones which are implemented in a new test environment, because dealing with only a few design parameters is helpful for tuning the controller. A working, classic controller is also useful for the investigation of the closed loop. A state feedback controller is especially suited for the helicopter, because it is a system with multiple inputs and outputs. ‘The helicopter model has some immeasurable state variables which describe the movement of the servo rotor. Robust control is providing solutions for such systems. The theory of robust control in this thesis is based on the work of (Zhou et al., 1996), which offers different possibilities for applying it to a linear plant. The application of the controller requires a study of the plant, thus the tuning and the implementation is working after the specification of the designer. Reducing the time needed for the design procedure by using a small number of design parameters is certainly one specific goal of every control engineer. 1.2. Project background and structure of the thesis This thesis is embedded in a number of helicopter projects at the Measure- ment and Control Laboratory of ETH Zurich. The project of the helicopter 1. Introduction group started in 1986 with the construction of an indoor flying test bench which is described in (Weilenmann, 1994). In 2000, the concept of the helicopter test bench was introduced and the whole construction rebuilt (see Chapter 4). In 1995, the development of a free flight helicopter was started. The goal was to participate at the International Aerial Robotics Competition (IARC) (see Chapuis et al., 1996). Full autonomous flight was reached with a second prototype and demonstrated at the COSY meeting (see Chapuis et al., 2000). During the years 1996 to 2002, research at the Measurement and Control Laboratory was focused on navigation with a dissertation by (Eck, 2001) and on the application of robust controllers in this thesis. ‘The thesis is structured as follows: A detailed description of the helicopter model developed and investigated, including equations and identification of parameters, is given in Chapter 2. The controllers, based on the mathemat- ical models of the helicopter, are presented in Chapter 3. The background of the control design and the validation is also included in Chapter 3. In Chapter 4, the test bench is described including its main parts the helicopter and a wind generator. Various results of the performance and quality of the tested controllers are presented in Chapter 5. A summary and outlook are given in Chapter 6. 2. Modeling, identification, and system analysis 2.1. Introduction to modeling Mathematical descriptions and simulations of the helicopter’s flight dynam- ies are numerous in literature. The standard works are (Padfield, 1996), (Prouty, 1995), and (Bramwell, 2001) which describe the aerodynamics and the behavior very exactly. ‘The helicopter dynamics and solutions of aerodynamic equations presented in those works are mostly for large-size helicopters. The results are transferred here to a model of a small-scale helicopter which is used as an unmanned air vehicle. The aerodynamics of small-scale helicopters are quite similar, although they are designed for a different flight envelope than full-scale helicopters. Describing the flight behavior of a small helicopter presents a difficult task because the vehicle can be viewed as an arrangement of interdependent subsystems (see Figure 2.1) which have their own distinctiveness. Obvi- ously, the main rotor is the dominating part, and it has the most complex structure. Its influence on the force and moment equations is the main part of the modeling, Setting the main rotor model as main topic follows most concepts in literature. The remaining main physical parts of the helicopter, such as fuselage, drive, and tail rotor, are investigated and related to each other. Every item has its own characteristics and can be different from helicopter to helicopter, depending on the exact design of the component (e.g., electric or internal combustion engine). The resulting equations in this work are modeled on the test bench described in Chapter 4, but they can easily be transferred to any helicopter design with a two-blade rotor, a Bell-Hiller stabilizer bar, and a tail rotor. The frame of all components is given by the definition of the different 2. Modeling, identification, and system analysis fuselage empennage velocities forces m(a— rv — qu) = Fe Fe =Te+...[?]m(o— pw — ru) = Fy motor/engine ee ee transmission . [Euler angles main rotor ko Ke be f o rotational tail rotor velocities moments Tes — =D by b= Ms + Iya = M controller ad physical parts motion Figure 2.1.: The structure of modeling a helicopter coordinate systems, which is certainly a main part of the navigation block. Thereby, the motion of the aircraft can be specified with the geometric fuselage attributes. ‘The description of the forces of the main and tail rotor is given in the next section. In the last part, the engine is modeled and integrated in the mathematical model. 2.2. Motion of the aircraft In Section 2.2.1, the equations of the translational and rotational motions of a helicopter are derived. For this purpose, the aircraft is assumed to be a rigid body. This simplification also concerns the navigation theory which describes the necessary coordinate transformations. Furthermore, it is assumed that position, velocity, angles, and angular rates are mea- sured in every design unmanned rotorcraft vehicle, because it is normally equipped with an inertial measurement unit (IMU) and a global position 2.2. Motion of the aircraft system (GPS). 2.2.1. Navigation The navigation process is the calculation of coordinates that are relative to a known reference point. Usually, this reference is a specific coordinate system, and for different applications, such as attitude or position control, the definition of several frames of reference is necessary, as described below: Earth-fixed frame: This frame is a Cartesian coordinate system with 2°, y®, 2%, pointing to north, east, down, with the origin at the top of a fixed point on the earth. It can also be regarded as an inertial frame, because the origin is not considered to be accelerating. Body-fixed frame: The Cartesian coordinate system is oriented along the helicopter axes x°, y®, z° (corresponding velocities: u, v, w); point- ing forward, sideward right, downward, and having the origin at the center of gravity (helicopter). Taking the center of gravity as the loca- tion of the body-frame origin simplifies the derivation of the kinematic equations. Navigation frame: The third Cartesian coordinate system is oriented along the helicopter axes (like the body frame). Its origin is at the same place as the earth frame (2", y”, 2”). For the description of the attitude based on navigation and body-fixed frames, there are two common ways used in navigation: One is the quater- nion formulation, the other is to use Euler angles (see (Farrell and Barth, 1999)). The Buler angles have a singularity at the transformation of angu- lar rates (see equation (2.2) and (2.3)) when the helicopter rolls or pitches to 90 deg. This phenomenon does not apply to the quaternion formulation. ‘As Euler angles are easier to understand and the helicopter not expected to make rolls, the quaternion formulation will not be used here. Hence, the transformation matrix from the navigation frame to the earth-fixed frame with the pitch 0, roll @ and yaw % angles is given by the direction cosine matrix C? ‘curc) —sucdtcpsbsh ss + esbed Cl = |such chch+sysbsd —cpsd + spsOcd (2.1) —80 cbsb cbc 2. Modeling, identification, and system analysis with ¢ = cos and s = sin. The standard sequence used in flight dynamics js to first rotate about the z-axis (r” — r™ with angle w), then about the y-axis (r™ — r™” with angle ¢ ), and finally about the z-axis (r™"” > r with angle @) (see Appendix A.1). Since all transformation matrices are orthogonal, the transposed matrix corresponds to its inverse ort ign The relationship between the rate of change of the orientation angles and the angular velocities of the fuselage in the system of body axes is a Pp Pex» + Geyr +7,» = Glq» + Oeyn’ + Pen = To |G (2.2) r with 1 sing tan cosd tanf T,=|0 cose —sing (2.3) 0 sind cos8 cost coad Appendix A.1 shows the calculation of Ti5'. 2.2.2. Motion The body frame moves with time-varying velocity components wu, v, w, and angular rate components p, q, and r. It is under the influence of the applied forces F;, Fy, F,, and the moments M;, My, and M,. The equations of motion can be derived by equating the rates of change of the linear and angular momentums to the applied forces and moments. Assuming constant mass, the equations are conveniently constructed by selecting an arbitrary reference point P inside the fuselage and by deriving the expression for the absolute acceleration of this point. The center of the moving axes is located at the helicopter’s center of mass G. The process is initiated by considering the position vector of the point P relative to G, (24) lpg = Upc = (2.5) 2.2. Motion of the aircraft Figure 2.2.: Systems of the orthogonal axes systems for helicopter flight and the reference points O (center of the earth frame), G (center of gravity and body-fixed frame), and P (point on the vehicle) Next, small changes in the angles 59 about each axis in the body frame are considered. They are depending on time (body rates) Gg | Oye _ 50 ae tt (26) and can express the derivation of the unit vector from the body frame W, X ead eab = Wy X Lyb (2.7) with angular velocity: w, =p-€,) +q- es +7-€.6 Since the fuselage is assumed to be rigid, the distance of the reference point P from the center of mass is fixed and the velocity of P relative to G can 2. Modeling, identification, and system analysis be written as Upg = (a2 — ry )ee + (rz? —p2 ew + (py? G2" )ee UPG * ab + UPG * eyb + WPG * e, (2.8) pg = bpg = (ipa — Fr UPG +4 WG) * ex6 + (pg —p wea +7 Upa) - ey + (wpa —q upa +p vPa) -e,6 (2.9) The velocity (relative to body-fixed frame) of the aircraft center of mass G in components is Ug =U eb+U- ep + Wes (2.10) and the velocity of the point P is vp =(u-ry?+q2’)-en+(v-p2't+ra2’)-ep +(w-qa’+py')-e. (2.11) = apg t (wy X Yp) (2.12) ap The sum of the external forces acting on the aircraft is Fo = Fe @qp + Fy eye + Fe €8 (2.13) The following translational equations of motions of the aircraft can be extracted from (2.12) and (2.13): F, = mu(i—rvt+qw) Fy =mu(i—pwtru) (2.14) = mut —qu+pv) The external moment vector about the center of mass can be written in the form Mo= [ re xar=[[ (ya. — 2 0y)dm ody he + Uf (z ae a.) , hs + Uf oer) a) a a =Le,+Me,+Ne, (2.18) 10 2.2. Motion of the aircraft where, e.g., Lap] (ye +2%)dm—ar [ (2 -y")am +a | vedm—(pq+%) [ azdm body Nb ody torre [ ay dm (2.16) Jvody Defining the moments and product of inertia J, are x — axis Ion = [ (y? +27)dm body aT [ (2? +22)am Pa (2.17) z—axis ‘zz =I (y? +27)dm body xz — axis Ip, = [ azdm body The shape of the fuselage has a characteristic asymmetry in the «z plane. The comparison of the products of inertia to the other products of inertia (Jen, Tyy, and I,z), demonstrates that only Iz: has an influential size. The rotational equations of motion of the aircraft thus may be written as follows: L=Is2p — (Iyy — Iez)q 7 — Inz(p q+?) (2.18) M = Iyyq— (lez — Iew)p 1 + Ine (p? — 1”) (2.19) N= Lz¥ — (Ine — Iyy)q Pp ~ Ixz(B— 9) (2.20) 11 2. Modeling, identification, and system analysis 2.3. Physical nonlinear model The flight characteristics of a helicopter can be described in two different distinct regimes — hover or low speed (up to about 10 m/s), including ver- tical maneuvering, and medium or high-speed flight. The low-speed charac- teristics are unique to the helicopter as an operationally useful regime. No other flight vehicle is so flexible and efficient in maneuvering slowly. Flight at low speed can be gentle or aggressive depending on aircraft performance and the urgency required by a particular maneuver. The single helicopter rotor is unstable at low speed. The coupled pitch/roll *pendulum’ instability is a product of the flapping rotor’s response to ve- locity perturbations. While designers of a craft can do very little about the instability at slow flight, they can make a high speed flying helicopter with fixed stabilizers almost as stable as a fixed-wing aircraft. In this work, only the hovering case will be considered. Small helicopters are mainly designed to fly at low speeds and to hover under difficult circum- stances. From a control engineer’s point of view, the design plant becomes less demanding for modern controller at increasing forward flight speed because the helicopter is more stable (see (Bramwell, 2001)). The nonlinear model will include the external forces and the physical prop- erties of the main and tail rotor. There are more nonlinear terms than in the derived equations of Section 2.2.2. 2.3.1. External forces and moments Figure 2.3 depicts the main forces of interest in helicopter research. These are the weight force F, the thrust force of the main rotor 7’, and the thrust force of the tail rotor Tp. The two additional drag force D and fin side force Pr do not affect hover and are thus neglected here. The weight force Fy is assumed to always act in the vertical sense. It is con- tained in the components of the body-fixed frame with the transformation matrix C?. Fy —sin mne 9 = |sindcos@ mre g (2.21) cos $0080 Me g 12 2.3. Physical nonlinear model T Fy ls Figure 2.3.: External forces and moments Moments caused by the main rotor hub @ and by the tail rotor Qr have to be taken into account as well. They occur due to the drag of the rotor blade profile and the mounting of the flapping blades. These terms of the main and tail rotor will be deduced in Section 2.3.2. The moment of the drive M4 of the rotors is a quite important area variable, because the tail rotor force depends on this moment (calculated in Section 2.3.3). 2.3.2. Main and tail rotors The main rotor blades of every helicopter of any size bend and twist under the influence of unsteady and nonlinear aerodynamic loads. Therefore, the most important mechanism of the rotor behavior is the cyclic blade flapping, which provides indirect control of the direction of the rotor thrust and the rotor hub moments. Hence, it is the primary source of the helicopter maneuverability. Body motion of a rotor blade Generally, the stiffness of rotor blades of small helicopters is considerably larger than of that of normal-sized helicopters. From this point of view, the helicopter blade is rigid and has two main movements: flapping and lagging. Let’s first look at the flapping motion: it results from the changes in aerodynamic loads (like lift) and can be described through the motion 13 2. Modeling, identification, and system analysis of its tip (tip-path plane models). The flap angle f is defined as the angle between the blade and the hub plane (see Figure 2.4). Figure 2.4.: Flapping angle and Coriolis force at a lagging movement ‘A blade which is free to flap experiences large Coriolis forces in the plane of rotation. The blade tries to lag in the direction of the Coriolis force. However, rotors with two blades can be constructed in a way that the lag- ging movement is not activated. Small helicopters, such as those described in Chapter 4, are constructed in this way. Finally, the blade can be feath- ered about a third axis, usually parallel to the blade span. Because of the stiffness of the rotor blade, this movement is ignored. The position of the rotor blade is estimated by the flapping angle and by the azimuth angle W, defined as the angle between the blade span and the rear center line of the angle. The derivative of the azimuth angle is the rotor speed (2. Actuator disc theory The calculation of the thrust T(t) requires information about the loading of the rotor blades caused by the airflow at the rotor blade. This in turn requires knowledge on the air velocity induced by the lift of the blades. The classical actuator disc theory allows for certain assumptions to be made about this topic which will be deduced here briefly. A detailed discussion on this part of actuator disc theory may be found in (Prouty, 1995, Chapter 2) and (Bramwell, 2001, Chapter 2). In the actuator disc analysis, the following assumptions are made for a simplified approach: The rotor has an infinite number of blades, so that it may be regarded 14 2.3. Physical nonlinear model as a disc. « The jump of pressure is uniformly distributed over the rotor disc. The velocity of the air flow through the rotor is constant. © There is no rotation of air in the flow. The velocity is only axial. Figure 2.5.: Actuator disc theory for hover, climb, and descent Three flight conditions are distinguished: hover, climb, and descent. During hover, the minimal induced velocity Vjo at the main rotor is achieved by using the Bernoulli theorem, _ T __ [Mhetitotat 9 Vo=Va55u ~\ 2p Su (2.22) where Spy is the plane of the rotor disc. ‘The minimal induced velocity Vzo is also termed thrust velocity. At a small disturbance or change in thrust, the induced velocity changes. This has to be considered in the model of the helicopter dynamics. A first-order model is used to model these dynamics: 1 Ve) = A, ; T(t) ) =Vr(t) + Vio—-—_ 2.23 ( 1 ) + 1 TRnetivotal g. ( ) Equation (2.23) is the basic equation for the dynamics occurring during a climb, but also for very slow descents. The induced velocity will be augmented by the vertical velocity. w(t) Viv = + (2.24) 15 2. Modeling, identification, and system analysis In vertical descent, some of the air recirculates through the rotor because of drag effects. This effect is known as the vortex ring state. In the extreme case, when w is equal to the induced velocity, there is no mass flow through the rotor and thrust cannot be developed. Pilots always try to avoid this condition. When the upflow velocity is increased to the point where it is higher than the induced velocity, the air slows down on when passing through the rotor and the condition is known as the ‘windmill bake state’. With reference to Bernoulli, this case can be characterized as follows: vp 5 View = 2 - + VAC (2.25) This state is also called ‘ideal autorotation’ and can be described through the induced velocity. Generally, autorotation is defined as the specific ro- tor state in case of an engine failure. Pilots use the last kinetic energy of the rotor system to land the helicopter. ‘The airflow during helicopter de- scent provides additional energy, which is stored in the turning rotor. The induced velocity in this case is characterized by the incoming airflow. The induced velocity for the tail rotor may be calculated in an analogous way. The yaw rate r and the lateral velocity v also affect the induced velocity. Later in this work, the connection to the lateral velocity will be neglected. at r(t) +d) ae View = + Vern(t)? (2.26) The dynamics at the tail rotor is also modeled as a first-order system: , 1 Tr(t) Vr t) = =— ( -Vrilt Vi 2.27) reall) = pe (-Veu() + Vero ee (227) where Tro is the side force of the tail rotor necessary for compensating the driving torque of the motor in hovering. The determination of this value is iterative. ‘The time constant for the induced velocity for the main and tail rotors, Ty, and Ty,, respectively, are values which have to be identified. In (Weilen- mann, 1994) they are estimated with the help of a high speed camera. The time constants range from 0.1 to 0.15 s. 16 2.3. Physical nonlinear model For the calculation of the lift force T, the local velocity at a certain blade radius is required, for instance by assuming a linear distribution of the induced velocity with the maximum at the tip of the blade. According to the theory of (Bramwell, 2001, Chapter 2.2), one may receive a better approximation by taking the square of the radius in the numerator. 3Vi0(Rin — Pin) 2(Rin — Tin) w= (2.28) Aerodynamic forces at the blade The rotor main force component along the perpendicular of the tip-path plane is usually called the thrust 7’. The component perpendicular to this axis and pointing rearward is called the H force. The third component, pointing sideways, is the Y force. Usually, the Y force is very small. The H force can be regarded as a rotor drag and has some influence on forward flight. These forces may be neglected at hover, whereas in forward flight they have a stabilizing effect (see (Bramwell, 2001)). For these reasons, the Y and H force are not part of the discussion in this work. However, the calculation of the H force is needed of for the determination of the rotor torque Q. For the calculation of the thrust force T, the lift force of a small blade element dL has to be evaluated. The rigging angle of incidence is the blade pitch © which can be expressed as a function of the rotor blade azimuth U with Q(W) = Oo — Arcos ¥ — Bi sin¥ (2.29) consisting of collective pitch (Qo, or Aj), the lateral (A;) and longitudinal (By) cyclic pitch. They are controlled with a swashplate mechanism (see Figure 2.8) Because the direction of the flow relative to the blade is quite small, the elementary thrust dT’ is approximately equal to the finite lift force dL of the small blade element of length J and at the distance r from rotor center. ar = 5caapU2(OW) +Up/Ur)ldr (2.30) The lift slope caa is constant along the blade. ‘The horizontal velocity of the air flow consists of the speed of the rotor and the velocity of the helicopter. Up = Or; — usin ¥ — veos ¥ (2.31) 17 2. Modeling, identification, and system analysis dL Ur 8 Up dD Figure 2.6.: Lift and drag at a blade section The vertical component of the air speed is coupled to the induced velocity and flapping angle. The roll and pitch movement may be neglected because they are assumed to be very small at hover with p =, q = 0: Up = —rif — v1 + uB cos U — vB sin ¥ +w (2.32) The total thrust integrated over the whole blade for two rotor blades may thus be expressed as follows: T(t) = af” ft" Feaap?(0(¥) +Up/Ur) ldr av (2.33) ‘The calculation of the drag force H was made in a similar way, integrating the drag force at a blade element dD over the whole rotor disc. The rotor torque @ can be deduced from the T and H forces (or directly form dD). In Appendix B.33, the deduced value for the H force and the rotor torque Q is described in detail. At the tail rotor, the same calculation is made for the thrust force Tr. Also, the drag moment at the tail rotor Qr is calculated (see Appendix B.34). Flapping movements The mechanism of blade flapping provides indirect control of the direction of rotor thrust (see equation (2.32), hence it is the primary source of ma- noeuvrability of the craft. The blade flapping is partly under the control of the pilot and depends strongly on the local incoming air flow. 18 2.3. Physical nonlinear model The mathematical description of blade flapping is deduced by an angular momentum about the rotor blade, equal to the aerodynamic moment of the main rotor. Rp Mr aah rdv (2.34) This derivation is found in literature (e.g., (Johnson, 1994), (Prouty, 1995)). In order to understand, the resulting equation of motion is regarded in a reference frame which is rotating with the rotor (V = t). In (Bramwell, 2001, chapter 1.6), the flapping equation is a differential equation of second order, 8B 1B yp yn oc gu? t aap 18 = 2beos (#) — Gsin(w)) oat + Z(@(W) + Geos () + psin (Y)) if the velocities of the helicopter are neglected. ‘The Lock’s inertia number is 7. It is the ratio between the inertia and aerodynamic forces. The values p and @ are the angle ratios relative to the rotor speeds Q, p/ and q/Q. The Lock number is calculated by (Johnson, 1994) as: pCaaltcaveR* a (2.36) Leave Taking a typical value for 7 of 6 for a normal-sized helicopter results in a value for 7/8 of about 0.7. In reference to the damping of equation (2.35), it demonstrates that the time constant in terms of the azimuth is } of a revolution. Virtually, any force which affects a rotor blade at a certain azimuth generates a deflection of the rotor blade 90° later. For these reasons, it can be said that the flapping motion is very heavily damped. When the flight condition is steady, the flapping angle f can be expressed by a Fourier series = ap — a1 00s (Mt) — by sin (Mt) — az cos (22) : (2.37) — basin (2Mt) — ag cos (3M) — bs sin (3NL).... because the blade motion must be periodic for fulfilling (2.35). In these terms, ag represents the coning angle and a; and b; represent the longitudinal and lateral flapping. According to (Prouty, 1995, Chapter 3), the higher harmonics ap, by, a3, ... are relatively small and have little effect on rotor thrust and torque. 19 2. Modeling, identification, and system analysis Utilizing the substitution of @ (2.37) in (2.35), the differential equation is transformed in the terms of a, a, and b;. According to (Chen, 1980) the rotor flapping response can be accurately approximated by a first-order system, and the blade accelerations & and b can be neglected. After equating coefficients of sin VW and cos and transforming them into a non-rotating coordinate system with Y = 9, the flapping equations read as follows: I by A Tit = a1 — 5 = + an — Bi- tm (2.38) re a Bi _ 9.9 Tmb1 = —b1 + Qm tom + Ai —Tmp (2.39) with small simplifications and with tm as definition of the time constant of a teetering rotor (see (Prouty, 1995). 16 may (2.40) For a small helicopter, a value of 0.03 s is achieved at a constant rotor speed of 1500 rpm. The relevant frequency is thus above 30 Hz. The blade dynamics can be reduced and damped by the servo rotor (Bell-Hiller stabi- lizer bar). Therefore, the equations for the flapping of the main rotor are omitted. In a steady flight condition the flapping angles can be calculated as follows: ao = 122 a =-Bi +2 —rq by =Ar- 2-1 (2.41) ‘The term 7/8 multiplied by the (collective) pitch angle Qo represents the constant flapping angle ag. The blade traces out a shallow cone, which is why the angle is called the coning angle. The blade tips trace out the ’base’ of the cone, which is called the tip-path plane. "The flapping of the blade causes a torque on the helicopter, proportional to the flapping angle in 2? and y coordinates, multiplied with flapping hinge offset los and the stiffness of the attachment ko: Me(61) = ko 2b, (2.42) li M,(a1) = ko-Ghar (2.43) Stabilizer bar and geometry of cyclic and collective pitch control Due to the high rotor speed, small helicopters have dynamics in a higher frequency range than helicopters of normal size. In order to help them fly, 20 2.3. Physical nonlinear model Figure 2.7.: Interpretation of flapping and definition of the tip-path plane a Bell-Hiller stabilizer bar is attached above the main rotor. The phys- ical interpretation of the stabilizer bar is to increase the damping of the helicopter attitude dynamics. In addition, this attaching changes the way the helicopter responds to the pilot commands and to atmospheric distur- bances. The stabilizer bar can be interpreted as a second rotor, with two simple short blades, called paddles. They are fixed to the rotor shaft and have no hinge possibilities. Like the main rotor, they are engaged by the swash plate mechanism. The swash plate consists of one plate which does not rotate with the shaft but can be tilted in any direction by the servos. Clearly, if the swash plate is tilted, the rotated paddle chord remains parallel to the swash plate. Because the blade rotates with the shaft, cyclic engaging takes place in the direction of the tilted plate (see Figure 2.8). Through the cyclic feathering of the paddles, the auxiliary rotor is also flapping. This flapping angle in turn changes the cyclic feathering of the rotor blade. The main rotor receives the cyclic pitch (41, By) from the swash plate (A1,, Bi,) and from the paddles (a1p, bip)- At = kparbip + ks2rAis (2.44) Bi = kpar@ip + ksarBis By the levers of the rotor head mechanism, the stabilizer bar can also be regarded as a mechanical amplifier of the command of the swash plate. The angular transformation in the equations (2.44) is actually a simplification, because the gains kyor, ks2r, and kszp depend on the collective angle Ans. In addition, the angular transformation (2.44) is in fact nonlinear and the levers have a certain slackness. However, measurements have shown that the gains kpor, Kear, and ksgp are always quite equal. ‘The angular trans- 21 2. Modeling, identification, and system analysis Figure 2.8.: Swash plate (a) with servo rotor. The levers of the paddles (b) are connected with the swash plate through (e) and with the mounting of the main rotor blades (c) through (f). ‘The main rotor blades and the swash plate are also directly connected with (d). formation (2.44) is also transferrable to other constructions of the rotor head. The dynamics of the flapping angle can be calculated analogously at the main rotor. Because of the stiff attachment, there is no cone angle. Tpitplt) = Aap) + heap Bas(t) + PO — pat) (2.48) robip t) = ~bip t) + Baap Are) ~ $2 — appt (2.46) with a definition of the time constant Tp, including the Lock number for 22 2.3. Physical nonlinear model ‘Table 2.1.: Measured values of the rotor parameters gain | range _| value nearest operation point Rpor | 0.5...0.8 0.73 kor | 0.3...0.7 0.61 kesop | 0.9...1.5 1.28 the paddle 7p. ,-_16 2 yp Qae (2.47) For the time constant 7) a value of 0.09 s is calculated. The dynamics of the servo rotor have to be considered in the mathematical model because of the low eigenfrequency. 2.3.3. Engine model Normally the drive for the helicopter motor is a combustion or electrical engine. In comparison to combustion engines, electrical motors have the advantage of generating fewer vibrations. However, free flight helicopters with electric drives have a weight problem because of the amount of cells needed for a long flight. Mostly, the engines are optimized for high power at minimum weight. The dynamics of the motor are quite important and have to be taken into account, because the nonlinear model in this work needs the motor torque and rotor speed. Mood k Mi we {a}. —_{ 2 Kumot Figure 2.9.: Motor model of an electric motor 23 2. Modeling, identification, and system analysis Figure 2.10.: Output signal of the identification experiment, measurement and simulation of the identified model An electrical drive (as the one described in Section 4.4) can be modeled as a second-order system (see Figure 2.9). The motor model has five unknown parameters in two equations. . 1 1 1 Te eae ee a aaa oer (2.48) 9 = ka(Ma — kt Mroad) (2.49) Identification methods like the prediction error method (PEM) are used for estimating the parameters. The input signal used in an identification experiment can have a significant influence on the resulting parameter esti- mates. The motor system is not very complex and can be identified without any specific problems. For this reason, the input signal can be chosen as pleases. In the first step, the parameters are identified without the load influence. A chirp signal (an sinus with increasing frequency) is given as input signal on the control voltage (see Figure 2.10). Similar measurements as in Figure 2.10 are used for the identification of the parameters. In order to the identified model, a step response with the real and the mathematical model is calculated in Figure 2.11. The dynamics of the electric motor is quite high (w. = 73 Hz). A combustion engine, which has a similar structure of the model, would have a smaller bandwidth. 24 2.3. Physical nonlinear model Table 2.2.: Values of the motor model parameters variable name | identified value | variable name | identified value Ta 0.0134 ko 127.76 ka 40.131 k, -0.003 time (s) Figure 2.11.: Output signal of the identification experiment The influence of Mioaa can only be evaluated in flight. It was determined by the mathematical model and the operation point of the tail Rotor Ar. ‘The correction factor ky, could be put on 1 in (2.49). 2.3.4. Coupling of all forces and torques The equations (2.13) through to (2.49) represent all the forces, torques, and relationships required for the nonlinear model. The main part are the equations of the rigid body motion with the forces and moments. The forces are gravity and drive forces. F(t) = Fox(t) — T(t) sin (a1) Fy(t) = Foy(t) + T(t) sin (b1) + Tr(t) (2.50) F(t) = F2(t) — T(t) cos (a1) cos (b1) For the lift force 1, the induced velocity V; is needed, and for the compo- nents in the body frame, the angle of tip-path plane a and b, which are calculated from the flapping angle of the paddles ai, and bip. The torques of the loads on the main and tail rotors are the inputs for the 25 2. Modeling, identification, and system analysis motor. Mr(t) = Qr - nrirQ(t) 1 1 Miooa = —Q - —Qr 1M nT The drive torque is calculated tn the motor equations. Marive y(t) = Mr Marive 2(t) = 1m Mmotor(t) The resulting torques in all directions can be calculated as follows: L(t) = Mz — 2m T(t) sinb1 — 2t Tr(t) M(t) = My + Marivey(t) — zm T(t) sinal +.2m T(t) cos a1 cos b1 N(t) = Marivez(t) — 2m T(t) sin bl + at Tr(t) All nonlinear equations are listed in Appendix B.1. 26 (2.51) (2.52) (2.53) (2.54) (2.55) 2.4, System description for control design 2.4. System description for control design The systems of Section 2.2 and 2.3 describe the aerodynamics of the heli- copter with ordinary differential equations (ODE). The nonlinear systems can be presented by a state space description of the form f(x,u); y=g(x,u) (2.56) with m inputs (vector u) and p (measurable) outputs (vector y). The inter- nal description is given by the n state variables (vector x). The following order is chosen for these system vectors: state vector of the model with = [2° y? 2 FOb wv ap ary bip wr Q Ma Vi Vir * i IK Ie input vector of the model with | w= [Bis Ais Au Ar Umot)” y measurement vector of the model with Y= [2° y' 2° pO duvgpwrQ)’. The measurement vector with the selected signals is available on UAV sys- tems. Filters may be used to improve the quality of these signals. 2.4.1. Linearization ‘The nonlinear system can be linearized in an operating point by calculating the Jacobian matrix which presents the full derivative matrix of f at x (or u). First, the operation point for hovering has to be defined. Hence the following state variables are set to (2.57) The velocities and rates are zero, the position and yaw angle can have any value, and the rotor speed is in a typical range for model helicopters. Other state variables which belong to the stationary point of hovering, particularly the Euler angles @ and ¢, are found by solving the following equation system. 0=f(x,u) (2.58) 27 2. Modeling, identification, and system analysis Table 2.3.: Values of the operating point for hovering variable name | operating value | variable name | operating value 9% ~0.392 deg Biso 0.376 deg, 0 -5.399 deg Aiso 0.042 deg A190 0.532 deg Amo -9.692 deg bipo 0.060 deg Aro 11.15 deg Vi0 4.241 m/s Vri0 6.348 m/s Mao 0.834 Nm Muo 5.549 Nm It is solved numerically for x and u by MAPLE. The starting values for the solver have to be sufficiently close to the solution for the algorithm in order to succeed. Table 2.3 shows the state variables for hovering. They have realistic values, which shows a correct physical behavior of the developed mathematical model. The helicopter has a roll angle for compensating the horizontal force of the tail rotor. Also, the induced velocities of the main and the tail rotor are verified with measurements and are in the same range. In fact, the calculation of this operation point is related to the estimation of the trim point of the helicopter for hover. It can be compared to the test results in Chapter 5. The Jacobian matrices can now be calculated with the operation point for hover. The state space form of the linear system is =Ag+B a+Bu (2.59) Cx+Du A= SE uw ERS Ba SE oe eR’ ‘3 a (2.60) 135, = 9 13x18 ag C= Fileomo €R Gy louo ER where the notation for the deviations from the operation point is without a (eg., x instead of Oz). The matrices are listed in Appendix C.2.1. In order to study the linear system to be controlled, the first step is to scale the plant. Scaling is quite important because it makes model analysis and controller design (weight selection, see Chapter 3) simpler. The design engineer has to make a judgement about the performance of the system, 28 2.4, System description for control design Scaling factors for inputs and outputs inputs and outputs | scaling value Zz 0.2m 04,0 0.2 rad u,v, w 0.2 m/s Par 1 rad/s Q 57 Am, Ats Bis 0.02 rad Ar 0.25 rad Umot 2V arp, bip lrad Vi, Vir 1 m/s 1Nm even though the design process has just started. In order to achieve this goal, decisions have to be made about the expected changes of reference value and about the allowed magnitude of each input signal. For hovering, typical deviations on the inputs and outputs are presented in ‘Table 2.4, based on a small helicopter with a length of 2 m. In principle, it is not necessary to scale the state vector 2, because the frequency response is independent of the scaling of the interior state vector. G(s). = Ty 'G(s)Tu = T; (C(sI — A)'B+ D)T, (2.61) However, for a better condition number of the matrix A and in case the state variable (or the eigenvectors) are studied, it is advisable to scale the state vector z as well. A=T;'AT, B=Tz;'BT, C=T)*CT; D=Ty'DTu 2.62) - y ¥ The diagonal matrices T,, Ty, and T,, contain the scaling factors in the correct order of inputs and outputs in the diagonal. For the investigation of linear systems, singular value plots are used when jt comes to the consideration of frequency domain issues. Singular values allows some generalizations of SISO systems (Bode plots) to MIMO systems (see (Zhou et al., 1996)), unfortunately without getting any information about the phase. The maximum (and the minimal) singular value is defined by [GGw)ull a(GUGu) = may 9((40)) = min, EDA hal (2.63) 29 2. Modeling, identification, and system analysis 100 ~ a a g 80 4 = = 2 60 =o . & Ae NY 3 40 -50 S 8 5 0 s 0S 0 3 10" 10 10 107 10 10 frequency (rad/s) frequency (rad/s) Figure 2.12.: Singular values and condition number of the linearized plant In the singular-value plot, however, only the maximum singular value is of interest. It is a good measure of the maximal gain of the system cor- responding to the input direction, but it may lead to very conservative results. A good measure for the quality of the scaled model is the condition number fk. It is defined as the ratio of the system’s maximum and minimum singular value (GGw)) a(GGw)) If «is large, the gain of the system matrix heavily depends on the direction of the input. For control purposes, a system with a condition number close to one is desirable. Otherwise, it is not possible to define an exact bandwidth of the system and it is difficult to set specifications for the singular values. (2.64) The whole helicopter plant is ill-conditioned despite the scaling (see Figure 2.12). By changing the scaling matrices, a better condition number is not achieved and the scaling values are not obvious for control design anymore. Because the model includes the model of the motor and the dynamic of the helicopter, the singular values for low frequencies spread and the condition number increase. To achieve a better starting point, the system is split in two main systems: the horizontal and the vertical movement, respectively. These two main 30 2.4. System description for control design 100 80 x a g7 Vee : 3 Sool s 2 50 / \ § 5 / g g : gop \ 5 Feo \ 5 0 ‘S c) 5 0 S 10” 10 10 10” 10 10 frequency (rad/s) frequency (rad/s) Figure 2.13.: Singular values and condition number of the plant parts of the system are only connected by small couplings. The connection can be confirmed by looking at the steady-state gains for different frequen- cies. Because the systems describe different physical structures, it makes sense to identify and to develop the controllers for the two subsystems sep- arately. They can later be combined using the design parameters of the individual designs. Another problem lies in the linearization of the equations for calculation the position. The transformation of the earth-fixed to the body-fixed velocities is a part of these equations which is a nonlinear transformation with no specific operating point. Through the linearization of the equations, the transformation of equation (2.1) is reduced to a useful range of +10 deg for every angle. The transformation is problematic particularly for the yaw angle w, because the helicopter may be aligned in any azimuth direction. For these reasons, the identification and the control design will not include the state variables of the position and the yaw angle. If the system is reduced by these state variables, the system shows different singular values (see Figure 2.13). The condition number does not change much. The next step will be to look at the subsystems. 31 2. Modeling, identification, and system analysis gain (4B) frequency (rad/s) frequency (rad/s) Figure 2.14.: Singular values and condition number of the zy plant 2.4.2. Horizontal motion The motions in the horizontal plane are the forward and sideward move- ments. Both dynamics are coupled physically by the main rotor. ‘The state vector x, the input vector u, and the measurement vector y for this subsystem are hor = (0 UVa arp brp]” Unor = [Bis Ars)” (2.65) Ynor = (bu 4 pl” The roll and pitch attitude of the helicopter system is described in this first part and shows an interesting behavior for control design. The linear system is state controllable and state observable. As the singular values in Figure 2.14 show, the condition number is small over the whole frequency range of importance. The system behaves as a second order system, with a resonance peak and a subsequent roll-off. The resonance indicates an oscillation in the attitude of the helicopter. This is also shown by the poles of the system on the imag- inary axes (see Table 2.5). The subsystem is unstable because it has two poles on the right half plane. For the physical interpretation of the eigen- values, the eigenvectors are regarded. The system’s dynamics depends on its eigenvalues (poles) and its eigenvectors. The eigenvalues determine the time constant or frequency and damping of each mode, and the eigenvec- tors determine how much each state variable of the system contributes to 32 2.4. System description for control design Table 2.5.: Poles s; to s7 and zeros of the xy plant poles zeros: .0633 + 21.7604i | none — 23.0633 — 21.7604i —12.4049 + 16.0214¢ —12.4049 — 16.0214¢ 0.0148 + 0.4668¢ 0.0148 — 0.4668¢ —0.0539 + 0.43047 —0.0539 — 0.43047 a given output. The eigenvalue associated to its eigenvector offers valuable clues to helicopter characteristics (see (Padfield, 1996, Chapter 4)). According to Figure 2.15, the eigenvalues describe the following character- istics: 51,2 = —23.0633 + 21.7604i: Naturally, the fastest pair of poles is at the roll rate p. On this axis, the helicopter has the smallest moment of inertia, To control the roll movement, a controller with an high bandwidth is necessary. 83,4 = —12.4049+16.0214i: The pitch rate is connected with this pair of poles. It has also quite a high dynamics, but because of the length of the tail, the moment of inertia affects the pitch movement. 85,6 = 0.0148 + 0.468%: The unstable pair of poles has its primary components of its eigenvector in the lateral and longitudinal velocities. Small components of the eigenvector are also in the roll and pitch angle. 87,3 = —0.053940.4304i: This damped mode mainly affects the lateral and the longitudinal velocities. The run of the curve of the the singular values (see Figure 2.16 for the x direction) in the different paths shows that the oscillation is found in the angular dynamic and in the velocity. The same behavior is found in the pitch and in the roll angles. This confirms the imaginary poles of the plant. 33, 2. Modeling, identification, and system analysis 1 1 05 05 Bo so 0/04 05 05 -1 -1 -1 0 1 -1 0 1 1 1 + 0s 0s Eo 0 05 -05 -1 -1 -1 0 1 -1 0 1 Figure 2.15.: Components of the eigenvectors for the eigenvalues 51, 83, 85, and s7 in a phase plot; the components are marked with x for 0, 0 for ¢, + for u, * for v, © for q, and ¥ for p a $ singular value (4B) ° singular value (dB) b $ Z 10 10 10 10” 10 frequency (rad/s) frequency (rad/s) Figure 2.16.: Singular values of the output of @ (left) and u (right) from the single (--) and combined (-) inputs of By, and Ai. 2.4. System description for control design 30 10 ——= 20 Tt \ ze a 2 10 g 3 5 eS £6 do 5 2 -10 54 aay 2 2. 7 2 = 2 10” 10° 10 10" 10° 10 frequency (rad/s) frequency (rad/s) Figure 2.17.: Singular values and condition number of the zw plant 2.4.3. Vertical motion The vertical and the yaw motions are strongly coupled in the model. The collective pitch of the main rotor causes lift and creates a reaction in the rotor speed because of the changing load. A sequential result: is the fact that the yaw motion also changes as a reaction to the decreasing engine torque. For the vertical subsystem, the state vector is yep = [wr & Ma Vr Vir|? In most: of the helicopters, the motor speed is controlled separately. It is assumed that the rotor speed is constant and the coupling between the state variables of the motor model and the remaining state variables can be neglected. At the helicopter test bench described in Chapter 4, the dynamics are also in a range outside of our interest. The state variables, inputs, and outputs of the vertical subsystem are Byer = (wer Vi Vir|” Uyer = [Am Ar]” Yee = [w rl” (2.66) Figure 2.17 shows the singular values with a low-pass behavior of first order. 35 2. Modeling, identification, and system analysis The system is also state observable and controllable, and it has all poles in the left half plane. As for the horizontal case, the eigenvalues and the eigenvectors are calculated and investigated. © s; = —33.536, so = —23.066: The poles describe the dynamics of the induced velocities for the main and tail rotors. © s3 = —1.842: The eigenvector of this pole has components in the induced velocity of the main rotor and in the vertical velocity. © sq = —0.852 Alike to the pole s3, the dynamic describes the coupling between the yaw rate and the induced velocity of the tail rotor. 36 2.5. Identification of deterministic models 2.5. Identification of deterministic models 2.5.1. Identification methods Identification is the second step after the mathematical modeling of the plant used for the control design. If the structure of the system is known, the identification procedure is confined to identifying the parameters of the plant. Identification is an established field in research, and the theory and methods used are well described in (Ljung, 1999) and (Séderstrém and Stoica, 1989). The approaches in this work are based on these books. For the helicopter plant, the identification procedure has more difficulties than in the regular case. The helicopter is an unstable system, and in order to collect the necessary flight data, a pilot or a controller is used for flying. The identification step is considered in closed-loop instead of the closed-loop case. The fundamental problem with data logged in a closed loop is the correla- tion between the immeasurable noise e(t) in the measurement signal y(t) and the input signal to the plant u(t) which is calculated with the con- troller K(t). If input and noise are correlated, the model of the plant and the model of the disturbance cannot be separated. In addition, the con- troller produces his own dynamic in the signal u(t) which perhaps does not cause the important characteristics of the plant. In this case, parameter identification is very difficult. However, it is sometimes necessary to use closed-loop data due to safety or economic reasons. For closed-loop identification, there are many different methods. A part of them are described in (Forssell and Ljung, 1998). Four different approaches can be classified. Direct approach The most straightforward approach is to neglect the feed- back. ‘The data are treated as if it is logged in an open-loop environ- ment. The command signal for the plant u and the measurement vector y is used for the parameter identification. y(t) = Go(t)u(t) + Hel) with G, as the plant to be identified and H(t) as a filter for the white noise e(t). The method is comparable to the classical identification 37 2. Modeling, identification, and system analysis approach, and in (Séderstrém and Stoica, 1989, p.389) the identifia- bility is warranted if a high-order noise-free (or non linear) feedback is used. The identification step will be improved through an additional (external) input u,(t) on the command signal u(t). Indirect approach The whole closed-loop transfer function is identified in this method. It is assumed that the external command signal u,(t) is measurable and the regulator K(t) is known. Hence, the open-loop system parameters from the closed-loop model are obtained, using the knowledge of the feedback. ult) = Ga(tur(t) + Heel) Ga(t) is the closed-loop transfer function, which has to be identified, and H, is a fixed noise model. Joint input-output approach The third approach is the joint input-output identification. The input u(t) and the output y(¢) of the plant are jointly regarded as the output from a system driven by the external command signal u,(t) and the noise e(¢). It is assumed that the feedback controller K,(t) is known and linear. The extended plant is (28 ]=[°20 Jose where S,, is the sensitivity (see Section 3.2.1) of the closed loop and Ho is a complex prefilter consisting of 5, G.(t), Ks(t), and H(t). Projection approach This method is similar to the joint input-output ap- proach, but instead of the true controller a frequency weighted approx- imation of the controller is used. The identification process consists of two steps. 1. Identifying the sensitivity function S, in the form of a non-causal finite impulse response filter (FIR filter). "Non causal” means that future data is also considered. k=M ti(t) = Su(Qur(t) = > s*(@)ur(t) M00 fi This can be viewed as a projection of the plant input u(t) onto the external input u,(t). 38 2.5. Identification of deterministic models 2. The signal a(¢) is constructed and the open-loop system using the output error model y(t) = Ga(t)a(t) = G.()S' (hu, (t) is identified. In this approach, there is no information needed about the feedback controller. Because the controller has certain nonlinear components which cannot be included in the indirect or joint input-output meth- ods, this procedure has a clear advantage. All approaches need the prediction error method (PEM) which is included in Marta. If these approaches want to be applied, the plant and the signals have to be preprocessed and a first approach by a least square method is worked out. Preparatory work Through linearization of the nonlinear model in Section 2.4, a linear model for the motion of the helicopter is derived investigated in Section 2.5.2. The models correspond to the general linear state-space structure, with an additional model of noise K(@) a(t) = A@)x(t) + B(O)u(t) + K(O)e(t) u(t) = CO)x(t) + D@)ult) + e(#) with © as a vector containing all of parameters to be identified. The term K(@) is the Kalman gain of a stationary Kalman filter (see (Forssell and Ljung, 1999, p. 98), or (Séderstrém and Stoica, 1989, p. 157)). (2.67) The modeling of the noise can be neglected by filtering the measurement data. In this case, the Kalman gain is set to zero. This is very useful for the direct approach because the gain K(@) is difficult to estimate and it can influence the identification process. The data is filtered with a 10-point averaging finite impulse response (FIR) filter. SS vise Yi siterea =) EEE k#0 (2.68) kes After the data was filtered with the medium filter in the forward direction, the filtered sequence is reversed and run back through the filter. The filtered signal is without any phase loss. 39 2. Modeling, identification, and system analysis The filter is used on the whole data set, which is used for identification. It is sampled at a specific time T, and brought to the form Z=[y(1)..-y(N) u(1)...u(N)] (2.69) consisting of measured input u(t) and output y(t) signals for a number of sampled points NV. The mean value of the data is subtracted and the new data can be used for the discrete-time model which is deduced from (2.67). In order to receive a good result in identification, the helicopter has to be moved so that the dynamic is visible in the measurements. There are two possible inputs for achieving this result: (a) ‘The reference signal r(t) is used for an input with a high frequencies. In this case, the reference values for velocity in the body-fixed frame (Urefs Vref; ANd Wyep) are provided with dynamics which are excited from steps or a white noise. (b) A disturbance signal (white noise) is superposed to the input signal of the plant u. The command signal, then includes not only the dynamics of the controller, but also a uncorrelated part. The data in this work are measured on the test bench described in Chap- ter 4, and for the experiments of case (a) the helicopter flies with the stabilizing Ho. controller described in (‘Tanner and Geering, 2001). The perturbations in case (b) are implemented with the LQ and H.. controllers introduced in Chapter 3 Least square method In a first step, in order to get some start values for further identification, the least square method is used. The model of the system has to be an autoregressive with extended input (ARX) model, which can be related to a system without any unmeasurable modes. #(t) = A@)a(¢) + B(O)u(t) (2.70) The system has to be identified as a time discrete model. The use of the method of Euler integration with sampling time T’ changes the system description to Tree1 = (14+A-Ts-)ee + B-Ts-u (2.71) 40 2.5. Identification of deterministic models With the equation error ¢(t), this form can be brought to Y = © Ors +¢(t,0) (2.72) with Y containing all the measured and @ all the regression parameters. The least square estimate of © is defined as the vector © that minimizes the loss function 1” V@1s) = 5 ye (2.73) ti The vector of parameters @ is here calculated with O15 = (@7)"'8"Y = [(1+ATs) BTs] (2.74) Prediction error method The prediction error method (PEM) is described in detail by (Séderstrom and Stoica, 1989, p. 185ff) or (Ljung, 1999, p. 197M). ‘The method works with a structure of the model as equation (2.67), transformed in a time discrete system. The modeling error is e(t,®) = y(t) -— 9(t, ©) (2.75) that is the starting point for the least square optimization. By defining a monic prefilter L that can be used to enhance certain frequency regions, the modeling error is extended with ev(t,©) = L(k, @)e(t, @) (2.76) The definition of the loss function Vy includes the modeling error and a symmetric, positive definit weighting matrix A N iy (O, ZN) = vie F(t, O)A~ter(t, O) (2.77) The estimated parameter vector is the minimum of the Viv Opem = arg minVy (0, Z”) (2.78) The application of the prediction error method was programmed in MAT- LAB and includes the following steps: 1. The structure of the model should be correctly defined. Also, the unknown coefficients which may first may be kept constant should be correctly listed in the MATLAB program. It makes sense to identify the coefficients step by step, depending on the affected dynamic. 41 2. Modeling, identification, and system analysis 2. The system is defined with the continuous-time form (2.67) in a MaT- LAB file which delivers the system matrices for a set of input variables (see (Lennart, 2001)). For better numerical stability, the parameters are scaled with the expected values. The input variables have the same length as the input of the identified coefficients and have a value of about one. The required system matrices can also be calculated in the discrete time format. 3. The pem function in MATLAB requires the state space form and decides whether the proposed time discrete form will be used or the sampling will be done by an internal function. The parameter vector will be estimated by optimizing the prediction error. Closed-loop identification in the prediction error framework The PEM described in the sections above is the base of the identifica- tion approach. The direct approach amounts to applying it directly to input-output data, ignoring possible feedback and it coincidences with the standard (open-loop) PEM. This approach is used first and delivers the majority of all identification results. In an additional step, the projection method is used and the results are compared with those of the direct ap- proach. While the direct approach can use the experiments from (a) and (b), the projection approach is limited by theory to experiment (b). The indirect and the joint input approach are not used in this work. Because of the large system used for identification, the predictor in the PEM is often unstable and a solution is rarely obtained. 2.5.2. Model structures and identification The linear model of the helicopter obtained on the basis of the nonlinear model has a lot of coefficients in the system matrices. The system for the horizontal movement alone has 36 coefficients. The question is whether all of these parameters are need to be identified. For this reason, the struc- ture of the linear model and the influence of the odds parameters on the helicopter dynamic are investigated. Again, it is necessary to divide the system into a horizontal and a verti- cal movement. In Section 2.4, it is shown that the system has no strong 42 2.5. Identification of deterministic models coupling between these two subsystem parts. The number of parameters will be reduced by the division into the two subsystems, and the remaining derivatives are named according to (Padfield, 1996). System in the horizontal movement The linear system for the horizontal movement has more state variables than the system of the vertical movement. So there are more parameters to identify and the equations are regarded separately. The first two equations are the calculation of the attitude angles by integrating the angular rates. 6=q b=p (2.79) If the helicopter is not exactly horizontal, equation (2.79) does not corre- spond to the linear approximation of (2.2), but for small angles there are not any additional deviations. So there are no unknown coefficients in these equations. The equations for the velocity and the angular rate describe the main dy- namic of the system. Xo + Xuut Xyv+ Xqq+ Xpp + Xerpaip + Xtipbip + XiBis + X2Ais 6 =Yo0+Yoo+ Yuut+Yout Yq Ypp + Yayy@ip + Yoyybip + Yi Bis + YoAts Myu+ Myv + Mgq+ Mpp + Ma,,p + Moy,bip + MiBis + M2Ais p= Lyut Ly + Lag + Lpp + Lay dip + Loy, dip +L, Bis + L2Ais (2.80) The coefficients for each state variable parameterize the model. To avoid an over-parameterized model, the number of parameters can be reduced further through the following assumptions and conclusions: Xq and Y, are in the model of the rigid body equation the coefficients of the gravity force Fy. For small angles, the coefficients of 9 and ¢ do not change in the 2 and yz coordinate. The division by total mass m results in —Xg and Yi being the gravity acceleration g. © The cross coupling of the forward and sideward velocities can be ne- glected so that X, and Y,, are equal to zero. 43, 2. Modeling, identification, and system analysis The velocity changes because the deviation of the operation point in the Euler angles of the helicopter are equal to the acceleration. The derivatives of the angular velocity have almost no influence. So Xq, Xp, Y; and Y, can be neglected. The equations of the servo rotor are the same as in the nonlinear model and can be used here because they are linear, too. frp = (dip + G+ keap Bra) — 4 . ae (2.81) bap = B(—bip — Go + Kap Ata) ~ B with k = 1/7 which is the inverse of the servo rotor time constant. The system for the horizontal movement is now considered in two different ways. Full information model A linear system without any immeasurable state variables is called a full information model. If the model has such a structure, a state feedback controller can be implemented (LQ design, Section 3.2). To reach a full information system from the equation (2.79) - (2.81), the two state variables for the servo rotor are dropped. Additionally, the number of system parameters is reduced by dropping some of the coefficients without changing the poles of the model described in (2.80). o=q =p = 90+ Xuut XiBis b=9¢+You+YoAis q = Muu + Mvv + Mqq + Mi Bis + M2Ais p=Lyu+ Lyv t+ Lppt+ Li Bis + LoAis (2.82) The cross couplings for the inputs Xz and Y; are dropped, because the coefficients are very small. In addition, the coupling parameters for the angular rates M, and Lq have no influence of the poles or eigenvectors, so they are dropped, too. Because of the inputs to the main rotor blade over the swash plate and the servo rotor, the coefficients of the inputs will be different compared to the coefficients of the model received through linearization. 44 2.5, Identification of deterministic models 415 420 425 time (s) time (s) 6 20 9 (deg) 20 415 420 425 415 420 425 time (s) time (s) Figure 2.18.: Measurement and simulation of the identified model of the zy plant for the full information model; 6 (upper left), q (upper right), ¢ (lower left), p (lower right) 45 2. Modeling, identification, and system analysis 08 0.6 04 u (ms) “415 420 425 “415 420 425 time (s) time (s) Figure 2.19.: Measurement and simulation of the identified model of the xy plant for the full information model; u (left), v (right) The identification is successful as can be seen in Figure 2.18. The valida- tion of the identified model is made with a different data set. The Euler angles 6 and ¢ are well following the measured signal, although they are the integrated signals of the angular rates q and p. The curves of the angular rates are also well followed. The run of the velocity curve in Figure 2.19 shows that the simulation has an offset in comparison with the measurement signal, although the dynamics in the measurement data are included in the identified model. These characteristics can be reduced to a model error in the Euler angle calculation. The missing servo rotor model also causes an additional offset. The roll and longitudinal movements are better identified than the pitch and lateral movement. The coefficients are listed in Table 2.6. Model with servo rotor The model with the servo rotor is identified based on the linearized sys- tem from Section 2.4.3. Some parameters are dropped in the continuous comparison of the poles of the system with the poles of the system with all 46 2.5. Identification of deterministic models coefficients. b=4 o=p 7 99+ Xuu+ XiBis b=gh+ Yout YoAis G= Muu+ Myv + Mgq+ Maaip + Mobip + MiBis + M2Ais (2.83) b= Lyu+ Lyv + Lyp + Laaiy + Lobip +1) Bis + LoAis day = b(—aip + ns + keapBis) 4 inp = k(—bip — na + keopAts) — P The dynamics of the angular rate are coupled with the flapping angles of the servo rotor a1p and b1p. The input to the dynamics of the helicopter to the main rotor blade is the swash plate and the servo rotor. This dupli- cation of influence causes certain difficulties in the identification, because the parameters of the servo rotor in the dynamics of the angular rates (Mu, Mp, La, and Ly) are related to the input coefficients of the swash plate (L1, Lz, My, and Mz). The experience of the identification from the full information system iden- tifies the parameters and validates the model (see Figure 2.20). The simu- lation results for the Euler angles and the angular rates match the experi- mental data. The dynamics are closer followed than in the full information model. An increase of the identification quality is seen in Figure 2.21 in comparison with Figure 2.19. The run of the measured velocity curve shows a better matching with the simulation than in the full information case. There is also no discrepancy between the longitudinal and lateral movement anymore. The identified coefficients (see Table 2.6) are closer to the coefficients of the linearized model. While a PEM step is executed in the identification procedure, the numbers of identified parameters are always limited so that not all of the parameters can be identified in one step. In particular, the parameter of the servo rotor k cannot be identified with other parameters because of the duplicity of input through the servo rotor and swash plate. AT 2. Modeling, identification, and system analysis @ (deg) 5 415 420 425 time (5) time (s) (deg) P (deg/s) 4 -20 41s 420 425 415 420 425 time (s) time (s) Figure 2.20.: Measurement: and simulation of the identified model of the zy plant with servo rotor modeling; 6 (upper left), q (upper right), @ (lower left), p (lower right) System in the vertical movement The system in the vertical movement has less state variables than the sys- tem for the horizontal movement, and it is stable. The PEM works better with stable systems than with unstable systems and the iteration converges faster. This could be noticed in the identification procedure for this sub- system. 48 2.5. Identification of deterministic models 0.2 02 u (m/s) v (m/s) . = -03 415 420 425 415 420 425 time (s) time (s) Figure 2.21.: Measurement and simulation of the identified model of the zy plant with servo rotor modeling; u (left), v (right) Again, the linear model from Section 2.4.3 is used as a reference model: w= Zyw + Z2v,Vi + ZAM t= Now + Ner + Nv, Vi + NvppVir + NaAm + NaAr Vi = koww + koVi + kusAm Vir = kot, + kutVir + kugAr (2.84) This model includes the induced velocity of the main and tail rotors which are coupled with the vertical velocity and the yaw rate. For a full informa- tion model, the two state variables describing the induced velocity of main and tail rotor are dropped. w= Zywt Z3Au 2.85, t= Nwow+ Nor + NgAr (285) This system is now a low pass of first order in every path, with the coupling term Ny. In Section 2.4.3, the linearized plant is recognized with a low pass behavior. A model based controller, e.g. a LQ controller, does not have a very complex structure. In (Séderstrém and Stoica, 1989), it is mentioned that the identification becomes quite difficult for systems of low order with simple controllers, because the dynamics of the controller and the dynamics of the plant are not distinguishable. The data set is also made by using the test bench described in Chapter 4. Because of the limiting space of the test bench in vertical direction, a high excitation with the velocity w is not possible. From theoretical 49 2. Modeling, identification, and system analysis r (deg/s) time (8) time (s) Figure 2.22.: Measurement: and simulation of the identified model of the zp plant; w (left), r (right) point of view, the yaw angle seems uncritical for stimulation. Because the transformation of the coordinates to the body-fixed frame is strongly depending on the yaw angle, however, large deviations in yaw angle can cause higher displacements of the helicopter in all three directions. Though of the limiting possibilities for a high perturbed signal, an identi- fication of the parameters is possible, but with a large standard deviation (see Figure 2.22). Identification tests with the subsystem did not yield any useful results. 50 2.6. Helicopter models 2.6. Helicopter models In this chapter, three different models for the horizontal and two for the vertical movement have been presented. One for the horizontal and one for the vertical system are the linearized form of the nonlinear model. The re- maining models are introduced in the identification part. They are different in the number of state variables, although they describe the same subsys- tem in every case. The models, whose parameters are identified in the last section, have a modified structure from the model achieved through the linearization of the nonlinear plant. They have been simplified by reducing the number of parameters. The horizontal subsystem is the more interesting part of the helicopter system and all methods described in this work are applied to this model. Table 2.6 shows the results of the estimated parameters. Table 2.6.: Coefficients of the ry-plant fall information | identified model | linearized from system with servo rotor | nonlinear model Xu “5.3 =1.0724 ~0.012 Yo -1.1767 -1.1887 -0.067 Mu. -24.173 44.224 52.269 My 116.06 9.4572 0.816 M, -21.81 -3.7446 -13.81 Ma -1554.7 242.54 My 2067.4 64.313 Lu 10.848 44.224 52.269 is 19.354 -110.61 -95.143 Ly -6.489 -34.615, -34,702 La 158.55 -161.52 Ls 1587.2 609.11 M 81.441 41.324 2.65 Me -34.029 -39.869 0.7046 Ly -7.6115 32.265 -1.7697 La 35.292 82.95 6.673 k 11.2 11.2 The parameters of the model in the last column of Table 2.6 are from 51 2. Modeling, identification, and system analysis 20f- 2 4 x = 10 : : u Ss : Hx 2 oF -O tet OQ +OHp 0 2 @ ? : xx = -10 a x -20} - x * + 25-200 15-10 =5 0-05 0 05s real(s,) real(s) Figure 2.23.: On the left the poles of the zy plant, on the right a detail view of origin; x linearized model, + model with servo rotor, o full information model the linearized model based on the aerodynamic equations in this chapter. The parameter of the identified model with the servo rotor is very similar in comparison with the parameters of the linearized model. ‘The comparison of the parameters of these two models results in the fact that some parameters have the same value or they are in the same relation to the other parameters of the plant. The parameters of matrix A (eg., Xu, My, Lu, .) in particular are in the same range, while the coefficients of the matrix B (e.g., Mi, Mg, Ly, ...) differ strongly in value and sign. ‘The parameters of the full information system cannot be directly compared to the other models, because the missing of the dynamics of the servo rotor requires a change of the parameter values. To compare the three models, the poles of the transfer functions of all models are plotted in Figure 2.23. ‘The verification of the systems (and particularly of the parameters) is made by comparing the stable and unstable poles. ‘The poles of the linearized model have components on the imaginary axis, while the others models have poles with only real values. In all models, the unstable poles are close together and have a real and an imaginary component. The eigenvectors can be calculated for every model similar to Section 2.4.2 for the linearized model. For the identified model with servo rotor, they cannot be assigned to a special mode because they have components in all state variables. For this reason, the linearized plant is chosen for the control design of the H,. controllers. The full information system is used for the design of the state feedback controller. 52 2.6. Helicopter models Table 2.7.: Coefficients of the 2v-plant full information system linearized from nonlinear model Zw Nw N; 43 Ns -17.568 162.8 -0.5283 3.486 -10.83 -9.2631 -2.584 -1.9669 3.7864 -39.7568 The coefficients of the vertical plant are presented in Table 2.7. The dy- namics for the vertical velocity and yaw angular rate are very close for the full information and the linear model. In both models, the eigenvalues only own real parts. Both of models are used in the control design. 53, 2. Modeling, identification, and system analysis 3. Control design 3.1. Controller for helicopters Control theory distinguishes between classical control, robust control, adap- tive control, and other approaches which are not so popular or too scholas- tical in comparison to the first three mentioned. Most control design meth- ods desire or require a mathematical model of the plant to be controlled. Sometimes, the model is only used to specify the control parameters or give simulation results. In robust control, a (linear) model is the observer part of the controller giving the necessary information about the dynamics of the plant. These controllers give better results than PID controllers in cases where the system has many cross coupled inputs and outputs (MIMO), or where there are modes which cannot measured. Small-scale helicopters are known to have a higher control sensitivity and faster angular dynamics than full-scale helicopters. ‘This is shown in Chap- ter 2, particularly in the modeling of the flap dynamics, depending of the rotor speed and their time constant, which cannot be measured. Model- based controllers are well-suited for this application and will be presented in this chapter. Furthermore a state feedback controller, which is classified in classical design control theory, will be tested and compared to the other controllers. For classical control approaches in particular, the performance of control results depends of the quality of signals which are calculated in tasks of the navigation routine. The navigation part in our system determines the ac- curate position, velocity, attitude, and angular velocity relative to a known reference. The calculation of the necessary signals depends on the utilized sensors and systems. An example for a navigation routine is found in the real-world model used in this work and described in Chapter 4 and 5. The methods are not a main topic of this work and are discussed in (Farrell and Barth, 1999). 55 3. Control design encoder navigation flight controller t—> GPs ‘actuators ref pos pect pos 7 be bay path i commander i is Fo ae, LE | controller structure _ free] ' 1 1 1 1 H 4 1 1 Figure 3.1.: Structure of the controller in the implemented autopilot For control design, the structure of the controller has to be defined and reconceived under dynamic aspects. It is required that the compensator is applicable on every real plant. ‘The controller also has to be integrated in the environment of the implemented autopilot, the sensors, and the actuators (see Figure 3.1). The flight controller needs all available information about position and at- titude. This includes the desired reference position and velocity. The nav- igation part should deliver these signals accurately, without a bias, and noise-free (as far as possible). In particular, the attitude control is respon- sible for the stability of the helicopter and should not be received with a time delay. In a free flight helicopter, the quality of the signal of the ac- celeration and the angular velocities mostly depend on the motion pack, whose measurements are sampled with high frequency (normally at 300 to 400 Hz or higher). The position (and in some cases the velocity as well) is estimated with a global position system (GPS) that has a small data rate (1 to 5 Hz). An intelligent navigation routine includes all signals and filters the signals at the highest possible rate. The technical components of the rotor sensor are generally not as sensitive as the sensors of the navigation part. The rotor speed is about 1400 to 1600 rpm. It is required that the measurement is sampled as least at 60 Hz for cancelling any anti-aliasing effects 56 3.1. Controller for helicopters On the other side, the actuators of the helicopter have a limited bandwidth. In the normal case, they are servos with an integrated controlled step motor. ‘The servos are controlled by a pulse width modulated signal of a frequency of 50 Hz. Experiments have shown that command signals to 25 Hz are accepted with sufficient dynamic accuracy. This information has to be included in the control design. Because of the different sample times at the navigation and actuator part, the sampling of the control routine is set to 100 Hz. There should be no time delays in the helicopter system. A time delay would limit the upper possible bandwidth of the closed loop system and has to be modeled for the control design. In addition, the constraints of the commands signals have to be known to the controller. The path commander for hovering is not so complex and gives the refer- ence set points. Some simple trajectories like steps are made to test the robustness of the controllers (see Section 3.5). 3.1.1. Controller structure For reasons of safety, the controller is divided into two sections: motor and flight controller. The rotor speed has to be held constant at 1200 to 1500 rpm, depending of the type and size of the helicopter. If the rotor speed fails, the dynamic of the helicopter changes rapidly. The vertical velocity behavior is sensitive to the rotor speed when varying. The mathematical model is calculated for a constant rotor speed. For this reason, the motor controller is separate and has inputs of gas (in the electric case control voltage) and collective. ‘The flight controller is also separated into two main parts: the position and the attitude controller. This splitting is mentioned and justified indirectly in Section 2.4.1, where the position is factored out from the model. In Figure 3.5, the chart of the signal flows in the flight controller is presented. The position controller transmits the difference of the velocity and the yaw rate to the attitude controller. The attitude controller is more complicated to design and has a wider in- fluence on the behavior of the helicopter than the position controller. It gets all the necessary signals of attitude and error velocity. Because of the path for velocity control, the quality of position control is also affected. All 57 3. Control design OP hover Auer) attitude and Velocity control Figure 3.2.: Structure of the flight controller the physical characteristics of the helicopter are included in the attitude controller. Different control design methods will be applied to the velocity and atti- tude controller. First, there will be a design of a least: square controller. This is not too complicated to achieve and brings forward a discussion of the stability of the plant. Then, H. controllers are used to get better con- trol performance. Two methods of H. control are presented: the mixed- sensitivity approach, which will shape the closed-loop transfer functions, and the open-loop shaping approach, which relies on the classic control design procedure. Helicopter configuration with an electrical drive is considered in this control design study. The control design for combustion engines will not diverge much from this case, because the dynamic from command signal to rotor speed is not so different and can be adapted with the same model. The separated motor control design is addressing the structure and the parameters of the motor model which are described in Section 2.3.3. It is brought in state-space format for investigation and linear control design. The system has no poles in the right half plane and no transmission zeros 58 3.1. Controller for helicopters 200 50 150 a |g 0 3 10 vot S y ~ 3 50 y 3 -50 Fo t a 5 -100 2 0 2 ae 2 oO 2 4 10° 10 10 10° 10” 10 10 10 frequency (rad/s) frequency (rad/s) Figure 3.3.: Sensitivity and complementary sensitivity with the inverses of the weighting functions for the mot input. Ma] _ [74.66 -8.91] [Ma] , [2996.2 0 ] [umoe (‘s'] = [are 0 1(*]+[ 0 ovd| [x2] ($4) Control design For a stable SISO system of second order (like the abouve motor model with neglected collective input Agr), the theory in (Astrém and Hagglund, 1995) shows that a PID controller cannot be matched by another linear controller. Different approaches exist for the estimation of the optimal control parameters. In this work, a Ha. controller is used for the parameter of the PID controller. A GS/T scheme described in Section 3.3.3 is applied and the specification are thus set for the bandwidth and for the behavior of follow-up control. The system reaches a bandwidth of 13 rad/s (see Figure 5.2). Increasing the bandwidth of the closed-loop system does not give any considerable advantages. Experiments on the test bench of Chapter 4 have shown that the power amplifier has limited dynamics and the engine is not breaking down at increasing load. As expected, the amplitude of the Hz. controller matches that of the PI controller. The two controllers have the same behavior in the low frequen- 59 3. Control design gain (4B) bog & 86 8 | | | 10 10° 10° 10° frequency (rad/s) Figure 3.4.: Comparing the two controllers H. and PI Kaw EY AQ |} Figure 3.5.: PI controller with an anti-windup structure cies. In the higher frequencies, the amplitude of Hz. controller is falling slightly down. The PI controller has the continuous transfer function K(s) = Kp (1 + a) (32) and the controller parameters can be set to Kp = 0.0016 and Ty = 0.0315 s according to the H.. control design. Another advantage of the transformation into a PI controller is the simple adjusting of the control parameter in case the motor parameters change. The implementation takes into account a saturation for the output of the controller (anti-windup). The I-part of the controller is limited with a feedback from the saturation (see Figure 3.5). The feedback gain K4w is set to 400. 60 3.1. Controller for helicopters The rotor speed decreases for huge changes in the collective pitch. This does not affect the case when the helicopter is hovering. If the collective input is used as an additional disturbance input, an Ho. controller with a disturbance observer could be designed (see Section 3.3.3). However, experiments have shown that no additional design procedures are required. 61 3. Control design 3.2. Classical control design Control design not only depends on the chosen design approach, but also on the structure and characteristics of the available mathematical model. Different classes of systems can be defined regard to signal dimension: single and multiple inputs (SI and MI) in combination with single and multiple outputs (SO and MO). In addition to this classification, systems with and without unmeasurable state variables need to be distinguished. Systems without unmeasurable state variables are called full information systems. Most tools for the investigation of closed-loop systems are available for SISO (single input — single output) systems. In case of MIMO (multiple input — multiple output) systems, full information systems give the best control performance. Elementary or low level control design tools are PID for SISO and least square design methods for MIMO with full information systems. With the approaches mentioned, it is possible to give statements about stability, robustness, and performance. The system for a small helicopter should be investigated on this low level for a first controller design. The stability margins achieved are criteria for other control designs. 3.2.1. Stability and robustness margins For the comparison of these different control designs, the closed loop is regarded under certain stability and robustness criteria. These character- istics are independent of the (linear) control design method, because they can be calculated from the linear transfer functions of the plant and of the controller. Transmission poles and zeros The closed loop system is said to be asymp- totically stable if all poles lie in the open left half plane, Re(s;) < 0. The poles s; are equal to the eigenvalues of the state space matrix A of the closed-loop state space representation. Transmission zeros on the right half plane are only of interest in the transfer functions of the design plant and of the controller, e.g. when the controller inverts the plant. Positive transmission zeros are pointing to a phase delay in the system. Systems without positive transmission zeros are called minimum-phase systems. 62 3.2. Classical control design Singular values of loop gain The loop gain L is the product of plant G and controller K», and depends on the order if the MIMO case is considered Le=KvG Lu =GKy (3.3) In the first case, the loop breaking point is at measured signal e; in the second case, the loop breaking point is at the command signal u. The most interesting information from the loop gain L is how primarily the bandwidth (or gain crossover frequency we). We = arg (A(L(jwe)) = I) (3.4) The loop gain L, is regarded if the information of the bandwidth is required in a certain path measured. For the whole system, the open- loop system L,, is magistral, because the minimal order of the system in the helicopter case is at u and L, has not full rank. Singular values of sensitivity and complementary sensitivity The sensitiv- ity S is the transfer function from the reference input to the control error or the transfer function from a disturbance input in the plant output to the output of the control system. There is also a difference on which breaking point S is calculated. Se=(Ipt+Le)' (or Su = (Im+Lu)~*) (3.5) with m being the signal dimension of u, and p being the signal di- mension of e. The same difference is made by the calculation of the complementary sensitivity T. It is calculated by the relation to the sensitivity $ Set+Te=Ip (or Sut+Tu = Im) (3.6) ‘The peak values (H. norms) of S' and T are defined as [loo = maxa(T(jw)), —||Slloo = max a(SGw)) (3.7) ‘They are equal to the maximal singular value. It is required that the peak of S, which is equivalent to the inverse of the distance of the critical point (in the Nyquist plot), is less than 6 dB, and for T to be less than approximately 2 dB. A larger value indicates poor performance as well as poor robustness, i.e. very close to instability. Additionally, the singular values of $ give information on a possible steady-state control error. If S, is quite small at low frequency in a certain path, a steady-state offset will be cancelled. However, a general statement with the sensitivity S,, is not possible. Only the 63 3. Control design quality of the disturbance rejection on the command signal u can be evaluated with the transfer function Sy. ‘The singular values of T describe the final behavior of the closed-loop system. A low-pass filter of first order is normally the desired form of this system. The maximal bandwidth wy of the system can be determined with the singular value plot, and the cut-off region in the high frequency area is apparent. ‘The singular values of the sensitivity and of the loop gain can show sim- ilar characteristics of the closed-loop behavior, e.g., it is possible to see steady-state offsets in the sensitivity S and in the loop gain L. However, all robust criteria can be used with both sensitivity and complementary sensitivity, while the loop gain does not give much information about ro- bustness. Depending on the control design method, both singular values plots are regarded in most cases. 3.2.2. State feedback control design The first attitude compensator is a state feedback controller for a linear plant which minimizes a quadratic cost function (LQ method). In Section 2.5.2, a full information system is derived and investigated. The linear system is in state space form written as follows: a(t) = Ax(t) + Bu(t) (3.8) with x as state variable which describes deviations of the operation point. The cost function J(u) is defined as J(u) = ft [2 ex(t) + u™(ORu(O)] dt (3.9) with Q > 0 and R > 0 as design parameters (see (Geering, 2001)). This least-square method has a solution which minimizes the cost function J(u) with the optimal control law u(t) = —Kox(t) = —RB" K..2(t) (3.10) with K.. as the only positive definite solution of the Riccati equation 0=-ATK—KA+KBR™"BTK-Q (3.11) The relation of Q and R is important for design and not the absolute value of each control parameter. The bandwidth of the closed loop system can be best influenced by R, while the weighting of each state variable by Q. 64 3.2. Classical control design The Riccati equation can also be solved for a plant which is in discrete state-space presentation. The cost function and the Riccati equations are adapted to discrete form. Differences between the continuous and the dis- crete solution are shown at very slow sample time at a plant with dynamics in high frequency. This does not occur in the case of the helicopter. 3.2.3. Attitude control The resulting controller K, from (3.10) is used for stabilizing the helicopter at hover and can be adjusted by the weighting matrices Q and R. In Qcy, the weighting matrix for horizontal motion, the weighting factors for velocity are bigger than those for the angular velocity or the Euler angles. The idea behind this choice is to do control of velocity first. The angle does not need to be kept, because it is an operation point for hover and can change as a result of the configuration of the helicopter. w ° ° ° ey = Rey = 0.55 [ ° ] (3.12) coo ° x The dynamics of the system which are not modeled and the high frequency noise should not be amplified. For this reason, the control parameter R is as small as possible in order to limit the commands signals. The command signal for rolling, Ais, is also more weighted because it allows a higher bandwidth and consequently a faster follow up control. 8 S0fnn nA 5 g 50 B Ops zo 4 3 3 -sot..-< 3 -50 go |- 2 -100 -100 2 2 2 2 s S 5S S 5 5 frequency (rad/s) frequency (rad/s) Figure 3.6.: Loop gain Le, left of longitudinal motion (a), right of lateral motion (b) 65 3. Control design Figure 3.6 (a) shows that the loop gain in forward velocity u reaches a bandwidth of 5 rad/s, in the Euler angle @ a bandwidth 11 rad/s, and for the angular rate q a band pass from 0.4 to 12 rad/s. There is no high gain in the low frequencies of the paths concerning the attitude, and the rate q has no gain at w = 0. This behavior in the derivative is typical for the helicopter, because the angular derivative reacts to high dynamic inputs. A similar picture can be seen in Figure 3.6 (b) for the lateral velocity v, where a bandwidth is reached of 6 rad/s and of 10.5 rad/s for angle ¢. The integral part in the controller is missing and causes the sensitivity Sy not to be very small at low frequency (see Figure 3.7). A steady-state error has to be taken into account, which is not desirable for velocity. 20 fo 10 10° 10 frequency (rad/s) singular value (4B) Figure 3.7.: Closed loop T,, and sensitivity S,, of horizontal motion The sensitivity S,, shows a good result of robustness with a peak of 2.01 dB, but the complementary sensitivity T,, has peak of 2.6 dB. Because the angular rates are statically well suppressed, there is a potential steady- state error and the closed loop transfer function T, is not 1 at s = 0 for every path. The control design for w and r is not so critical and is only limited by the bandwidth of the actuators. Qu = lit to] Ree (3.13) For the dynamics in vertical direction w, the loop gain has a bandwidth of 3 rad/s with a low-pass characteristic. It shows a small gain at low frequency and does not increase for higher frequencies, because the gain is critical for the actuators in case there are disturbances of fast dynamics. The sensitivity for the altitude (see Figure 3.8) is more significant, which is a little under the 0 dB line for low frequencies. Small changes in the collective 66 3.2. Classical control design Aw causes some changes in the vertical velocity at the high rotor speed of 1500 rpm, although the loop gain is quite small. ‘This disadvantage of the controller shows that at low frequency, disturbances are not well suppressed and good follow up control is missing, 8 8 & i 8 & & 5, singular value (4B) ° singular value (4B) ° 2 5, 10 10 10° 10 frequency (rad/s) frequency (rad/s) Figure 3.8.: Loop gain L., left for vertical motion (a), right for yaw mo- tion (b) An opposite behavior of the vertical movement is found in the path, where the gain and the bandwidth of the open-loop transfer function is quite high. The scaling is not as critical as the scaling at vertical velocity. The singular values of the complementary sensitivity show good robustness and a bandwidth of 12 rad/s. For the path, it is obvious that the behavies of complementary sensitivity matches the singular value of a low pass. This Part of the plant is not sensitive from the control point of view. _ 20 s 20 3 19 2 a 2 0. 2 107 10 10" frequency (rad/s) Figure 3. motion Complementary sensitivity 7, and sensitivity S,, of vertical The controller K, from equation (3.10) gets the difference of measured in- 67 3. Control design puts and operation point or reference signal and has static coefficients (see Table 3.1). The gains of the attitude controller are different for every path, particularly for velocity and Euler angles due to the choice of the weight- ing matrix Q. The results of this control approach are depending on the Table 3.1.: Gains for the attitude control 0 | 6 w v p4 ]? [wl By, | 1.515 | 0.206 | -10.395 | 1.901 | 0.05 | 0.026 0 0 Ajs | 0.119 | 1.044 | 0.526 | 9.308 | 0.005 | 0.063 0 0 Ay 0 0 0 0 0 0 3.438, 0 Ar 0 0 0 0 0 0 0 -0.09 accuracy of the operation point. Because there are no integral parts in the controller, a steady-state offset is possible through an incorrect: operation point. 3.2.4. Position control Since the velocity control is achieved through the attitude controller, po- sition control does not have to be very complex. In the implemented au- topilot, a PI controller receives the position error in body coordinate. The resulting signal is a reference velocity. An additional reference forward velocity u can be added to the controller output (see Figure 3.10). The in- tegral part of the PI controller is limited and can be reset. An anti-windup structure for the whole controller may also be possible. Figure 3.10.: Position controller in u direction The structure is the same for every velocity path and heading. The values 68 3.2. Classical control design are given in Table 3.2. The gains in the vertical velocity are different. In case of fast maneuvers, the position control structure allows for references in velocities and rates to be given, and it is quite obvious to tune. Table 3.2.: Gains for position control ely [2 lulv[wl ge rs Kp|}2}2/07/1]/1/1/25]1 Ty | 5 | 5 | 15 | - | -{ - 5 | - 69 3. Control design 3.3. H. mixed sensitivity design 3.3.1. Introduction Feedback is used to reduce the effects of uncertainty which can appear as disturbances or as other imperfections in the used control-design models. Model uncertainty and robust control have been a central theme in the de- velopment of the Hy. control theory. Robust control theory applying Ha. control was very popular since the beginning of the 1990s, when a state- space realization was proposed by Glover and Doyle. A detailed derivation of this H.. solution is treated in (Zhou et al., 1996), and a very good sum- mary is given in (Zhou, 1998). From a practical point of view, good aspects are listed in (Skogestad and Postlethwaite, 1996), and the engineering as- pects are described in (Christen, 1996) and (Roduner, 1997). 3.3.2. H. theory A short derivation of the Ho. method ask for the definition and verification of the H, norm for systems. It can be interpreted as an induced norm of signal norms. If the definition of the Ly norm of a vector signal z(t) lellp = ([.Srore) P€ [1,00) (3.14) on a time interval t € R (often t > 0) is considered, the space of all measured functions x(t) with finite L, norm is called a Lebesque space. The most popular norm is the Ly norm \lzll2 = (f° > ln(orat) = ft xT (t)a(t)dt (3.15) isi The Lz norm contains the signals of bounded energy. The definition of the A norm of a linear time invariant system with matrix G(s) € C?*™ is IGlloo = sup a(G(s)) (3.16) Re(s)>0 where G(e) denotes the maximum singular value. According to (Zhou et al., 1996, page 104), it can be written as a induced norm a sup WGule guy Gulla (3.17) ihullzz0 ella gulia=a IIGlloo = 70 3.3. Hoo mixed sensitivity design This is valid in the time domain as well in the frequency domain and denotes the maximum of a system by which the energy of the input signal can be amplified. Hoo is the name of the Hardy space containing the transfer functions which have a finite Ha. norm, are analytic in the right half plane, and are proper. co = {G(s)| Jim G(s) < 00, G(s) analytic in Re(s), ||Glloo < oo} The idea of the Ha. control design method starts from this definition. The plant, which should be controlled, is augmented by dynamic weighting functions which are in general in the frequency domain and specify the closed-loop behavior. The augmented plant G, has two input vectors with w (dimension m:) as external disturbance and u (dimension mg) as output of the controller. There are also two output vectors with y (dimension p2) as the measurement vector directed to the controller and z (dimension p1) as an error signal desired to be small (see Figure 3.11 (a)). a we 2 wt Se a el ds « a0 Ke y Ky (a) (b) Figure 3.11.: Transfer function T;,, without (a) and with model error sys- tem (b) The pair z and w is used for the formulation of the control problem: The goal is to design a stabilizing controller K in order to constrain the transfer function Tz.) in the Hoo norm. This is called the Hoo problem and can be generally formulated for a two-port linear system G, with A| BB; AeR™", Be R™™ Gp=|G [Du De |, pixn pixmy (3.18) Co | Dar Das eae or Gu Gy 3.19 [at ef] (19) 71 3. Control design for which a controller Ky is sought for the closed-loop system Tew = Gir + GiaKo(I — Ga2Kv) Gar (3.20) to stabilize and to constrain its Ho. norm. Ky = arg(||Teulloo <7) (3.21) There are different mathematical solutions presented for this problem. In this work, the algebraic method based on Riccati equations is used. The augmented plant G, has to fulfill the following assumptions A1 to AB: Al (A, Bz, C2) is stabilizable and detectable. This is intuitionally neces- sary and concerning the plant G and the additional weights. A2 a(Di1) < 7, because the feed through from w to z has to be smaller than the desired boundary 7. A3 rank(Di2) = mo, i.e., D¥,D12 is invertible. In other words, every input signal of the controller u has an influence on the output z. A4 rank(D2;) = po, ie., DF, Dz; is invertible. In other words, every input signal of the disturbance w has an influence on the output of the plant ys A5 The matrices above have full rank for all w A~jwl Bp A-jol By [ CG > || Co Da The last assumption is made for the solvability of the Riccati equations. The important property of the Ha. norm results from the application of the small gain theorem: If ||Twlloo < y then the system with the block diagram in Figure 3.11 (b) will be stable for all stable A with ||Alloo < 1/7. A is the uncertain model which can be described in the frequency domain and stands for unknown linear systems. The descriptions and characteris- tics of uncertain models are described in (Zhou et al., 1996) and (Weinmann, 1991). There are multiple types of uncertainty in a closed loop system, de- scribing different variants: Additive uncertainty The model uncertainty can be represented by an ad- ditive perturbation (sce Figure 3.12 (a)). The system Tio, which is in 72 3.3. Hoo mixed sensitivity design the loop with the modeling error A, is made up of the compensator Ky and the sensitivity S., respectively Sy. Tio = —KypSe or Tio = —SyKy The additive uncertainty describes the model error of uncertain right half plane zeros and shows how the high frequency dynamics in the loop is neglected. If the signal through A passes through the opposite direction, the lower composed system is made up of the plant Gy instead of the controller Ky. Tio = —SeGs or Tip = —G5Su It can be shown that the error model in this case describes uncertain poles on the right half plane and low frequency parameter errors. (b) Figure 3.12.: Additive uncertainty (a) and multiplicative uncertainty at the output (b) Multiplicative uncertainty The multiplicative uncertainty can be set at the input or at the output of the plant (see Figure 3.12 (b) for a multi- plicative uncertainty at the output of the plant). If the uncertainty is at the input of the plant, input actuator errors can be modeled. The complementary sensitivity —T, is the system which is in the loop with A. If the uncertainty is at the output of the plant, output sensor errors are modeled and —T, is in the loop. Tio=—-Te or Tio =—-Ty In both cases, the high frequency dynamics are neglected and un- certain transmission zeros of the plant on the right half plane are modeled. Also, parameters uncertainty can be translated into mul- tiplicative uncertainties, e.g, the set of plants with gain uncertainty 73 3. Control design described by Gy = 96G5(8) Goin < Ge < Imax where g. is an uncertain gain and G,(s) is a transfer function with no uncertainty. It can be rewritten as multiplicative uncertainty Gy = goG.(s)(1+4rg) FER, |5] <1 where go is the average gain and r, is the relative magnitude of the gain uncertainty. Divisive uncertainty The divisive uncertainty (or inverse multiplicative un- 74 certainty) bears resemblance to the multiplicative uncertainty, be- cause it has the same structure, except if the signal flow goes in the opposite direction. Instead of the complementary sensitivity in the closed loop, we have the sensitivity —S,, and —S, respectively. Tio=—-Se or Tio = —Sy Low frequency parameter errors and uncertain poles of the plant on the right half plane are tested by this uncertainty. In addition it can be used for poles crossing between the left- and right-half planes. The following example shows such uncertain system with changing numbers of right-half plane poles described by SER, |s] <1 (3.22) Then G,1 = ;4; has one right-half plane pole and Gy 2 = ;4; has no right-half plane pole, but both systems are described by’ (3.22). Nevertheless, the set of uncertain plants CG, can be covered by a di- visive uncertainty description as shown in Figure 3.13 with the plant Gs=} —aat Figure 3.13.: Divisive uncertainty at the output of the plant 3.3. Hoo mixed sensitivity design Coprime factor uncertainty Another uncertainty occurs when a left co- prime factorization of Gy is calculated (e.g., Gp = M-1N) and the plant is perturbed by the uncertainties [Ay Ajy] (see Figure 3.22). In this special case, uncertain right half plane poles and zeros as well as low frequency parameter vectors are taken into account (more in Section 3.4). If the different perturbed closed loop systems are considered, the model er- rors are only one uncertainty block at a time. For more known uncertainties at the places where they occur, one unstructured block can be used with additional off-diagonal blocks in the uncertainty matrix. For applying the Hoo norm on these perturbed systems, which should be smaller or equal to 7, weights W, are appended to the system. These weights have to reflect the upper bound on the magnitudes of the uncer- tainties and are dynamical systems which approximate the bounds for the various uncertainties. Depending on the robustness tests chosen, the bounds describe the inverse of the required limit. For different uncertainties, there is the corresponding transfer function of the closed loop. The shape of these lower transfer functions is specified for the control design. Often, the multiplicative and divisive uncertainty for weighting the sensitivity and the complementary sensitivity are chosen. A solution for the posed augmented system can be calculated when the assumptions A1 to A5 are satisfied. The popular $/KS/T scheme (see Figure 3.14) is an example of a scheme with an augmented plant which weights the sensitivity S. and the comple- mentary sensitivity T, and fulfills all assumptions. The main control design is determined by these two closed-loop transfer functions, thus they are the desired shapes to specify. The sensitivity S, is bounded by the inverse of the weight W, and the com- plementary sensitivity T, has the inverse of the weight Wy as the limiting system. The transfer function Tey is WeSe Tew = | WukoSe (3.23) WyTe from which the H,. norm has to be calculated. An additional weight W,, is placed and has the command signal as input. The closed loop system is K Se, which is equivalent to an additive uncertainty. Using this additional 75 3. Control design Figure 3.14.: S/KS/T scheme with error model constant weight or a dynamic weight with feed-through is a way to satisfy Assumption A3. The solution constrains the transfer function T.,, to y, which is normally chosen as one. From the first row of T,,,, there is \|WeSelloo<1 +> Wet>s, (3.24) where S, is the desired shape for S, in the relevant frequency ranges. The transfer function T. can be affected in the same way with W,. W, can be chosen to be constant and arbitrary small (the gain of the controller can be slightly influenced). Because of assumption A1, all of the weights must be asymptotically stable as they are not detectable from the second output of the augmented plant. Every scheme has some difficult points in terms of implemention, particu- lary this S/KS/T scheme. If the plant G, contains poles on the imaginary axis, assumption A5 is violated because the transfer matrix of the plant cannot be influenced from the input w. The helicopter plant in particular has this problem, because the calculation of the angles happens by inte- grating the angular rates, which causes poles at s; = 0. A further problem is that the H,. controller designed using this augmentation scheme tends to invert the plant and makes a pole/zero cancellation. This is not optimal because uncertainties in the poles and zeros can cause slow transients and the control system is less robust. 76 3.3. Hoo mixed sensitivity design 3.3.3. The GS/T-scheme The structure of a scheme and can be chosen at random only depends on the fulfilling of assumptions A1 to AS. A scheme, which is more suitable for the helicopter problem, is the GS/T-scheme. Not only does it weight the output of the plant, but also the input and an additional disturbance input at the command signal. ‘There are now four different weights for the closed loop shaping (see Figure 3.15). , Ly ww, ah fw, H=$ 1 j 1 1 T. Figure 3.15.: GS/T weighting scheme W, and W, are similar to the S/K'S/T scheme, Wz weights the input of the reference signal and W,, lies in the input of the disturbance path. By calculating the H.. norm of the transfer function —WiTwWy —WaK S.Wa Tew =| WiG.SuW, — WyTeWa (3.25) the effects of the weights on the closed loop system are more obvious. The complementary sensitivity and the sensitivity appear in more than one ele- ment of the T;» system matrix. The designer can now choose if the system is designed by loop shaping at the input or at the output of the plant. If Wa is chosen to be very small, the other terms in the row can have rather big singular values as long equation (3.21) is fulfilled and T, (shaping at the output of the plant) can be specified independent of T,,. The sensitivity is weighted at y and the plant G is included in the term. This complicates the design. The transfer function from w, to zy can be transformed as follows: W,G.S.W, = W,5-G,W, If the design weight for the sensitivity is W,, the product W,G, is the boundary to S, (with W, = I). Otherwise, S, can be shaped by GW, 77 3. Control design (Wy = 1) in case this weight can be chosen. Weighting with the plant is an advantage as long as the plant has integral behavior in the desired paths and the weight can be set more or less static. If not, the sensitivity might not be able to be weighted as specified in the low frequencies, hence often causing a problem of steady-state error. A solution for better tracking is to design a two degree of freedom controller (2DOF). The GS/T scheme is extended with an additional weighting path (see Figure 3.16). ol We wml ar Figure 3.16.: Extended GS/T weighting scheme The feed-forward part is superposed over the standard scheme and the transfer function T,,, in the 2DOF case is a [aMaTW. We Suk yWe Wy SukyWe — —WakeSeWe Tow = [igiaih, “EEREGET ikea, —yategs’ (3.26) ‘The path for the feed-forward in the second and third column of the second line is not together with the sensitivity or the complementary sensitivity. 3.3.4. Attitude control When applying the GS/T scheme for attitude control, the first step is to concentrate on the closed loop specifications. They have to be defined for the sensitivity and the complementary sensitivity by regarding the transfer matrix (3.25). 78 3.3. H. mixed sensitivity design Again, the plant is separated into the horizontal and vertical movement for a clearer overview of the design procedure. In the horizontal movement, there still is the difference weighting idea for the attitude angle (inclusive angular velocity) and the velocity. The velocity path is quite important and determines the behavior in the position loop. The weight Wa, which is multiplied with the complementary sensitivity T, and the controller product KsSe, is chosen to be small so that the influence on the last column of T:» can be neglected. Additionally, the input weight W, is set to the unity matrix so that the design parameters are the weights Wy and Wa: Wa=10° Is Wy=hb (3.27) The product WyG, is the weighting of the sensitivity 5, and is the main design weight of the controller. The bandwidths of the velocities, Euler angles, and angular rates path are adjusted by the weight W,. Three con- trol variables can be introduced for the tuning of the controller: go, and 4a. ave for the bandwidth of the Euler angles and angular rates, g..y for the bandwidth and the follow-up control of the velocities. The additional dynamics in the velocity path improve the control quality. 90,6 T2 0 0 W=| 0 7th 0 (3.28) 0 0 Ga,pl2 ‘The Euler angles are more weighted than the angular rates. The variable 90,¢ With factor 3 is higher than the variable gj, with factor 0.3. For velocity control, the bandwidth is set to 10 rad/s with g,,) = 1.5. Figure 3.17 shows that the (lower) bandwidth of the sensitivity 5, is limited to 10 rad/s with this choice of W,. The resulting bandwidth of S,, is about 11 rad/s, which is quite high. The main goal of the weight W, is to limit the bandwidth of the closed- loop transfer function T,,. There is no reason to give an upper bound for the complementary sensitivity, because a maximum of T,, larger than about 2.2 dB is never reached in the design process. ? + 40s + 400 001s? + 2s + 108-7 (3.29) W, multiplied with the term K»S, should influence the bandwidth of the controller Ky, as far as possible. Additionally, the bandwidth of the con- troller can be adjusted by the weight Wq. The difference between attitude and velocity control can be investigated by replacing the weight Wy at 79 3. Control design $ s & a Ee 8 3 ao g 3 3 Zo 3 5 3 3 £ 27 2 0 2. a 2 0 2 107 10 10 10° 10 10 frequency (rad/s) frequency (rad/s) Figure 3.17.: Sensitivity S,, with limiting weight W,G., and closed loop T,, with limiting weight W,, for horizontal motion (3.27) by =1073| %e 0 Wa = 10 (iG Al (3.30) and varying the variable bg, between 0.1 (more velocity control) and 1 (more attitude control). In this design, it was not necessary to change this value. The crossover frequency of the loop gain L,, is 10.1 rad/s. The controller has no integral part, and the plant has no integral behavior either. Thus the open-loop singular value is limited at low frequencies and a steady-state error has to be taken into account. 3 8 a Ee z 50 3 50 3 3 : = 0 so & - & _ > -S0} b> -50 s 2 me 2 oO 2 roe 2 0 2. 10° 10 10" 10° 10 10" frequency (rad/s) frequency (rad/s) Figure 3.18.: Open loop L for horizontal motion 80 3.3. H. mixed sensitivity design A picture quite similar to the LQ-design shows the loop gain Le for the different paths in Figure 3.18. The main difference between these control designs lies in the higher bandwidth in all paths. The bandwidth for the velocity u and v is about 7 rad/s. It is not as high in the LQ case. Increasing the bandwidth in the velocity control was only possible in the control design by reducing the quality of the attitude control concerning the Euler angles. The desired specification for control of the Euler angles concerns an ignored offset of steady state. This is, however, a contradiction to velocity control, because according to the system equations in Section 2.5.2, the Euler angles are more or less linear to the accelerations. ‘The controller is primarily made for good attitude control; hence the band- width for the Euler angles is about 10.5 rad/s and for the angular rates about 12 rad/s. For better performance in velocity, the extended scheme in Figure 3.16 was used, which includes an additional feed-forward path: _ 0.0015 + 0.8 a s+0.24 In Wr =06 In (3.31) The important weight is W,, which weights a ,sensitivity“of the feed- forward path, because the term Ty, — I in the transfer matrix (3.26) is similar as the relation (3.6). If W, is high in the low frequency range, the follow-up control is good and errors in steady states do not appear. The bandwidth of W,, thus limits the lower bandwidth, the weight W; can limit the upper bandwidth of the feed-forward path. _ 50 _ 50 3 ol 3 3 g 0 = -s0 H 3 5 5 2-100 4 ae 2 0 at 2 0 2. 4 10” 10) 10° 10” 10 10 10 frequency (rad/s) frequency (rad/s) Figure 3.19. Feed-forward path Ty, and controller Ky and Ky for horizontal motion Figure 3.19 shows that the singular value of the feed-forward path Ty indicates good follow-up control with a value of 1 at w = 0. The feed- 81 3. Control design forward controller also depends on the design of the feedback controller and is calculated with the weights from (3.27) to (3.31). The choice of these weights explains that the design goal for the feedback controller has to be good attitude control, the remaining duties have been given to the feed-forward compensator. A better performance is expected from the control design in the vertical movement than from the LQ-design controller because of the consideration of the unmeasurable state variables of the plant (induced velocity). The range of the achieved solutions is not limited by the chosen weights and a controller for -y = 1 is always calculated with no problems. _et15 0.10450 Wa= [ Tose OS ] y= | #00 on, wot 0 ait 0 SStior 8413.3 0 (3.32) ss we= | PBF ssatee | Woh The complementary sensitivity shows a low pass characteristics of first or- der (see Figure 3.20). There are different bandwidths for the yaw and the a) _ 2 a g 8 8 5 10 3° 3 0 ¥ _20}- 3 3 : ao 2 oO 2 ne 2 0 2. 10° 10 10 10° 10 10 frequency (rad/s) frequency (rad/s) Figure 3.20.: Sensitivity S,, with limiting weight WG, and closed loop Ty with limiting weight W, for vertical motion altitude path. The slower path is for the vertical direction, which is less sen- sitive to disturbance as in the LQ case. This is the reason why the weights Wg and W, limit the bandwidth. The singular value of the controller in Figure 3.21 decrease for higher frequencies. In this way, high dynamic dis- turbances are well suppressed. With an additional feed-forward controller, the performance on vertical movement is improved. W, and W; are the weights for the feed-forward path. They are completing the scheme 3.16 82 3.3. Ho. mixed sensitivity design with the weights 3.32. W= 0.001s + 10 8+8 (3.33) The singular values of the compensator in Figure (3.21) shows the higher bandwidth of the feed-forward path. _ _ 30 3 oF 8 3 -20 3 3 \ a 3 \ 4 10° 10° 10° frequency (rad/s) frequency (rad/s) Figure 3.21.: Feed-forward path Ty, and controller Ks and Ky for vertical motion Table 3.3.: Gains of the attitude controller @ d u v q Pp w r Biz | 1.797 | -0.618 | -23.011 | -4.241 | 0.288 | -0.051 0 0 Ais | 0.406 | 1.724 | -5.289 | 20.116 | 0.073 | 0.217 0 0 Amu 0 0 0 0 0 0 3.751 | -0.001 Ar 0 0 0 0 0 0 3.872 | -0.079 Table 3.4.: Gains of the feed-forward part u v w r By, | -25.232 | -4.804 | 0 0 Ais | -5.815 | 22.684] 0 0 Au | 0 0 | -5.579 | -0.001 Ar | 0 0 | -3.453 | -0.08 83 3. Control design 3.4. H. loop shaping design 3.4.1. Introduction The H.. loop shaping procedure was introduced in (McFarlane and Glover, 1992a). It does not directly minimize the sensitivity and the complemen- tary sensitivity functions. It is close to the classic control design for PID controller because it uses the open-loop gain L, or Ly, to apply the desired closed-loop specification. This procedure has proven to be a popular and intuitive means of designing robust feedback systems. 3.4.2. Theory of the H,, loop shaping method The aim of this section to give the theoretical background of the method and the necessary explanations of the used control parameters. The first step is to give the definition of coprimeness and to apply it to a given plant model G. The product G=M"'N is a normalized left coprime factorisation of G if « N and M are asymptotic stable © There exist two asymptotically stable systems X and Y such that MX—NY=1 ¢ N and M are normalized in the following way NN*+MM* =1 Any rational transfer function G can be represented as the quotient of the two stable transfer functions M and N. If G has any poles on right half plane, then they will appear as zeros of the denominator of Mf. N will be a_minimum-phase transfer matrix, because all positive zeros of G are in M. In (McFarlane and Glover, 1992a) an algorithm is presented for the calculation of a state-space representation of N and M. It is based on 3.4. Hoo loop shaping design the minimal state-space realization of G and utilizes an algebraic Riccati equation. [Wa] = (3.34) RIC R?D RZ A+HC|B+HD HW with R=I+DD™ H=~-(ZC™+BD")R™ and Z = Z™ > 0 being the solution of the Ricatti equation (A- BS"'D"C)Z + Z(A- BS“'D™C)" — ZCTR™CZ + BS"BT =0 with S=I+D"D The class of perturbed models of interest is given by Gp = (M+ Ay)(N + Ag) (3.35) in form of Figure 3.22, where Ay, and Ag are asymptotically stable, un- known transfer matrices representing the uncertainty and satisfying 1 AwAslll < a with y > 0. An interesting feature of the framework of the coprime factors is that many Ay, and Ax describe the uncertain plant Gp. If Gp has any poles in the right half plane, then no simple additive uncertainty exist so that Gp = G+ A (see (Vinnicombe, 2001)). However, with coprime factor uncertainty, it will usually be the case that any Gy is described by equation (3.35). Additionally, Ay, describes the uncertainty of all right half zeros of the plant. an ako a eo 2 Ke st > Figure 3.22.: Perturbed system of left coprime factors The design goal is to find a feedback controller Ky which stabilizes all perturbed plants G, for a given 7. According to (Zhou, 1998, p.317) or 85 3. Control design with regard to the input w to the outputs z; and 22 in Figure 3.22, this can be rewritten in the framework of an Hy. optimization problem: Find a stabilizing controller K such that I[E]e-emrw <7 (3.36) where K is chosen over all controllers which stabilize G. If G is stable, NV can be chosen J and (3.36) changes to (3.37). Il k ] (ext all <7 (3.37) In this case, the transfer matrix can be interpreted as a signal based Hoo design method described in Section 3.3. The resulting weighting scheme would, however, be quite complicated to specify and the shaping of the closed loop transfer T’ would not materialize. The solution of the smallest y can be iterative, but because the left coprime factorization of G is normalized, it is possible to show according to (Me- Farlane and Glover, 1988a) or (Zhou, 1998) that a minimum value of 7 can be obtained by a non-iterative method. It is given by min = VT — Xmax(XZ) (3.38) where X > 0 is the solution of the Riccati equation (A- BS“ DC)" X +. X(A- BS~'D™C) — XBS"'BTX + CTR™C =0 The controller achieving this bond is given by _ [A+ BF +7Q712Z07(C + DF) | QuZcT Ls [Seer eee ee (3.39) where F -S"\(D™'C+BTX) Q=((1-7*)1+ XZ) Until now, the control design is made with the normalized left coprime fac- torized plant G. No specifications for the closed-loop behavior were given. In this method, the principal idea of ”loop shaping" is used. So the maxi- mum singular values of the sensitivity $ and the complementary sensitivity T are directly determined by the singular values of the corresponding loop gain matrix L. This design technique is similar to the Loop Transfer Re- covery (LTR) method in LQG design, where the designer specifies a desired loop shape. The following steps are required: 86 3.4. Hoy loop shaping design 1. Open-loop shaping — The use of a precompensator W; and a post- compensator W2 with the plant G shapes the singular values of a control law to give a desired open-loop shape L = W2GW,. It is assumed that W, and W2 are such that £ contains no hidden modes (all state variables are completely observable). 2. Robust stabilization — A feedback controller Ko. which robustly sta- bilizes the normalized left coprime factorization of G, is synthesized. The stability margin 7" is an indicator of the success of the loop shaping. 3. The final controller K’ is then constructed by combining the H. con- troller Koo with the weights Wi and W such that K = WiK..We However, this open-loop shaping approach becomes more complicated by the need to ensure robustness of the resulting closed loop system. The cal- culated value of 7 indicates how far the desired robustness is granted. This requires that the phase properties of the plant are considered as well and the loop shape can be shown to be limited by certain robustness requirements. This solution for a 1DOF controller has the following advantages compared to the mixed sensitivity method: © There is always a solution of the control problem. The + iteration is not necessary. * The procedure is applicable to stable or unstable, minimum-phase or nonminimum-phase plants as long as the shaped loop gain L is completely controllable and completely observable. © There are not many parameters for the control design because only the weights W; and W are available for the loop shaping. A 2DOF design with the NLCF method is proposed by (Limbeer et al., 1993 and Green and Limbeer, 1995) to enhance the model-matching prop- erties of the closed loop. With this additional path, the feedback part of the controller is designed to meet robust stability and disturbance rejec- tion requirements similar to the general procedure of loop shaping design except that only the precompensator W is used. A specified model Ty, is introduced which should force the closed-loop behavior to the desired specification (see Figure 3.23). 87 3. Control design Figure 3.23.: NLCF scheme with 2DOF extension The design problem lies in findig the stabilizing controller with a feed- forward part K =[K; Ky) for the shaped plant G, = GW; which constrains the Ho. norm of the transfer function pSuKy KySeM~* Tow = pSeGsKy SeM~* (3.40) P’(SeGsKy—Tyr] pSeNf~? from [r w;]” to [z1 22 23)". The problem is solved with the same 7 iteration as used in the mixed-sensitivity design. The purpose of the prefilter Ky is to ensure that IW - G. Bu)" @.K 5 — Tyrlloo S10? where Ty, is the desired closed loop transfer function and pis a scalar pa- rameter used to increase the model matching in the optimization at the expense of robustness. The compensator K is calculated in detail in Ap- pendix D.1. Compared to the case 1DOF, the design process changes as follows: © Open-loop shaping — Using only a precompensator W;. The de- sired open-loop shape £ = GW, is designed with the same procedure described at the IDOF design. © Closed loop — The desired closed-loop transfer function Ty, is se- lected between the referenced and controlled outputs. If there are more plant outputs available that need to be controlled, the signal dimension can be reduced on the required controlled outputs. 88 3.4. Hy loop shaping design © Robust stabilization and feed-forward design — The H,. optimization for the augmented plant T,,, (defined in equation (3.40)) is solved for the shaped plant G,, the desired Ty,, and the scalar parameter p. © The final controller with the feed-forward part is put together with the augmenting transfer function Wy. 3.4.3. Attitude control ‘The control design for the feedback compensator is limited to the weight W, with regard to the 2DOF design. ‘The crossover frequency and the roll- off of the loop gain L can be adjusted with this parameter. It is a low pass of first order for the horizontal movement, and the same for the vertical movement. The heading can be adjusted with a constant, because there is no additional roll-off necessary. soi 2 0 0 Wi = 480 We = Is (3.41) 0 ool The specific paths could be adjusted with the weight W». In this case, however, only by the scaling the plant can an individual correction of the paths be achieved. For the investigation, the plant is separated into the horizontal and the vertical (with heading) movement. The crossover frequency is close to the other design. For the loop gain Ly, a bandwidth of 10.5 rad/s is required ~ 100 ~ 20 3 3 0 : : 4 3 -80 5 10° 10 frequency (rad/s) frequency (rad/s) Figure 3.24.: Loop gain L,, and preshaped loop gain £,,, and the sensitivity S,, and complementary sensitivity T,, of the horizontal movement by the chosen weight W;. The run of the curve of the open-loop singular 89 3. Control design value matches the specified singular value very closely. A ‘ymin of 2.43 is calculated. If ymin is very small, then the specified loop shape is incom- patible with robust stability requirements and a new 7 = 1.1ymin can be adjusted (see (Zhou, 1998, p.326)). The resulting cross-over frequency is 11 rad/s. For both velocities u and v, there is a bandwidth of 7 rad/s. This > 100 > a 5 oh 3 an) - 3 ze } 5 [ae 5 © 109 ° 10° frequency (rad/s) frequency (rad/s) Figure 3.25.: Open loop Le of 8, u, and q, and also of ¢, v, and p bandwidth is in the same range as the other controllers. The cross-over frequency for the Euler angles and angular rates are also higher with 10.2 and 10.6 rad/s. There are no large differences to the other Hz. control design concerning the singular values. The experience of the earlier control design influence in the choice of Wi for the part of the vertical movement and the heading. The cross-over fre- quency for the vertical movement should not be too high, and the controller for the heading has the same behavior as the other controllers. The insight of the control design in bandwidth and steady-state gain is used for the calculation of the feed-forward controller part. The desired model for the feed-forward part Ty, is included in the scheme of Figure 3.23. For the horizontal movement, the feed-forward path should not be too fast and the following Tyr gives good results: 0.7 Ter = Oa5gi 2 (3.42) A low pass is the desired model with a cut-off frequency of 5 rad/s. The static gain of the low pass is not of value one, because the gain of the con- troller would be too high and the bandwidth in a range where disturbances would be amplified. 90 3.4. Hy. loop shaping design _ 50 g é a g = 5 4-0 3 2 2-40 ae 0 2 0 2 4 10 10° 10 10" 10 frequency (rad/s) frequency (rad/s) Figure 3.26.: Loop gain L. and preshaped loop gain Le, and Sensitivity S. and complementary sensitivity T. for the vertical movement. Table 3.5.: Gain of the attitude controller a @ u v q P w r By, | 0.576 | -0.091 | -10.846 | -1.591 | 0.037 | -0.01 0 0 Ais | 0.085 | 0.598 | -1.538 | 10.9573 | 0.01 | 0.038 0 0 Am 0 0 0 0 0 0 3.835 | -0.001 Ar 0 0 0 0 0 0 6.749 | -0.163 The same arguments are valid for the feed-forward path of the second part of the systems: (3.43) The design parameter p can be used to increase the performance at the expense of robustness. The value of p is between 1 and 3. In this case, this was not necessary. The controls is designed with p = 1. 91 3. Control design 3.5. Conclusions and results In order to compare the three different controllers, design characteristics and test results are evaluated. The properties of the quality of the design methods are shown and confronted to each other in the following sections. In addition, experiments give any rates about success of the applied method and information about stability and robustness. The test bench described in Chapter 4 is used to make the maneuvers for checking the reliability and robustness of the applied controllers. A step of 30 cm in positive x; and y; directions is executed without a reference in velocity. This step can be interpreted as a disturbance in position or velocity measurement or as an adjustment of the position. The steps are also made in negative «; and y; direction. However, there are no striking differences remarkable in the measurement for all tests. Primarily, the performance for hovering is tested, rather than the follow- up control for fast maneuvering. The velocity in this case is also a long during impulse which is controlled by the velocity and attitude controller. ‘Tests with a similar reference of velocity are made in simulation before the experiments on the test bench are conducted. These simulation results are found in Appendix E. In addition, the same steps as in the experiments mentioned above are simulated and presented in Appendix E. 3.5.1. LQ controller and position controller The LQ controller is the most suitable controller for easy tuning and for fast implementing. ‘The design is made by a full information model which does not match the physical model of the helicopter. The controller has, however, shown good robustness and good behavior. Because the bandwidth of the controller is not limited, not modeled disturbances with high dynamics are not rejected and the bandwidth of the closed loop can be quite high. It can be seen in Figure 3.27 and 3.28, where the LQ controller always has a fast response on the given input. ‘The position controller was mainly adjusted with the LQ compensator and remained unchanged for the other controllers. The PI controller is the sim- plest solution and is sufficient from a practical point of view. ‘The integral part is for a steady-state offset in position and the gain 1/T», can be kept 92 3.5. Conclusions and results very small. The gain K, can reduce a fast position offset and is also kept small. The influence of the position loop is held small. The primary goal is to give the performance of control to the attitude and velocity controller. In this configuration, a gain scheduling of controllers for different operating points could be realized. 3.5.2. H,, controllers The two tested H.. controllers differ from the LQ controller through the model they are designed with. The auxiliary rotor is included in the second model and describes important dynamics in the model. The effects are well seen in the step of Figure 3.27, considering at the attitude angle @. An oscillation of the angle exists which also affects the position. This behavior is not remarkable in the step responses with the H,. controllers. The Hoo controller developed with the mixed sensitivity design method has the task to hold the attitude at the expense of the performance of velocity control. A look at the system equations in Section 2.5.2 shows that the velocity is calculated by integrating the Euler angles. This correlation between velocity and attitude angle prevent good attitude and fast velocity control on the same time. For this reason, the bandwidth for position is also smaller, which is shown in Figure 3.27 and 3.31. One of the differences between the two design methods of Ho is the num- bers of parameters for adjusting the closed loop characteristics. The method with the mixed sensitivity design has more design weights and they are more difficult to choose for the right specifications, particularly in the GS/T scheme. There are more possibilities to shape the closed loop at the plant input or at the plant output. However, both cannot be applied at the same time, even though it seems possible with this scheme, e.g., the complemen- tary sensitivity can be shaped at T,, with Wz or T. with Wa, but they are multiplied with each other in the additive model uncertainty which cannot be achieved. The sensitivity can only be shaped with the plant, and the setting of the weights can be very difficult. The open-loop shaping procedure only has the open-loop transfer function for closed-loop shaping. In this case, the control design parameter was reduced to the parameter Wy. The limited number of parameter is simpli- 3. Control design fying the control design. Normally, the second weight W is used and would increase the performance if it were not to be omitted in the two degree of freedom design. 3.5.3. Feed-forward controller The controllers with a feed-forward part improve the control performance compared to a 1DOF controller, which is shown in Figure 3.31 and 3.32. It makes sense to only weight the velocity in the additional path because there are the same numbers of control inputs at the plant, independent of the design approach. The scheme in Figures 3.16 has this feature. The feed-forward compensator can easily be designed. Figure 3.33 and 3.34 show that the improvement in velocity and the good attitude behavior from the 1DOF design is retained. The 2DOF control scheme of the open-loop shaping method increases the quality of the step response. The bandwidth of the precompensator in Figures 3.37 and 3.38 is smaller. There is a smaller overshot in position, but also a slower step response. 94 ines ref and measured pos (m) eseco b 3.5. Conclusions and results 8 10 10 A 5 i 2 ., | 3 0 if oo ALL ous Fey i 10 a 0 2 8 10 time (s) Figure 3.27.: Step response in 2; direction with LQ (——) and Hx. 1DOF controller (—) based on the GS/T scheme: The Ho. controller overshoots in position more than the LQ controller, but the Euler angles are not oscillat- ing. This shows that the Hoo controller is primarily an attitude controller. 95 3. Control design , ref and measured pos (m) sso e v (m/s) Figure 3.28.: Step response in y; direction with LQ (-—) and H,. IDOF controller (—) based on the GS/T scheme: The lateral movement is more oscillating in position and velocity than in the longitudinal movement. The cross coupling is well suppressed because the pitch angle is not oscillating with the H,, controller. 96 3.5. Conclusions and results eeee 8 ok ref and measured pos (in) time (s) Figure 3.29.: Step response in 2; direction with LQ (-—) and H, 1DOF loop shaping controller (—): The higher bandwidth of this H.. controller is visible in position and velocity. The Euler angles are very smooth. 97 3. Control design = 04 3 ~ z 03 — 2 02 fi 3 1 fl 3 ol rh 5 cee : & An SN es BOERS ONS] 0.1 i 0 2 4 6 8 10 10 5 RB 3 0 md ° -5 -10 0 2 4 6 8 10 1 & > ~ -0.5 - oO 2 4 6 8 10 time (s) Figure 3.30.: Step response in y; direction with LQ (——) and H, 1DOF loop shaping controller (—-): The lateral movement is better controlled than in all other cases. There is no overshoot and no steady-state offset. 98 = 04 B o3 g g 0.2 = o1 5 2 -0.1 10 5 a 30 be ° 5 3.5. Conclusions and results -10f- + 2 8 10 ey aN fi bo { \ oF a ot = ~ b<3/N\. Te Oo 2 8 10 time (s) Figure 3.31.: Step response in <; direction with LQ (-—) and H.. 2DOF controller (—) based on the GS/T scheme: The response is smoother than in Figure 3.27. This is the advantage of a second input path of the controller. 99 3. Control design Bo Ff ref and measured pos (m) eecee time (s) Figure 3.32.: Step response in y; direction with LQ (——) and Ha. 2DOF controller (—) based on the GS/T scheme: The lateral movement is better controlled with the 2DOF controller. The roll angle is still oscillating. Both controllers have the same velocity response. 100 3.5. Conclusions and results ref and measured pos (m) ° u (m/s) time (s) Figure 3.33.: Step response in 2; direction with Hy IDOF (—) and Hoo 2DOF controller (——) based on the GS/T scheme: The same 1DOF con- troller is also included in the 2DOF controller, but the reference comes in through the feed-forward path. That is why there are different responses. The 2DOF design has a faster response. 101 3. Control design 2 ES es ref and measured pos (m) so ° bo 8, $ (deg) v (m/s) time (s) Figure 3.34.: Step response in y; direction with Hj. IDOF (—) and Hoo 2DOF controller (~—) based on the GS/T scheme: The lateral movement is better controlled in the 2DOF case. The improvement which is determined in Figure 3.32, is more visible by comparing the two responses. 102 3.5. Conclusions and results ref and measured pos (m) 8, (deg) u (m/s) time (s) Figure 3.35.: Step response in 2; direction with LQ (——) and H, 2DOF loop shaping controller (—): The 2DOF controller is faster and smoother than in all other controllers. The high deflection in the pitch angle is used for fast manoeuvering. 103 3. Control design ee of ref and measured pos (m) 2° £6 ° b ©, (deg) v (mis) time (s) Figure 3.36.: Step response in y; direction with LQ (——) and H.. 2DOF loop shaping controller (—): There is no overshoot. A small steady-state error is visible because the I part of the position controller is limited. Also, for the lateral movement, this controller is the fastest and smoothes of all controllers. 104 3.5. Conclusions and results &8 oF ref and measured pos (m) eseo ° & 05 u (m/s) 2 05 -1 time (s) Figure 3.37.: Step response in 2; direction with Hx. 1DOF loop shaping controller (—) and Ho. 2DOF loop shaping controller (-—): Also, in this case the 2DOF controller is slower than the 1DOF controller. 105 3. Control design ee So & ref and measured pos (m) © & 0.1 0 0.1 0 2 4 6 8 10 10 5 5 hte: 3 LIA 3 oft iy 2 es Boas ey S ish. dal p -10 0 2 4 6 8 10 v (m/s) time (s) Figure 3.38.: Step response in y; direction with Ha, 1DOF loop shaping (-) and H.. 2DOF loop shaping controller (-—): The same high dynamic of Figure 3.37 is determined in the lateral movement, too. 106 4. Helicopter test bench 4.1. Overview of the helicopter test bench Different small-scale unmanned helicopters have been developed in num- bers over the last few years. The main reasons are that the unmanned aerial vehicle (UAV) market is permanently growing and the applications for helicopters are extended from military use to the civilian sector. An overview of different small UAVs can be found on the internet, e.g., on the NASA home page, which is updated daily. The typical instrumentation of an UAV consists of an inertial measurement unit with gyros and accelerometers, and a GPS system which measures the position and the velocity. All this data has to be processed in a naviga- tion routine turning the measurements in noise-free and bias-free signals. The programming of the navigation routine requires good knowledge of the sensors and a correct calculation of the applied filters. The accuracy of velocity, attitude, and position depends of the quality of the sensor sig- nals, the precision, and the power of the onboard computer unit, and the mathematical level of the navigation routine. For a small air vehicle, the navigation part is quite complicated and often a source of error. The small-scale helicopter test bench described here avoids this problem. It is rigged to a mechanical structure that allows the deter- mination of the vehicle state variables with a small amount of on-board instrumentation. The helicopter body is attached to a mechanical frame consisting of three aluminum arms. There are three optical encoders below the detachable connection between the helicopter and the frame which are measuring the attitude of the helicopter with high precession. The measur- ing of the angles of the arms allows for the calculation of the position of the top point (see Section 4.5). The structure of the test bench is illustrated in Figure 4.2. All sensor and 107 4. Helicopter test bench Figure 4.1.: Helicopter mounted on the frame command wires are linked to the computer unit. It is a flexible development system allowing controller design and validation directly on the test bench. The dSpace real-time system allows rapid control prototyping, which is sought for the control of this test bench. ‘The single board computer board DS1103 with a fast processor provides the necessary computing power. For the use at the stand, an integrated incremental encoder subsystem is used for the attitude angles. Multiple analog-to-digital-converters (ADC) and digital-to-analog-converters (DAC) are available for the measurement of the rotor speed and the control unit for the servos. The system has a slave subsystem with additional I/O interfaces used primarily from a control panel. With a Real-Time Interface, programming with SIMULINK and MATLAB is straightforward. 4.2. Sensors and Actuators There are six optical encoders installed at the test bench with a resolution of 0.18 deg. The encoders give the angular signal in a digital form to the real-time system. The quality of signal would never be reached by any inertial measurement unit or analog encoder. Optical encoders have an index which orients the start of the rotation, During the installation of the encoders, the index is positioned near the start position. Before flight, the 108 4.2. Sensors and Actuators vans wack slave module u power supply DC 12V servos and electric board DAC power supply S23 for the Enel DC motor me pas nc eee Bnet control panel Bncs 9 DA potentiometer Eeeraoe toggle switch AD20 as Space Figure 4.2.: Schematic of the helicopter test bench 109 4, Helicopter test bench helicopter has to be initialized by being moved in all directions. An inertial measurement unit is attached at the head of the helicopter body. It measures the acceleration and the angular rates. These signals can be used with an Kalman filter estimating the Euler angles. The bias and the noise of the measurement unit are estimated exactly by comparing the Euler angels of the Kalman filter and the encoders. The quality of the sensors at the measurement unit may be investigated with this method. ‘The helicopter is controlled by three servos which drive the swash plate and one servo which drives the helicopter’s tail rotor. The swash plate moves when the actuators are mechanically pushing or pulling. A standard way for controlling servos is using a pulse of variable width, repeated periodically to specify the position (angle) of of the servo motor. The angle is determined by the duration of a pulse. The method forming the signal for the servos is called Pulse Width Modulation (PWM). The servo expects to see a pulse every 20 milliseconds (PWM period). The length of the pulse will determine how far the motor turns (PWM duty cycle). For example, a 1.5 millisecond pulse will make the motor turn to the O-degree position (often called the neutral position). The rotation of the shafts in normally limited from -60 to 60 degrees. The bandwidth of the servos is limited to 25 Hz. Because of the Shannon theorem, the bandwidth of the controller should not be larger than half of the maximal possible frequency of the actuators. The mechanical range of action is also limited by the construction of the rotor head. Normally, this is not critical because the servos are not reaching the limited stop. 4.3. Helicopter modules ‘The production of the signals for the servos has to be made quite properly and cannot be made directly at the real-time system. The modulation signal would have to amplified and the additional cable at the frame would add extra weight to the load of the helicopter. ‘The same problem exists when reading the signals of the inertial measurement unit. In (Vuskovic, 2001) a controller board was developed for this test bench to allow wireless bi-directional data communication between the computer unit (dSpace) and the model helicopter. The data coming from the motion 110 4.4. Drive pack is transferred to the real-time system and the helicopter receives the command signals for the servo actuators from this board. Thus the micro controller on the board samples the analog signals of the IMU and generates the PWM signals for the servo motors. Figure 4.3.: Module for data transfer ‘The connection from the board placed in the helicopter (slave module) can go over a serial connection to a wireless link to a second board (master module) which is placed on the ground. The master module manages the data transfer. The two modules can also be connected directly by cable with the integrated serial port. Because there is a dead time of 0.1 s from the wireless module, the cable solution was used for flight and to have no time delay anymore. 4.4. Drive At the test bench, a DC motor from Plettenberg, HP 355, was used which has a automatic stator ring adjustment to displace the cross-induction field when under load. The weight amounts to approximatly 720 g. The number of windings depends on the operating point of the rotor speed. In this case, a motor with five windings was chosen. The maximal possible power consumption lies at about 1800 W. At the operation point of 1500 [rpm] for the rotor speed, an efficiency factor above of 80% is reached. For power supply, there is a transformer rectifier which provide a maximal voltage of 40 V and a maximal current of 60 A by cable. The input of the supply is a control voltage (0..10 V). The transfer function of the motor 111 4. Helicopter test bench control unit has a cross frequency of we = 1 kHz. A Hall Sensor is used for the measurement of the rotor speed. There are two magnetic spikes inside the pinion of the shaft. With a frequency converter (DAC) to an analog signal, which measures in the span from 0 to 800 Hz, the double frequency of the motor is recorded. The DAC gives a signal from 0 to 10 V. Qmotor = Upac -80-0.5-60 [rpm] (4.1) 4.5. Delta robot arm for position The geometry of the DELTA Robot was developed at the Swiss Federal Institute of Technology in Lausanne (Devaquet and Brauchli, 1992). The construction is qualified to let the helicopter fly with an attached frame by reducing the additional load and making the position calculation quite clear. The determination of the position is made from clear geometric relations. Looking from above at the frame (see Figure 4.4) shows that it is necessary to nominate some lengths (a, ..., g) and corner points (z1, 22, 23). % by a do | — 10 cm. 125 5. Implementation and results ol ol 0s 0.05 € > 0 0 2 0.05 0.05 0.1 0.1 01 -005 0 005 0. 0.1 -005 0 005 O01 8z(m) 1 0.05 ce : E yO 38 ~0.05 0.1 “0.1 0.05 0 0.05 8x(m) 0.1 Figure 5.6.: Hovering in the xy, zy, and xz plane with the Hq. controller based on the GS/'T scheme 126 5.3. Results 0.1 0.1 0.05 0.05 dy (m) ° & ° 0.05 0.05 -0.1 -005 0 005 O01 ~0.1 -005 0 005 O01 8z(m) ol 0.05 Eo «e 0.05 0.1 0.1 -005 0 005 01 3x(m) Figure 5.7.: Hovering in the zy, zy, and xz plane with the Ho. controller loop shaping method The test flight with the H.. controller designed with the GS/T scheme, as shown in Figure 5.6, yields a better result than the LQ controller. The derivation in the zy plane is less than 4 cm. The third experiment used an H.. controller designed with the open-loop shaping method. As Figure 5.7 shows, the maximum derivation is just 1.5 cm. In all three experiments, the derivation is highest in the z direction, al- though it decreases from Figure 5.5 to Figure 5.7. Controlling the height of the helicopter at this test bench presents some difficulties which are not typical for model helicopters. Because the test stand is used in a closed room, a certain amount of air circulation is caused by the rotor. It flows along the floor and up the walls until on its return from the ceiling and flows through the rotor again. This circulation can be disturbed through nets placed above the floor, but a certain downwash remains and lets the 127 5. Implementation and results @ (deg) 4 (deg) Figure 5.8.: Attitude angles @ and ¢ from the hover experiments with the LQ controller (--) in Figure 5.5, with the H, GS/T controller (--) in Figure 5.6, and with the H.. open-loop shaping controller (-) in Figure 5.7 helicopter oscillate around the reference point. This concomitant of the height is not modeled, and even the Hoo controllers are not able to cancel this oscillation. The height range for this helicopter is also quite limited. In all three cases, the attitude of the helicopter at hover does not change very much. The LQ controller has a slow oscillating signal on the yaw and roll angles, which is due to the fact that the auxiliary rotor is not modeled. The two Hoo controllers succeed better at holding the helicopter in the air smoothly. The implementation for the Ha. controller with 2DOF is equal to that of the 1DOF case, except that the velocity reference enters the loop only through the feed-forward path. The hovering result (see Figure 5.9 and 5.10) does not change much because the feedback controller is provided very similarly in both designs. An overview of the experiments can be found in Table 5.2. The calculated controllers for the yaw angle are not very different from the 128 5.3, Results 0.05 A 0.05 Sy (m) e ° 0.05 ~0.05 0.1 0.1 “0.1 -0.05 0 0.05 01 01 -0.05 0 005 O01 82m) oO. o.0s 0 8 -0.05 0.1 “0.1 -005 0 005 0.1 3x (m) Figure 5.9.: Hovering in the ey, zy, and ez plane with the Ha. 2DOF controller based on the GS/T scheme Table 5.2.: Deviation o in all directions for the different experiments shown in Figure 5.5 to 5.10 x (em) | oy (em) | o: (em) | oy (deg) LQ 461 647 | 10.49 8.17 H., 1DOF GST 4.30 4.16 8.16 7.24 Hq. 2D0F GST 3.88 3.77 4.62 7.50 Hz IDOF open-loop shaping | 0.96 141 7.89 3.97 Hoo 2DOF open-loop shaping | _ 2.10 2.15 7.63 8.50 129 5. Implementation and results Ol Ol 0.05 0.05 Sy (m) ° ae ° 0.05 0.05 0.1 0.1 0.1 -005 0 005 01 0.1 -005 0 005 O01 5z(m) 0.1 0.05 = yO 6 0.05 0.1 “0.1 -005 0 005 O01 8x(m) Figure 5.10.: Hovering in the xy, zy, and az plane with the Hy. 2DOF controller loop shaping method 130 5.3. Results Table 5.3.: Measured values of the operating point in hovering variable name | operating value | variable name | operating value 9% 0.798 DEG Biso -1.515 DEG ¢o -2.400 DEG Ao -0.783 DEG Amo -7.261 DEG Aro 12.410 DEG various design approaches. The significant difference lies in the fact that the H., controllers have a limited bandwidth, yielding results equal to those obtained with the LQ controller. With the hovering experiments, the operation point for hovering is deter- mined by the command signals A;,, By,, Ar, and Ar, and the Euler angles 4 and ¢. In Table 5.3, the values of the operation points are provided. They were composed by the mean values of different experiments. 5.3.2. Hovering with wind generator The wind generator described Section 4.6 above can be used to test the performance of the flight controller for hovering under difficult conditions. The generator blows the wind in the direction of the helicopter at a constant speed or at varying speeds. ‘The situation for any helicopter is the same as when it meets various wind gusts or when it flies with a constant velocity. The helicopter can be aligned for the experiments in different ways, depend- ing on the directional derivation requested. ‘The reaction is very different whether the wind comes from the rear or the front. For helicopters to fly backwards is quite complicated because the stabilizing effect from the fin (horizontal stabilizer) is missing, which renders the system even more un- stable. In this work and for safety reasons, the behavior investigation of the craft’s is limited to the air flow from the front. Figure 5.11 shows the effects of the wind speed increased by the operator to a value of 4.3 % at the point where the helicopter is standing. The LQ controller holds the helicopter as smooth as possible. As the wind increases, the pitch angle @ decreases, which is to be expected because the change of the attitude of the tip path plane (and with it the main rotor force) cancels the induced drag force. The helicopter shows the same behavior when it flies with a forward velocity, which proves that the LQ controller can be 131 5. Implementation and results wind velocity ms) = y iJ a \ ° 10 20 30 40 50 60 70 2 + 1 - - - : r 1 . Bo (neo Un : Aylin 3 \ i nm Ne 2 . wy fw \ 2 way PN 2 . NI ry 10 20 30 40 50 60 70 ol r + + 1 Y r time (s) Figure 5.11.: Hovering in the wind field of the wind generator with the LQ controller 132 \ wind velocity (m/s) i 0 20 40 60 80 100 120 140 160 180 0.1 T T T T T T T T 0.05} : : : g m Se Ofrapaat ares AA eying Vi & : 0.05 : : 0.1 i) 20 40 60 80 100 120 140 160 180 time (s) Figure 5.12.: Hovering in the wind field of the wind generator with the Hoo controller utilized even for slow forward flight. For safety reasons, the maximum possible wind speed was not tried out because at the velocities tested the attitude of the helicopter was too excited already. The streamed helicopter has also got an offset in position, although there is an integral part in the position controller. In addition, the LQ controller gets an offset in velocity which is cancelled by the offset in the pitch angle 0. The H.. controller that was designed using the GS/T scheme handles the wind disturbance better than the LQ controller. The wind speed could be increased to 5 m/s. Figure 5.12 shows that the pitch angle 0 increases to 133 5. Implementation and results the same maximal value of 2 deg as in the case of the LQ controller, but the offset in the position (2p direction) is smaller and the helicopter moves more smoothly. The same result could not be obtained with a H,, controller designed with the open-loop shaping approach. In an experiment, the controller was hold- ing the position exactly and the attitude was also very stable up to a wind velocity of 3 m/s, after which the helicopter became quite jumpy. Due to the high bandwidth of the controller, there are fast reactions to not modeled effects. 5.3.3. Hovering with a disturbance input For the identification of the system parameters introduced in Section 2.5, measurements with dynamic flights are required. These can be obtained by exciting the helicopter through fast trajectories in the reference signals (steps, ramps, white noise, etc.). Figure 3.27 to 3.38 show the results of these experiments which can then be used for identification. A better result for estimating the parameters is reached by exciting the helicopter at the command input as described in Section 2.5. The exci- tation is realized with a discrete white noise signal which is added to the command signal from the controller. Because the experiment requires that the helicopter remains stable, the controller has to manage the disturbance signal. From this point of view, it is also a test for the robustness of the controller. The first test was made with a LQ controller as shown in Figure 5.13, and a band limited white noise signal with a deviation of o=1deg/s was added on the command signal for pitch By,. After 5 s the input signal has to be stopped and for 3 s, the helicopter is oscillating. The LQ controller is not able to achieve stability with this disturbance input. The Hoc controllers have less trouble dealing with this type of input. Figure 5.14 shows the same signal used with an H.. open-loop shaping controller. The helicopter was visibly excited but did not become unstable. Later experiments with the same controller but with deviations up to 1.2 deg/s showed that the helicopter remained stable. 134 5.3. Results s b i = 7 | 4 3 Ss ° 8 (deg) rig Les wd Ay Figure 5.13.: Hovering with a disturbance input (white noise, with ¢ = 1) on the command signal Bi, using the LQ controller. In the first plot, the additional input (upper signal) and the resulting command signal (lower signal) are shown. 135 5. Implementation and results s ow noise and B, , (deg) 0 5 10 15 20 25 0 5 10 15 2 25 lA AMM aN \ i\ ty lyn WA x(m) y(m) 0.06 0.1 0 5 10 15 2 25 Figure 5.14.: Hovering with a disturbance input (white noise, with o = 1.1) on the command signal By, using the H, 2DOF open-loop shaping controller 136 5.3, Results 5.3.4. Discussion With these three different experiments (hovering, hovering with wind, and hovering with disturbance input), the ability to hold the helicopter at the same point under different circumstances is tested and investigated. The step responses in Chapter 3 are meant for the general view of cancelling the steady-state offset, bandwidth, and closed-loop behavior (particularly the characteristic of T, or T,). However, the influences of disturbances in signals and modeling are investigated in this section, in other words, the sensitivity S, or S, are regarded. The sensitivity S, is the transfer function from a disturbance input on the measurement vector. Since the signals are certainly not noise free, the normal hovering test is a first quality statement if the sensitivity is small up to frequency not perturbing. Also, not modeled influences or changes of the design plant could be affected by the sensitivity S.. The hovering in a wind field, which is equivalent to a slow forward flight, belong to this class of distraction. The disturbance input on the input signal in Section 5.3.3 directly affects Su, which is shaped in the control design with Hoo controllers. This could not be done directly with a classical or LQ design. Finally, the hovering quality was increased through using model-based ro- bust controllers, which allow a higher bandwidth of the closed loop and good disturbance rejection. The experiments show the clear advantages of these controller methods. 137 5. Implementation and results 138 6. Summary and Outlook The design of the controller part of an autopilot for an unmanned helicopter was successfully demonstrated in this thesis. The theoretical background of helicopter aerodynamics, identification, and controller design was presented and applied to a helicopter. The results in this work and the outlook of activities in the future are shortly described here. Summary In the first part of the thesis, a mathematical model of a small helicopter was deduced and investigated. The equations of translational and rotational motion were derived. They were referred to the helicopter body with the origin at the center of mass of the aircraft. The applying aerodynamic forces and moments were calculated with the actuator disc theory, which required the derivations of the induced velocity, the aerodynamic forces at the blade, and the flapping angle of the main blade. An important aspect was the modeling of the flapping angle dynamics of the servo rotor. While not being measurable, this angle is important for the description of the helicopter behavior. The resulting helicopter model was a nonlinear model of 19th order. Lin- earizing the model in the hovering operation point, the linear model was discussed and used in the identification step as a parameterized model. Dif- ferent aspects of closed-loop identification were discussed and the resulting identified parameters for models with varying structure were presented. In the second theoretical part of this work, the controller concept within the autopilot structure was explained. The flight controller consisted of two different parts: the position and the attitude controller. There were three different attitude controller designed and implemented. The first was a LQ controller able to stabilize the helicopter. Then, two H.. controllers were presented with two different design methods. One H,, controller was cal- 139 6. Summary and Outlook culated by shaping the closed-loop behavior (GS/T scheme), the other was designed by using the open-loop shaping method, which used less design parameters. Both approaches were realized with an additional feed-forward path. Measured step answers in position were compared with the LQ con- troller and the Hoo controllers, which showed a clearly better result. A test bench was presented which consisted of a small helicopter mounted on top of a mechanical frame that allowed free motion. All sensors and electric equipment were described as well as a wind generator, which produces an air stream and which was used for testing the robustness of the designed controllers. These hover experiments were discussed in the last part of this work. Outlook The controllers that have been designed and implemented a lead to good results, and the presented helicopter test bench proved to be a reliable construction giving signals of high quality. However, several aspects, par- ticularly in identification, should be investigated in more detail, e.g.: © The proposed off-line identification methods were applied and two applicable models (based on the nonlinear model) were presented. It would be interesting to see if another structure exists with less pa- rameters, so that an online identification method can be implemented. The new model would be adaptive for the different flight modes. Based on online identified models, the controller could be adapted au- tomatically in the different flight modes. A linear parameter varying Ho. controller could be implemented. Hover and forward flight are not the only challenging flight modes. Good pilots for small size-helicopters can fly rolls with the helicopter and turn the helicopter with the main rotor upside down. An acro- batic flight controlled with a feedback and feed-forward compensator should be possible. Another flight maneuver is autorotation, which is very difficult to perform for an experienced pilot, The aerodynamic of the main rotor has to be described by the induced velocity and by the blade drag. A sequential controller algorithm has to be developed and integrated in the autopilot. 140 The last two aspects cannot be investigated on the test bench described in this thesis. A free flight helicopter with a diameter of 1.8 m should be used for such tasks. 141 6. Summary and Outlook 142 A. Navigation The description of the navigation routines are based on the books of (Farrell and Barth, 1999) and (Rade and Westergren, 1997). A.1. Earth-to-navigation-plane transformation The relation between points and vectors in the navigation and earth-plane coordinate systems can be completely described by the rotation matrix C®. It can be defined by a series of three plane rotations involving the Euler angles $, 0, and ¢ for the x, y, and z-axis. ia) 0 Re=| 0 cosp —sing (A.1) 0 sind cosd cosé 0 sind Ry=| 0 1 0 (A.2) —sind 0 cos@ R.=| sinp cosp 0 0 0 1 [> sin’ "| (A.3) The first rotation rotates the earth coordinate system by ¥ rad about the z axis. The plane rotation for this operation is described as Rex” (A.4) 7 The second rotation rotates the coordinate system that resulted from the previous yaw rotation by @ rad about the y! axis. (AS) 143 A. Navigation And the last rotation rotates the coordinate system that resulted from the previous pitch rotation by ¢ rad about the x” axis. Rex” (A.6) So finally it is Ch = ReRyR: (A?) A.2. Euler angle derivatives The equation (2.2) shows how the transformation matrix for the Euler angle derivatives is calculated. It is a composition of three transformations of the unity vector. First, p=o (A8) when d¢/dt is the only rotation. Second, note that @ would correspond to the angular rate vector de’... Transforming this vector to the body frame by @ (see equation (A.5)) produces 0 cos @ (A9) r -sing P . . q | = bem = ORa Rye, for the 6 rotation alone. Third, note that ¢ corresponds to the angular rate vector Je.n. Transforming this vector to the body frame by ¢ and 6 produces Pp | _[ sing q | =en =bReRyRe,» =| cosdsind (A.10) r cos 0 cos 6 for the a rotation alone. Therefore, when all the angles change simultane- ously, the transformation is given by P 1 0 —sind 6 q|=| 0 cosd singcosé @ (A.11) r 0 -sing cosdcosd | | 144 B. Equations B.1. Nonlinear plant B.1.1. Constant Values Su = Rix Sr =aRp Neave = Meave~Z f mu 3 Rhy — rf 9M = P lieaveCaa—¢@—$* Treave 2 8 “My = P NcaveCaa TH Lteave 4, — rh P= P lpadCaap Ta wo = ¢ ymQu ~ 16w 26D ux $x = ~ arctan Ono B- (B.1) (B.2) (B3) (B.4) (B.5) (B66) (87) (B8) (B.9) 145 B. Equations Mhelitotal J Vinero =F 5 Sir Muwo at Fraitroto Vinero = fem B.1.2. Variable values Fraitrotoro = t Vie = 28 4 2" + vial — _3RiViv O° = RR, — re At = kparbip(t) + ks2rAis(t) By = kporaip(t) + hear Bis(t) Av Onew(t ee aap AM + oe 2 ant Geen ) a1 =—sindx ams Bimh B/ (OK, — WG)? + AC? wHN, amy Armn + cos 646 —— “8 /(OR, — way? + ACwROR, F amr Arm b; = —sin 2 Sn OI TE ak)? + 1a, cos dx mr Bimh 8) (Vy — WH)? + AWN, Ann = Ay — 20. — 1608), A(R, = re u(ta0 OQ ~ FOr > BRE — rh Qa ARs = rol) (4 aa + 3CRE = FE) Qar AMO + OP 146 (B.10) (B.11) (B.12) (B.13) (B.14) (B.15) (B.16) (B.17) (B.18) (B.19) (B.20) B.1. Nonlinear plant 16q(t) , 4(Riy — rie)v(t)a0 yOu) 8(RE, — rh) Qu ~ 02 ays(y 4 2% Bin = By + PO) On Fo, (t) = — sin (t) g mrewrotat Fy, (t) = sin 6(t) cos 0(t) g mitetirotat Fy, (t) = sin @(t) cos 0(t) g MHetiTotat Ri = Tit o2 T(t) = 0.80 cave Caa{ ——M EO, Au (t) — 0.22A0,0a¢ Rig =the u(t) v(t) 4 On (A Bi — At sul) + + ua _ Ob) ua , wah) 2 2 ztr(t) + v(t) 5 \ + Virr(t)? 3 Ri View 2(Ry = rp) View = — Avri = oe T(t) = 0.759 litcave Urcaa ( Re = (ar(y — 2 a) Rh — rh r(t)at + v(t) 2 Qr ) F,(t) = Fy, (t) — T(t) sinax Fy (t) = Fo, (t) + T(t) sin: + Tr(t) F..(t) = Fy, (t) — T(t) cos ax cos bi Metuten(t) = mom Matotor(t) — Nivom ImotorS2(t) — Mom IMotor*(t) (B.21) (B.22) (B.23) (B.24) (B.26) (B.27) (B.28) (B.29) (B.30) (B.31) (B.32) 147 B. Equations a 2, ( Ric=rhe (_ a, 4) — Avs) Ave O10 = Blanc ( Bt Au(O~ oer) Dae 4 (Rh = rho) Au(p — 20% u(y) ~ Bethe) 30m Oe A, - 14 Av + Bhatt ( Bart (cm — cut (auc + ae) +eu2 (amo + #)) + UN Bie — ed (em ere 42 (4uco + A “)) vette) (B.33) Or 2(Rp — 7) (en rat — (AZO — Ar (oat + vena") 3 OF FA Avi - Blatt ( Bt BET up (ano - 3 ) Ave \ Avis Ort) = ~Stnncnatt( BESTE (an - 3") vt 4 Re —2ReE th Catt (ars w)) (B34) Me = #eyob1 (B35) M, = ero (8.36) Muoit(t) = Qr(t) — nator Heat X(t) (B37) Marivey(t) = Meait(E) (B.38) Marivez(t) = —Mimotor(t) + (1 + ivoas)Metuten (t) + nor Mrait(t) (B.39) L(t) = Mz — 2m T(t) sinbl — 2t Tr(t) (B.40) 148 B.1. Nonlinear plant M(t) = My + Marivey(t) — zm T(t)sinal + 2m T(t) cosa; cos bi (B.41) N(t) = Marivez(t) — 2m T(t) sin bl + xt Tr(t) (B.42) B.1.3. Differential equations mors (iu(t) + w(t)a(t) ~ v(t)r(t)) = F(t) (B43) Mrorz (H(t) + u(t)r(t) — w)p(t)) = Fy(t) (B.44) Myert (w(t) + v(t)p(t) — u(t)a(t)) = F(t) (B.45) Texb(t) = (Ivy — Lez)a(t)r() + Lee (F(t) + p(t)a(t)) + LO) (B.46) TyyG(t) = (Tez — Inz)p(t)r(t) + Ina (r(t)? - p(t)”) + M(t) (B.47) Tez¥(t) = (Lez — Iyy)Q(t)p(t) + Taz (H(t) + p(t)a(t)) + N(t) (B.48) Mmotor = Cmotor Mmotor(t) + bmotor Umotor(t) (B.49) Q(t) = kmotor Mmotor(t) + beot Anr(t) + cn M(t) (B.50) Visenlt) = a (- Vaarn(t) + Viren PO ‘) (B51) Virr(t) = = (-virato + Ving nO (B.52) ap(t) = uae (-ante + csp Big (t) + x) -4(t) (B.53) big) = EBM (g(t) + esp Ault) 2) —pte (654) 149 B. Equations 150 C. Data of the helicopter and model C.1. Helicopter C.1.1. Parameters Weights Mhelicopter = 3.5 kg Mframe = 38kg Data of main rotor Rm = 0.725 m Tm = 012m Trotor = 0.066446 kg/m? Tyr 0.15 s Motade 0.211 kg lbiade = 0.062 m on = pat ers = ssszoaus esp = o7goz08 cro = 259406 ho 0.041 m Cac 5 wo = 0.01449 Cw = 0.08905 cuz = 0.39076 151 C. Data of the helicopter and model Data of tail rotor Re = 015m Te = 0.06 m Trotor = 0.934 1074 kg/m? Tyr = 01s Molade = 0.008575 kg Iotade = 0.032 m Cao = 5 CwitT = 21 Data of the servo rotor leervo = 0.05 m Ry = 0.35 m Tp = 0.24m Tserve = 0.001707 kg/m? Caas = 267 Drive data nMom = Bf "Mor = # Ou 163.363 [rad/s] Om = 653.452 [rad/s] ea = 27.309 Ta = 0.043024 s komega = 2.589 Kumot = ~0.0036 kam = 0.41098 152

You might also like