2007:086 MASTER'S THESIS UAV Stabilized Platform Martin Ernesto Orejas Luleå University of Technology Master Thesis, Continuation Courses Space Science and Technology Department of Space Science, Kiruna 2007:086 - ISSN: 1653-0187 - ISRN: LTU-PB-EX--07/086--SE UAV Stabilized Platform Ing. Orejas, Martín Ernesto Thesis Supervisors: Ing. Hromcik, Martin, Ph.D; Ing. Harlin, Gösta Master Diploma Thesis for Erasmus Mundus programme SpaceMaster Department of Space Science Kiruna Space Campus Department of Control Engineering Faculty of Electrical Engineering Czech Technical University Prague –May 2007 Abstract Abstract This work develops a performance analysis for a particular distribution of inertial sensors as they are assumed to be implemented on a prototype stabilized platform for a new UAV aircraft (under development at VTUL PVO and AVEKO now). The device to be stabilized is a 2-axis gimbal – inner and outer gimbal axis- that will have an attached array of optical and electric sensors at one of its axis. Gyroscopes are planned to be used to sense the angular rates of this highly non-linear and coupled system. The whole set of equations for the complete system is introduced with a specific focus on the inner gimbal dynamics. The use of different observers to filter the data obtained by the sensors is analyzed along with the effectiveness of various types of controllers. The inner gimbal system is modeled and several simulations are performed to study the effect of different practical contraints -actuator saturation, sensor bandwidth, sampling frequency, etc., in addition to the performance for different disturbance frequencies and controller parameters. Key words: LOS stabilization, gyroscopes, disturbance rejection. i Table of Contents Table of Contents 1. Introduction ............................................................................................1 1.1. Introduction to stabilized platforms .............................................................................1 1.2. Objetives.......................................................................................................................3 1.3. Thesis Layout ...............................................................................................................3 2. Gyros, motors and gimbal geometry ......................................................5 2.1. General characteristics of gyroscopes ..........................................................................5 2.1.1. Gyro specifications .....................................................................................5 2.1.2. Gyro modeling ............................................................................................6 2.2. Brush type DC motors..................................................................................................7 2.2.1. Motor modeling ..........................................................................................9 2.3. Gimbal..........................................................................................................................9 3. Dynamics and kinematics.....................................................................11 3.1. Basic coordinate frame transformations.....................................................................11 3.2. Eul e r ’ smome n te qua t i on s..........................................................................................14 3.3. Inner gimbal dynamics ...............................................................................................14 3.4. Outer gimbal dynamics ..............................................................................................19 3.5. Augmented inner gimbal dynamics –sensor dynamics .............................................22 4. Observers and controllers .....................................................................25 4.1. Observer for inner gimbal ..........................................................................................25 4.1.1. First order approximation .........................................................................25 4.1.2. Extended Kalman filter .............................................................................35 4.2. Controller for the inner gimbal...................................................................................41 4.2.1. Constraints for the controller ....................................................................45 ii Table of Contents 4.2.2. Alternative controller for the inner gimbal ...............................................53 5. Conclusions ..........................................................................................62 6.1. General conclusions ...................................................................................................62 6.2. Future work ................................................................................................................63 6. References ............................................................................................64 iii Figures List Figures List Fig. 1.1. Descriptive picture of LOS system................................................................................2 Fig. 2.1. Gyro circuit and block diagram.....................................................................................6 Fig. 2.2. 3D representation of the motor......................................................................................8 Fig. 2.3. Gimbal .........................................................................................................................10 Fig. 3.1. Different coordinate frames for a 2-axis gimbal .........................................................11 Fig. 4.1. Torque couse by Coulomb friction..............................................................................26 Fig. 4.2. Real vs. estimated friction torque ................................................................................27 Fig. 4.3. Zoom over transition area............................................................................................27 Fig. 4.4. System–observer diagram ...........................................................................................31 Fig. 4.5. Real vs. estimated CFT................................................................................................31 Fig. 4.6. Amplification of real vs. estimated CFT .....................................................................32 Fig. 4.7. Real vs. estimated CFT for high MN ..........................................................................32 Fig. 4.8. Amplification of real vs. estimated CFT for high MN ................................................33 Fig. 4.9. Re a lv se s t i ma t e dωIy ...................................................................................................34 Fi g .4. 10.Zoomov e rr e a lv s .e s t i ma t e dωIy .................................................................................34 Fig. 4.11. Real vs. Estimated CFT using EKF.............................................................................37 Fig. 4.12. Zoom over real vs. estimated CFT using EKF ............................................................38 Fig. 4.13. Zoom over real vs. estimated CFT using EKF for high MN .......................................38 Fig. 4.14. Zoom over real vs. estimated CFT using EKF for extremely high MN ......................39 Fig. 4.15. Se ns orou t pu tv s .ωIy....................................................................................................39 Fi g .4. 16.Re a lv s .e s t i ma t e dωIy ..................................................................................................40 Fi g .4. 17.Zoomov e rr e a lv s .e s t i ma t e dωIy .................................................................................41 Fig. 4.18. a) Angular velocity before and after control signal is applied.....................................42 iv Figures List Fig. 4.18. b) Zoom over angular velocity before and after control signal is applied...................43 Fig. 4.19. Angular velocity before and after control signal is applied (2º controller) .................43 Fi g .4. 20.Zoomov e rωIy using 2º controller ...............................................................................44 Fi g .4. 21.Zoomov e rωIy using 3º controller ...............................................................................44 Fig. 4.22. System response with and without quantization..........................................................46 Fig. 4.23. Sensor output after quantization..................................................................................47 Fig. 4.24. Zoom over sensor output after quantization ................................................................47 Fig. 4.25. Control torque with and without saturation.................................................................48 Fi g .4. 26.Zoomov e ra ng ul a rv e l oc i t yωI ywi t ho uts a t u r a t i on....................................................49 Fi g .4. 27.Zoomov e ra ng ul a rv e l oc i t yωI ywi t hs a t u r a t i on.........................................................49 Fig. 4.28. An g ul a rv e l o c i t yωI yf or50Hzs a mpl i ng....................................................................51 Fi g .4. 29.An g ul a rv e l o c i t yωI yf or300Hzs a mpl i ng..................................................................52 Fi g .4. 30.An g ul a rv e l o c i t yωIy for 1000Hz sampling .................................................................52 Fi g .4. 31.An g l eθ Iy for 50Hz sampling .......................................................................................54 Fig. 4.32. Complete system diagram for inner gimbal stabilization ............................................55 Fig. 4.33. An g l eθ Iy for 50Hz sampling and LQ controller..........................................................56 Fi g .4. 34.An g ul a rv e l o c i t yωIy for proportional controller .........................................................56 Fi g .4. 35.An g ul a rv e l o c i t yωIy for LQ controller........................................................................57 Fi g .4. 36.Zoomov e ra ng ul a rv e l oc i t yωIy for LQ controller ......................................................57 Fig. 4.37. Zoom over real angle vs. estimated angle ...................................................................58 Fi g .4. 38.An g l eθ Iy for 50Hz sampling and proportional controller ...........................................58 Fi g .4. 39.An g l eθ Iy for 50Hz sampling and LQ controller..........................................................59 Fi g .4. 40.An g ul a rv e l o c i t yωIy for 50Hz sampling and proportional controller.........................59 Fi g .4. 41.An g ul a rv e l o c i t yωIy for 50Hz sampling and LQ controller .......................................60 Fi g .4. 42.An g l eθ Iy for LQ controller in tracking mode..............................................................60 v Symbols Glosary Symbols Glosary LOS Line-of-sight PVG Piezoelectric vibrating gyroscope TFGYRO Transfer Function of gyroscope sensor KM Motor constant (motor performance index) RθPO Rotational transformation from platform frame to outer frame Rθ OI Rotational transformation from outer frame to inner frame OI Relative angular velocity between inner and outer gimbal PO Relative angular velocity between outer gimbal and platform wP(t) Angular rate of the platform frame wO(t) Angular rate of the outer frame wI(t) Angular rate of the inner frame M Applied moment h Angular momentum I Moment of inertia vector G Gravity gradient moment vector Tel Elevation stabilization control torque T fI Friction torque about the y-axis of the inner gimbal Kcf Coulomb friction coefficient Kif Viscous friction coefficient Taz Aximuth stabilization control torque ωIy Angular velocity over the inner gimbal y-axis vi Symbols Glosary e(t) Measurement noise Pk Estimated error covariance in time k Kk Kalman gain in time k MN Measurement noise CKT Coulomb friction torque EKF Extended Kalman filter SP Sampling period SF Sampling frequency KP Proportional gain L LQ optimal feedback gain vii Section 1. Introduction 1 Introduction 1.1 Introduction to stabilized platforms. In many engineering applications, a complete servomechanism system often comprises multiple channels, or axes; for example, a multiple-axis stabilized platform or a multiple-axis machine tool. The control of such multivariable servomechanisms is, in general, not a simple problem, as there exist cross-couplings, or interactions, between the different channels [6]. In the case of a multiple-axis stabilized platform, those cross-couplings include extra nonlinearities that make the task even more difficult. These nonlinearities come not only from the dynamics of rigid bodies but also from the nonlinear behavior of the bearing friction. This friction is produced between the motors, used for stabilization, and the axis of rotation. Stabilized platforms have various uses in many applications. One of the most common applications is for stabilization of the line-of-sight (LOS). The LOS stabilization system is a system that maintains the sightline of an electro-optical sensor when it is subjected to external disturbances such as base motion [5] Sensing equipment, such as electronic imaging devices, cameras, radars, navigation instruments, and the like are frequently carried by and operated in a moving vehicle, such as an airplane, that undergoes rotational motion about its center of rotation. In such an environment, the equipment is typically mounted on a movable platform that is stabilized with respect to vehicle movements. The stabilization may be about one, or more, of the vehicle axes. Particular applications involve the LOS stabilization of a camera, or other imaging device. In such cases, when the vehicle undergoes rotational motion about its axes, the LOS remains fixed with respect to an inertial reference frame. This is accomplished by sensing these angular disturbances and generating the 1 Section 1. Introduction necessary counter-rotations [1]. Typically, gyroscopes haven been used to measure these angular disturbances, although recently, the use of linear accelerometers has been proposed [7]. For this project, Piezoelectric Vibrating Gyroscopes (PVG) has been chosen as the sensors to use. These kinds of gyros overcome the disadvantage of the traditional jewel-bearing gyros without the need of the complex mathematics, involved with the linear accelerometers, of transforming linear accelerations in angular velocities. For our project, a 2-axis gimbal - that will contain one visible spectrum camera (Sony, weight approx.0.3 kg), one infrared camera (Miricle, 1.3 kg) and one laser rangefinder (Vectronix, 0.5 kg) - will be used. The electro-optical sensors should be able to keep their aim fixed at the target whether this is moving or still. Figure 1.1 shows a descriptive picture of the system. Figure 1.1 Descriptive picture of LOS system For achieving this, a multiple sensor array (of gyroscopes) will be attached to each of the axes, with the possibility to aggregate sensors in the platform itself and tachometers in the rotation points. In the last few years much research has been done in LOS-stabilization systems field. Many control techniques has been proposed, neural networks [], fuzzy logic [], robust control [], adaptive control []. Although most of these works fail in taking into account many real variables of the system, like gravity gradient torques, noise and frequency response of sensors or saturation level for servomechanisms, they have shown some improvements over typical control methods and could be useful for future upgrade of this work. Instead, we will circumscribe to 2 Section 1. Introduction more traditional controllers putting larger effort in analyzing how every variable of the system affects the overall performance. 1.2 Objetives The main goals of this work are: propose algorithms for the controller driving the motor that will be placed over the axis of rotation of the inner gimbal; and to analyze the effect of several practical constraints on the system performance. The controller will have the mission of stabilizing the Line-Of-Sight (LOS) of a cluster of cameras and pointing devices, which will be mounted in the 2-axis gimbal. The characteristics of both, motors and gyros, are provided by the companies producing the components. 1.3 Thesis Layout This work is divided in four sections. The first is this introduction to the principles and uses of multiple-axis stabilized platforms especially it’ sapplication to stabilization of the line-of-sight (LOS). Section two is devoted to briefly explaining the functioning and technical characteristics of the main devices to be used in the application, i.e. the 2-axis gimbal, the motors and the gyroscopes. Section three begins with an introduction to the geometric transformations used to go from one to another of the multiple-axis gimbal coordinate frames (inner frame, outer frame and platform frame). Then the rigid body equations of the inner gimbal are developed, describing each component of the torque vector. The same development is done in the next subsection for the outer gimbal dynamics. In subsection 3.5, after simplifying the equations using symmetry properties of the inertia matrix, a space-state representation that includes the sensor dynamics is developed. Section four is divided in two parts. The first deals with the design of the observer for the inner gimbal system where the second is focused on the controller and its constraints. 3 Section 1. Introduction In subsection 4.1, two different approaches for designing the observer algorithm are presented. First a first-order linear stochastic differential equation was incorporated to the system of equations to model the nonlinear friction and a discrete Kalman filter was used to estimate the angular velocity over y-axis and the nonlinear friction. In the second approach the complete nonlinear equations were involved and the estimations were performed using an Extended Kalman filter and assuming the parameter Kcf as one of the state variables. In subsection 4.2, two different control techniques are analyzed and several practical constraints for its implementation are studied. First, a simple proportional controller is developed. Three variations of this controller were simulated to study the improvements of canceling the nonlinear friction component and of using the estimated states, instead of the sensor output. The next subsection analyzes several practical constraints that appear when implementing the controller; these were: quantization step size of the A/D converter, s a t ur a t i onsl e v e l sf o rmot or ’ s tolerable maximum torque, sampling frequency used in the algorithms, measurement noise and bandwidth of the sensors. The influence and limitations posed by these constraints was analyzed. The last subsection presents an alternative augmented observer that will be used to design an LQ-controller for minimizing both variables: angular velocity and angular displacement around y-axis of the inner gimbal. Section five presents a summary with the main conclusions along for suggestions for future research in the field. 4 Section 2. Gyros, motors and gimbal geometry 2 Gyros, motors and gimbal geometry 2.1 General characteristics of gyroscopes A gyroscope is a device for measuring or maintaining orientation, based on the principle of conservation of angular momentum. The essence of the device is a spinning wheel on an axle. The device, once spinning, tends to resist changes to its orientation due to the angular momentum of the wheel. In physics this phenomenon is also known as gyroscopic inertia or rigidity in space. 2.1.1 Gyro specifications For the development of this project Piezoelectric Vibrating Gyroscopes (PVG) has been chosen as the sensors for retrieving information of the angular velocities around each axis. This kind of gyroscope is an angular velocity sensor that uses the phenomenon of Coriolis force, which is generated when a rotational angular velocity is applied to the vibrator. This device is surface-mountable so it is possible to be mounted to the desired location by an automatic surface mounter. Its ultra-small size, of about 0.2cc, and lightweight shape increases flexibility of installment and allows the sensor to be attached with no noticeable increase in the size of the apparatus where it is planed to be placed. Features Ultra-small and ultra-lightweight Quick response Low driving voltage, low current consumption Lead type : SMD 5 Section 2. Gyros, motors and gimbal geometry Reflow soldering (standard peek temp. 245 degree C) Characteristics Supply Voltage (Vdc) Maximum Angular Velocity (deg./sec.) Output (at Angular Velocity=0) (Vdc) Scale Factor (mV/deg./sec.) Linearity (%FS) Response (Hz) Weight (g) 2.7~5.25 +/-300 1.35 0.67 +/-5 50 max. 0.4 Table 2.1 Gyroscope characteristics Application To reduce the effect of temperature drift (due to change of ambient temperature), a high pass filter must be connected to sensor output to eliminate DC component. To suppress output noise component around 22-25 kHz (resonant frequency of sensor element), a low pass filter which has higher cut-off frequency than required response frequency must be connected to sensor output. The following figure shows a sketch of the gyroscope circuit along with a simplified block diagram that shows how the output of the sensor goes through the high-pass filter to eliminate the DC drift, then to the low-pass filter to reduce noise and from there to the A/D converter to have a discrete signal that the processing circuit and handle. Figure 2.1 Gyro circuit and block diagram 2.1.2 Gyro modeling To incorporate the dynamic of the gyros in the future models of the system we need to have a model of the gyroscope compatible with the rest of the models used. In this case, an attempt to obtain a transfer function with angular velocity as input was done. From the transfer function it 6 Section 2. Gyros, motors and gimbal geometry is straightforward to get the state-space representation if required. For this particular case a second order transfer function with cut-off frequency of 50Hz was chosen. TFGYRO r2 2 s 2r s r2 (2.1) where r : resonance frequency 50 Hz : damping factor 0.7 2.2 Brush type DC motors DC Torque Motor Characteristics One of the most useful rotating components available to the control system design engineer is the direct-drive DC torque motor. This versatile control element is a permanent-magnet, armature-excited, continuous rotation motor with the following features especially suited to servo system drive and actuation applications: No gear train Direct mounting on the driven shaft High torque at low speeds High torque-to-inertia ratio High torque-to-power Linear torque speed characteristics Low electrical time constant Convenient for factors Simple, rugged construction Smooth operation These features make it possible for the designer to obtain such system performance characteristics as: High coupling stiffness Fast response Precise positioning High tracking accuracy Excellent stability Low input power 7 Section 2. Gyros, motors and gimbal geometry Smooth and quiet operation Compact assembly Improved system reliability Figure 2.1 shows a 3D representation of the motor to be used in this project Figure 2.2 3D representation of the motor Motor data: Tp: Peak Torque. This is the maximum useful (non continuous) torque (in ounce-inches) that can be obtained at maximum recommended current input. KM: Motor Constant. This is the ability of a servo motor to convert electric power input to torque, a kind of figure of merit that can be used to compare motors in their ability to produce torque per unit of power input. It is the ratio of torque to the square-root of the power input. TF: Total Breakaway Torque. The friction contributed by the motor to the system determines the total breakaway torque (in ounces-inches). It is the sum of the brush-commutator friction, plus the magnetic retarding torques such as hysteresis drag and slot effect drag. JM: Moment of Inertia. The moment of inertia of the armature is measured about the torque mot or ’ sa x i so fr ot a t i on. TR: Ripple Torque. A small change in torque with armature position is caused by the switching action of the commutator. The armature rotates through a small angle before its field is returned to its original position through commutation. This variation is known as ripple torque and is usually expressed in percent of torque level. 8 Section 2. Gyros, motors and gimbal geometry 2.2.1 Motor modeling There are two main transfer functions concerning the dynamics of the motor. One related to the speed/power relation and the other one with the torque/power relation. In our case, because what we need is to cancel the disturbance torques, we are interested in the second relation. According to the motor manual this a static equation and therefore there are no dynamics involved. This can be easily understood if we consider that torques produced by electromagnetic forces act almost instantly compared to mechanical movements. Here is shown the equation that show this relation Torque K M Power (Watts ) (2.2) Where KM is the motor performance index From this equation it is clear that after calculating the necessary torque the motor has to exert it is straightforward to find the needed power. 2.3 Gimbal A gimbal, also called a gimbal ring, is a mechanical device consisting of two or more rings mounted on axes at right angles to each other. An object mounted on a three ring gimbal will remain horizontally suspended on a plane between the rings regardless of the stability of the base. Gimbals have a wide range of practical uses including aerospace applications. A gimbal may be used to keep objects level in unstable environments. Gimbals are also extremely valuable in shipboard and aircraft environments, when measuring instruments such as chronometers and compasses must be kept level with the horizon. Gimbals may also be used for aerospace navigation, as they can be set to provide a stable measurement from a specific reference point such as the earth or sun regardless as to their actual position in space. In this specific application, our gimbal will hold the cameras and instruments that instead the compasses will not be kept in horizontal position but will point to a fix position in the inertial 9 Section 2. Gyros, motors and gimbal geometry frame, or will perform a tracking of an object in that frame. Therefore the selection of the gimbal configuration is of special importance. Gimbals employed in aerospace navigation utilize Euler angles to orient an object such as a spacecraft. This work will also use Euler angles to develop the dynamics of the system. Euler angles are more intuitive, although not as robust as the quaternions, but can lead to gimbal locks and are not as efficient. A possible future work could be to develop a stabilization control based on quaterinions. However, assuming the real angles and angular velocities will not deviate much about the desired ones the behavior of both systems is rather similar. Figure 2.3 Gimbal 10 Section 3. Dynamics and Kinematics 3 Dynamics and Kinematics 3.1 Basic coordinate frame transformations For the 2-axis gimbal problem three different coordinate frames are used. The base or platform frame, the outer gimbal frame and the inner gimbal frame. All frames are related by transformation matrices, in this case the sum of a translational and a rotational matrix. Figure 3.1 shows the gimbal structure with the different coordinate frames axis depicted in different colors along with their angular relations. Figure 3.1 Different coordinate frames for a 2-axis gimbal 11 Section 3. Dynamics and Kinematics The relation between the platform frame and the outer gimbal frame is expressed by the sum of t wo ma t r i c e s .One ,Rθ PO, represents a rotational transformation and the other one, T PO, represents a translational transformation. For the sake of simplicity figure 3.1 does not depict the translation between these two frames although it does depict the translation between inner and outer gimbal frames. cos PO RPO sin PO 0 TPO sin PO cos PO 0 0 0 1 1 0 0 0 1 0 0 0 Tpoz (3.1) (3.2) Coordinates in Outer Frame RPO Tpo Coordinates in Platform Frame The same situation happens for the relations between outer and inner gimbal frame where now Rθ OI represents the rotational transformation and T OI represents the translational transformation. cos OI ROI 0 sin OI 0 sin OI 1 0 0 cos OI 1 0 0 TOI 0 1 0 0 0 Toiz (3.3) (3.4) Coordinates in Outer Frame ROI TOI Coordinates in Platform Frame Because our interest is focused on angular velocities, angular accelerations and angles, we can drop the Translational Matrix that has no influence in how the rate vector of one coordinate frame relates with the other. The outer frame rotates around the z-axis and the inner frame around the y-axis so the angles θ ndθ PO a OI will not stay fixed but will vary in time. The de r i v a t i v eofθ PO depends on the different velocities between the platform coordinate frame a n g ul a rv e l oc i t y( ωP)a ndt heou t e rc oo r di na t ef r a mea ng ul a rv e l oc i t y( ωO) around the z-axis. It is clear f r omhe r et ha tωO can not be solved explicitly without taking into account the forces that 12 Section 3. Dynamics and Kinematics produce the relative angular velocity between both frames of reference. The same goes for the i nne rc oo r di na t ef r a mea ng ul a rv e l oc i t y( ωI), but in this case between the outer coordinate frame a n g ul a rv e l o c i t y( ωO)a ndt hei nn e rc oo r di n a t ef r a mea n g ul a rv e l oc i t y( ωO) around the y-axis. . Asme nt i ona bov et h ed e r i v a t i v eo fθ PO represented by equations is PO Oz (t ) Pz (t ) (3.5) Creating the auxiliary matrix 0 PO 0 PO (3.6) We can get the relation between the platform frame and the outer gimbal frame as O (t ) RPO (t )P (t ) PO (3.7) Following an identical procedure OI Iy (t ) Oy (t ) (3.8) 0 OI OI 0 (3.9) And the relation between the outer frame and the inner gimbal frame is I (t ) ROI (t )O (t ) OI (3.10) Finally we need to calculate the angular acceleration rate of the inner frame. These angular acceleration rates will be used in the next section to develop the gimbal dynamics model. This model will be the basis for the rest of the work and will be used to perform the simulations and to analyze different controllers. Therefore, differentiating equation 3.10 we get I (t ) ROI (t ) O (t ) RAUX O (t ) OI OI 13 (3.11) Section 3. Dynamics and Kinematics Where RAUX sin OI 0 cos OI 0 cos OI 0 0 0 sin OI (3.12) 3.2 Eul e r ’ smoment equations The gimbal dynamics model can be derived from the torque relationships about the inner and outer gimbal body axes based on rigid body dynamics [1]. TheEul e r ’ smome nte qua t i onsa r e M h I h B h (3.13) Where M represents the applied moment and h is the angular momentum. The subscript I express a derivate in the inertial frame and the subscript B a derivate in the object frame. If the principal axes of inertia coincide with the coordinate frame, which is our case, performing the vector product give us three scalar equations M X I X X I Z IY Y Z M Y IY Y X Z I X I Z (3.14) M Z I Z Z X Y IY I X The following equations that will be used to develop the gimbal dynamics model follow closely the treatment given in [1] with major differences in the model of the friction. 3.3 Inner gimbal dynamics We will begin analyzing the dynamics of the inner gimbal. The sum of the kinematics torques about the inner gimbal is 14 Section 3. Dynamics and Kinematics M Ix t TOIx t TGGIx t M Iy t Tel t TGGIy t T fI t M Iz t TOIz t TGGIz t (3.15) With Tel : Elevation stabilization control torque. T fI : Friction torque about y-axis TGGIx , TGGIx , TGGIx : Gravity gradient torques about each gimbal axis. TOIx , TOIz : Torques exerted by outer gimbal on inner gimbal The elevation stabilization torque Tel - produced by the motor attached to the y-axis of the inner gimbal –consists of two parts. One has the main function of canceling the disturbances and therefore to nullify the angular velocity ωIy, this part will be called Tel, and the other has to control the inner gimbal in order to move up to the desired angle, this part will be called Tel. The gimbal-motor system has been built in a way to avoid or minimize all nonlinear of behavior of friction. It has proved to be a realistic assumption to consider the friction as the sum of a linear component, dependant on the relative angular velocity between the inner and outer gimbal y-axis and a nonlinear component called coulomb friction. The linear component of the friction is proportional to the viscous friction coefficient KIf. The coulomb friction has a constant value, Kcf, and its direction depends on the sign of OI . This constant can only be determined empirically. The outer gimbal will exert torques around the x and z-axis of the inner gimbal. This torque will be the necessary to produce the same angular displacement experienced by the outer gimbal due to the fixed relationship between these axes. One important external moment is the gravitational moment. An asymmetric body subject to a gravitational field will experience a torque tending to align the axis of least inertia with the field direction [4]. For the following development, we assume that the gimbal is at a distance R0 from 15 Section 3. Dynamics and Kinematics the Ea r t h ’ sc e nt e ro fma s s . The reference frame will be defined as follows: the origin of the reference frame moves with the center of mass (cm) of the inner gimbal. The zR-axis points towards the cm of the earth ( t he s ub s c r i p t“ R” s t a nds f orr e f e r e n c e ) . The xR-axis is perpendicular to the zR-axis in the direction of the unmanned plane’ sv e l o c i t y . The yR-axis is perpendicular to both, the zR and xR-axis. The aircraft’ sa x i sf r a mei sde f i n e dby xB, yB and zB where each axis coincides with the inertia axes ( t h es u bs c r i pts t a nds“ B”d e not e sbody). The Euler angles are defined as the rotational angles about the body axes as follows: φ,a bo ut the xB-axis; θ , about the yB-axis; and ψ, about the zB-axis, assuming an initial alignment between the reference and the body frame. The gravity gradient vector is defined as T G Gx Gy Gz (3.16) The force exerted on a mass element due to gravity is dm dF - 3 r r (3.17) Where r = R+ρis the di s t a nc ef r omt hee a r t h’ sc e n t e rofma s st ot hema s sdm. Since ρ <<R0, the moment about the center of mass of the body becomes dG dF = - dm r 3 r (3.18) where ρis the radius vector from the body’ scenter of mass to a generic mass element dm. Because ρ <<R0, 1/r3 can be approximated as 1 1 3R 3 1 2 3 r R0 R0 (3.19) Integration of equation 3.18 over the entire body of mass, together with Eq. 3.19, leads to 3 G 5 R R dm 2 R0 M After calculating the scalar and vector products we get the final results 16 (3.20) Section 3. Dynamics and Kinematics 3 Gx 3 I z I y sin 2 cos 2 2 R0 3 Gy 3 I z I x sin 2 cos 2 R0 (3.21) 3 Gz 3 I x I y sin 2 sin 2 R0 These are the gravity gradient moment components of G. The gravity moment vector G should be expressed in terms of the angles of the inner gimbal system of reference. This can be achieved measuring the body axes angular rates relative to the reference frame together with knowledge of the initial conditions of the Euler angles relative to the reference frame. As we will show later, the particularities of our case will make this unnecessary. If we combine equations 3.14 and 3.15 and replace the general inertia matrices and angular velocities by the inertia matrices and the angular velocities of the inner frame we get TOIx t I Ix Ix IyIz I Iz I Iy TGGIx t I Iy Iy IxIz I Ix I Iz Tel t TGGIy t T fI t (3.22) TOIz t I Iz Iz IxIy I Iy I Ix TGGIz t The first and third equation of 3.22 have no practical use since the gimbal rotation around the x and z-axis are will depend completely of the rotation of the outer gimbal. Therefore the torques will be the necessary ones to accomplish this. This leaves us with the second equation, which contains the derivative of one of the variables to be controlled (ωIy). Expanding the terms of the friction torques and the control torques we get I Iy Iy IzIx I Ix I Iz Telw t Tel t TGGIy t K vf OI t K cf sgn OI t (3.23) The goal is to control the elevation axis - inner gimbal y-axis - and the cross-elevation axis inner gimbal z-axis. Therefore, the variables to bec on t r ol l e da r eωIy a ndωIz. Because we do not 17 Section 3. Dynamics and Kinematics have direct control over the cross-elevation axis we have to do it indirectly through the azimuth axis - outer gimbal z-axis. Using the relations stated in equation 3.11 and 3.3 we get Ix cos OI Ox sin OI Oz Iz sin OI Ox cos OI Oz Ox cos POBx sin POBy (3.24) Oy sin POBx cos POBy The first three equations give us cos 2 OI Ox sin OI Iz sin 2 OI Ox Ox sin OI Iz Ix cos OI cos OI cos POBx sin POBy sin OI Iz cos OI (3.25) We also know that OI Iy (t ) Oy (t ) (3.26) Using the relations stated in equations 3.25 and 3.26 it is possible to represent equation 3.23 in terms of the controlled variables and the base disturbances. I Iy Iy K vf Iy K vf (cos POBy sin POBx ) Iz cos POBx sin POBy sin OI Iz I Ix I Iz / cos OI K cf sgn Iy cos POBy sin POBx Telw t Tel t TGGIy t (3.27) Equation 3.27 is useful to simulate and model the real system but for control purposes it is more practical to use the information of the angular velocities directly from outer gimbal axis instead of getting the data from the base. A second way to represent these dynamics is in terms of the controlled variables and the outer gimbal angular velocities as they will be seen by the controller. Later on it will be shown that the outer gimbal angular velocities will not be known exactly due to sensor dynamics and noise measurements. Representing equation 3.27 in terms of the outer gimbal angular velocities give us 18 Section 3. Dynamics and Kinematics I Iy Iy K vf Iy Iz Ox sin OI Iz I Ix I Iz / cos OI K vf (Oy ) K cf sgn Iy Oy t Tel t TGGIy t Telw (3.28) It can be seen that using the outer gimbal angular velocities not only simplifies the model but reduce the number of sensors used. 3.4 Outer gimbal dynamics Weus et h eEul e r ’ smome nte qua t i o nst og e tt her i g i dbodyt or quedy na mi c sf ort heou t e r gimbal. The total torque vector about the outer gimbal axis is M O M OT MI O (3.29) Expanding this equation we get the sum of the kinematics torques about each axis of the outer gimbal M Ox t TBOx t TGGOx t M Ix t O M Oy t TBOy t TGGOy t M Iy t O (3.30) M Oz t Taz t TGGOz t T fO t M Iz t O M Ix t , M t , M t are t het or que ’ sma t r i c e sof the inner gimbal referred to O Ix O Ix O the outer gimbal frame. As in the previous case we have Taz : Azimuth stabilization control torque. T fO : Friction torque about z-axis TGGOx , TGGOy , TGGOz : Gravity gradient torques about each outer gimbal axes. TBOx , TBOy : Torques exerted by the base on outer gimbal. In the same manner as before, the azimuth stabilization torque Taz - produced by the motor attached to the z-axis of the outer gimbal –consists of two parts: disturbance cancellation; and pointing. Because stabilization is required about the inner z-axis, called cross elevation axis, Taz is used to indirectly control t hea ng ul a rv e l o c i t yωIz. This fact poses a difficult problem due to the high nonlinearities that appear in the dynamics of the system. 19 Section 3. Dynamics and Kinematics As before, the friction can be divided in a viscous friction, proportional to BO and KOf, and the Coulomb friction, that has a constant magnitude equal to Kfc and whose direction depends on the sign of BO . The Outer and Inner torque vectors are defined as M O [ M Ox , M Oy , M Oz ] (3.31) M I [ M Ix , M Iy , M Iz ] Ap pl y i ngt heEu l e r ’ smome nte qu a t i ont oboth vectors M O h h I O O O I OO (3.32) ROI T M I ROI T h I I h ROI T I I I I I II (3.33) M I O Combining with equation 3.29 we can derive the rigid body torque dynamics for the outer gimbal body as M OT I O O O I OO ROI T I I I I I I I (3.34) From Equation 3.34 and the symmetry property that states that ROI 1 ROI T we can expand the first term of the right hand I O O = I O ROI T I I O ROI T RAUX O OI I O ROI T OI (3.35) With this result, equation 3.34 can be expressed as I O ROI T ROI T I I I + O I OO + ROI T I I II I O ROI RAUX O OI I O ROI OI M OT T T (3.36) As we are interested in the cross elevation axis dynamics, we shall consider only the third element of the vector shown in equation 3.36. Solving for the first term of the left hand we have 20 Section 3. Dynamics and Kinematics I Ox cos OI T T I O ROI ROI I I I 0 3 I Oz sin OI I Ix cos OI 0 I Ix sin OI 0 IOy 0 Ix I Ox sin OI 0 Iy IOz cos OI Iz 3 Ix I Iz sin OI 0 Iy I Iz cos OI Iz 3 0 I Iy 0 (3.37) I sin I cos Ix Iz Oz OI Oz OI I Ix sin OI Ix I Iz cos OI Iz cos OI I Iz I Oz Iz sin OI I Ix IOz Ix In this case []3 denotes the third element of the vector. As we can see below, the third element of the last term of the right hand is equal to zero. I Ox 0 0 cos OI T I O ROI OI 0 I Oy 0 0 3 0 I Oz sin OI 0 0 I Oy OI 0 0 3 0 0 sin OI 1 0 OI 0 cos OI 0 3 (3.38) After replacing the results obtained above for the cross elevation axis dynamics equation we get cos OI I Oz I Iz Iz = sin OI I Oz I Ix Ix O I OO + ROI T I I II 3 (3.39) + I O ROI T RAUX O OI M OT 3 3 From equation 3.11, the angular acceleration about the inner x-axis can be obtained. Substituting into 3.39, expanding the cross products terms and substituting the kinematics torques leads to 21 Section 3. Dynamics and Kinematics IT Iz = sin OI I Oz I Ix Ox I OzIx sin OI I IxOz OI cos OI OxOy I Oz I Ox sin OI IyIz I Iz + IyOx I Iy cos OI IxIy I Ix (3.40) cos OI Taz TGGz K Ovf OI KOcf sgn OI This equation could be expanded further to represent the whole dynamics only in terms of the controlled variables and the base disturbances. 3.5 Augmented inner gimbal dynamics –sensors dynamics All variables are measured by sensors which have their own dynamics plus noise added at the output. Therefore, the actual variables can not be known exactly. To deal with this, an augmented state-space representation of the system is done. From Eq equation 2.1 we know that the dynamics of the sensors are described as y 2500 TFSensor sensor 2 m s 70 s 2500 (3.41) Were the input m is the angular velocity to be measured. If we use a state-space representation x 0 1 1 x1 0 2500 70 2500 x2 x2 ysensor x1 1 0 x2 (3.42) Now we are ready to combine equations 3.42 and 3.28 to get a state-space representation of the inner gimbal dynamics including the sensor dynamics. This representation of the complete system will be used afterwards to develop the observers and controllers for the system. 22 Section 3. Dynamics and Kinematics x1 K vf / I Iy 0 0 x1 K cf / I Iy sgn x2 0 0 1 x 0 Oy 2 Iy 2500 0 70 2500 x3 x3 1/ I Iy 0 Telw Tel TGGIy K vf * Oy 0 (3.43) I Ix I Iz / I Iy 0 Iz Ox sin OI Iz / cos OI 0 ysensor x1 0 1 0 x2 x3 The gimbal used for this project has a special property: its axes of inertia are symmetrical. We can see that due to this symmetry the last term in the state equation will be canceled. Also due to this symmetry, the gravity gradient torque will disappear (see equation 3.21). Furthermore, this representation is useful for adding measurement noise at the output of the gyroscopes. Simplifying the state equation and adding the measurement noise we obtain x1 K vf / I Iy 0 0 x1 1/ I Iy x2 0 0 1 x2 0 Telw K vf * Oy 2500 70 2500 x 0 3 x3 K cf / I Iy sgn 0 Iy Oy 0 ysensor (3.44) x1 0 1 0 x2 e x3 Were e is the measurement noise of the sensor. This noise is assumed to be white noise with covariance Ecov. e(t ) N (0, Ecov ) 23 (3.45) Section 3. Dynamics and Kinematics It can be seen from the state-space equation that canceling the nonlinear term would leave as with a linear system and therefore able to apply a controller for LTI systems, e.g. PID controller. Two important obstacles are the fact that the parameter Kcf is unknown and need to be estimated in real time, and as stated before that we do not have the precise values of the variables ωIy and ωOy. It is clear the necessity to develop an observer to be able to apply a negative feedback to counteract the torque produced by this term. This observer will be developed in the following section. 24 Section 4. Observers and Controllers 4 Observers and controllers 4.1 Observer for the inner gimbal Besides the sensor dynamics mentioned before we consider here the common case of noisy sensor measurements. There are many sources of noise in such measurements. For example, each type of sensor has fundamental limitations related to the associated physical medium. In addition, some amount of random electrical noise is added to the signal via the sensor and the electrical circuits. Therefore analytical measurement models typically incorporate some notion of random measurement noise or uncertainty. When the variable being measured is planned to be used as an input for a controller an accurate estimation of it is of utmost importance. If we consider only the inner gimbal dynamics, there are two variables that are necessary to be estimated: the inner gimbal angular velocity of the y-axis - ωIy -, and the Coulomb friction torque (CFT), Kcf · sign( OI ). The first variable will be used to design the controller and the second is needed for canceling the torque is exerting over the gimbal and also for estimating the angular velocity. The main challenge is to estimate the nonlinear Coulomb friction. Two different, although similar, approaches will be used for this purpose. 4.1.1 First order approximation A simple method to estimate the Coulomb friction is described in [8]. A first-order linear stochastic differential equation is incorporated to the system of equations to estimate this friction. This linear equation in no way represents an accurate model for nonlinear Coulomb friction; however, it is of a form that is compatible with the Kalman filter equations. As it will be demonstrated, this stochastic differential equation characterizes a slowly changing random 25 Section 4. Observers and Controllers process that under the right conditions accurately identifies the effects of the actual nonlinear friction on the system. The algorithm was tested using numerical simulation techniques As shown in Fig. 4.1, the disturbance torque, associated with the representation of the Coulomb friction, is either plus or minus depending on the sign of OI . Figure 4.1 Torque couse by Coulomb friction If we assume that the switching of the friction from plus to minus is at a relatively low frequency compared to the sampling frequency, then it is possible to approximate the model for the friction as being exponentially correlated with time constant that is significantly greater than the inverse of the sample rate [9], [10], i.e., dK cf 1 K cf w dt (4.1) where w is white noise with zero mean and covariance Wcov. w(t ) N (0, Wcov ) (4.2) This simple first-order equation allows the use of the Kalman filter equations for estimating the Coulomb friction. A sensitivity study revealed that there was little change in performance once τbecame much greater than the inverse of the sampling rate. When the sensor dynamics and noise measurement are not considered, this approximation works extremely well. Although the equations for this first simple case are not shown, the results after simulating this model can be seen in figure 4.2. 26 Section 4. Observers and Controllers Figure 4.2 Real vs. estimated friction torque Next figure amplifies the plot in the transition point to visualize more clearly the convergence time and the steady-state error. Figure 4.3 Zoom over transition area Sampling period =0.001[sec] Di s t u r banc eωOy = 0.8 sin(2.28t)[rad/sec] Coulomb friction coefficient Kcf = 0.1 Viscous friction coefficient Kif = 0.56 Inertia moment IIy = 0.325 27 Section 4. Observers and Controllers The figures show that the convergence time is almost negligible and the steady-state error, although not zero, is small enough to have any significant influence. The estimation convergence rate and steady-state error are greatly deteriorated when the same approximation is used for the complete system, which means including the sensor dynamics and noise measurement. The continuous-space representation of the complete system using the first-order approximation for the Coulomb friction can be expressed as x1 K vf / I Iy x2 0 0 x3 2 r x4 1/ I Iy 0 1/ 0 0 0 0 2 r 0 x1 1/ I Iy 0 x2 0 1 0 Telw w x3 0 1 0 2 r x4 0 0 (4.3) x1 x2 0 0 1 0 e x3 x4 ysensor In order to discretize the system it was assumed Telw, w and e being piece-wise constant during a sample period, i.e. zero-order-hold for the input and disturbances. The new discrete system can be represented as X K 1 Ad X K Bd Telwk Gd wk ysensor k Cd X K Dd ek (4.4) where Ad e AT 1 sI A 1 t T T A Bd e dB A1 Ad I B if A nonsingular 0 Cd C (4.5) Dd D The covariance of the measurement and process noise is also influenced due to the discretization process, therefore 28 Section 4. Observers and Controllers wk N (0, Wd cov ) ek N (0, Ed cov ) (4.6) where T Wd cov e AQe Ad 0 (4.7) Ed cov E and T is the sample time. Within the significant toolbox of mathematical tools that can be used for stochastic estimation from noisy sensor measurements, one of the most well-known and often-used tools is what is known as the Kalman filter. The Kalman filter is essentially a set of mathematical equations that implement a predictor-corrector type estimator that is optimal in the sense that it minimizes the estimated error covariance, when some presumed conditions are met. Although not all the required conditions are met, because the system is only an approximation to the real nonlinear system, the Kalman filter has proven to be flexibly enough to be implemented. The Kalman filter estimates a process by using a form of feedback control: the filter estimates the process state at some time and then obtains feedback in the form of (noisy) measurements. As such, the equations for the Kalman filter fall into two groups: time update equations and measurement update equations. The time update equations are responsible for projecting forward (in time) the current state and error covariance estimates to obtain the a priori estimates for the next time step. The measurement update equations are responsible for the feedback—i.e. for incorporating a new measurement into the a priori estimate to obtain an improved a posteriori estimate. A whole demonstration of the Kalman filter equations can be found in [11]. A summary of the main equations is shown below The estimated error covariance to be minimized is defined as T Pk xk xk xk xk 29 (4.8) Section 4. Observers and Controllers The following procedure uses the system represented in equation 4.4. The first step is the time update calculations. The equations used are x k k 1 Ad x k 1 k 1 Bd uk (4.9) Pk k 1 Ad Pk 1 k 1 AdT GdWcovGdT For the measurement update we need first to compute the Kalman gain and then to correct the estimated states and to compute the new covariance matrix. K k Pk k 1CdT Cd Pk k 1CdT Ecov 1 ek k 1 yk y k k 1 (4.10) x k k x k k 1 K k ek k 1 Pk k Pk k 1 Pk k 1CdT Cd Pk k 1CdT Ecov C P 1 d k k 1 It is quite common for most applications to use the steady-state Kalman gain instead of the timevarying gain. All the following simulations were run using the steady-state Kalman gain. For the next simulation, a Kalman filter was also used to estimate the real value of the distur b a nc eωOy. In this case it was assumed a sinusoidal disturbance of a specific frequency (6.2832 rad/sec or equivalently 1Hz) and unknown amplitude and phase. Although this assumption about the disturbance characteristics could be seen as somewhat arbitrary it is quite common the case when information about the disturbance is known in advance. It is also possible to use a nonlinear estimator (e.g Extended Kalman filter) for a disturbance with uncertain frequency. For disturbances with frequencies much smaller than the s e n s or ’ s bandwidth simulations showed that it was unnecessary to use observers. Because the sensor transfer function is essentially a low-pass filter, low frequencies disturbances are almost not affected by it, pl u s ,t heKa l ma nf i l t e rf orωOy proved to be almost insensitive to its measurement noise. A complete diagram of the system and the estimator can be seen in Figure 4.4 30 Section 4. Observers and Controllers Figure 4.4 System– observer diagram The next figure shows a comparison between the actual Coulomb friction torque (CFT) and the estimated with the Kalman filter. It can be seen that although the measurement noise for the sensor was extremely low ( t hes e ns o rf o rωIy); the estimation error is considerably high. The convergence rate, although decreased, is still quite high. For all the simulations in this subsection the measurement noise variance for t hes e ns orofωOy was equal to 1-3. Figure 4.5 Real vs. estimated CFT Where measurement noise: Ecov = 1-10 Figure 4.6 displays an amplification of figure 4.5 were the convergence rate and steady-state error can be seen with more detail. 31 Section 4. Observers and Controllers Figure 4.6 Amplification of real vs. estimated CFT Simulations showed that the main reason for the decrease in the estimation performance is the measurement noise. Applying the Kalman filter to the complete system without measurement noise proved to be as efficient as for the simplified system. Furthermore, a minor increase in measurement noise greatly magnifies the estimation error. Figure 4.5 shows the increase in the estimation error when the covariance of ek, Ecov, was equal to 9.9-9 Figure 4.7 Real vs. estimated CFT for high MN 32 Section 4. Observers and Controllers In the next figure we can visualize that although the convergence time is practically the same as in the previous simulation the error in the estimation is too high to be use for canceling the Coulomb friction torque. Figure 4.8 Amplification of real vs. estimated CFT for high MN The measurement noise goes straight through the feedback gain K used in the Kalman filter. This gain has a high value, needed to be able to estimate a nonlinear behavior using a linear approximation. Therefore any noise will be highly magnified degrading the filter performance. One way to avoid this is to diminish the Kalman gain but what would cause a reduction of the convergence time. Because every time there is a switch in the sign the of OI , and consequently in the torque produced by the Coulomb friction, the estimation process start all over again this option has no practical use except for very low frequency disturbances. Even under these conditions decreasing the Kalman gain proved to be unadvisable because slowing down the convergence rate would cause a resonant response with a considerable overshoot. One option to improve the performance of the algorithm is to place a digital low pass filter for the measured variable before calculating the error signal yk y k k 1 . Simulations showed that placing a low-pass filter after the sensor succeeded in reducing the estimation error but the trade-off is the reduction in bandwidth. For low cut-frequencies a noticeably decrease in the convergence time was observed and also a phase-shift in the 33 Section 4. Observers and Controllers e s t i ma t i o no fωIy. If the cut-off frequencies were too high, almost no reduction of the estimation error of the CFT was achieved. However, one of the main advantages of the algorithm is the fact that even if high estimation errors are obtained for the Coulomb friction torque, the estimation of the inner angular velocity is rather accurate. This can be seen in fig ur e4. 6wh e r et hea c t u a la nde s t i ma t e dωIy are plotted. Figure 4.9 Re alv se s t i mat e dωIy The simulation was run with the same noise conditions, and system parameters, of the first simulation showed in this section. After zooming for a better visualization of the differences between both signals we have Figure 4.10 Zoom over real vs. e s t i mat e dωIy 34 Section 4. Observers and Controllers Although there is some noise corrupting the estimated ωIy, this is particularly smaller than the noise over the estimated CFT. As a result the filter provides us with an estimated angular velocity accurate enough to be used for the control algorithm. This accuracy tends to degrade when the Coulomb coefficient has higher values and therefore a major influence in the overall system dynamics. 4.1.2 Extended Kalman filter To overcome the main drawback of the previous algorithm an alternative method is proposed. Instead of using a linear approximation compatible with the traditional Kalman filter the extended Kalman filter for the nonlinear system will be developed. A Kalman filter that linearizes about the current mean and covariance is referred to as an extended Kalman filter or EKF. In something akin to a Taylor series, we can linearize the estimation around the current estimate using the partial derivatives of the process and measurement functions to compute estimates even in the face of non-linear relationships. The nonlinear system will be represented as X k f X k 1 , uk 1 , wk 1 yk c X k , ek (4.11) As same as with the discrete Kalman filter, the EKF algorithm is divided into two steps: time update and measurement update. The time update equation, for estimation of Xk based in the estimation of Xk-1, comes straight from equation 4.11. Consequently X k k 1 f X k 1 k 1 , uk 1 , 0 (4.12) Before presenting the whole set of equations we need to define the Jacobian matrices of the partial derivatives of f and c with respect to X, w, and e. These matrices will be used later in a similar way to how matrices Ad and Cd were used in the discrete Kalman filter. fi Ak i , j X k 1 k 1 , uk 1 , 0 X j 35 (4.13) Section 4. Observers and Controllers fi Wk i , j X k 1 k 1 , uk 1 , 0 w j (4.14) ci Ck i , j X k k 1 , 0 X j (4.15) ci Ek i , j X k k 1 , 0 X j (4.16) We are now ready to define properly the time update equations X k k 1 f X k 1 k 1 , uk 1 , 0 (4.18) Pk k 1 Ak Pk 1 k 1 A WkWcovWkT T k and the measurement update equations K k Pk k 1CdT Ck Pk k 1CkT Ek Ecov EkT 1 x k k x k k 1 K k yk c( x k k 1 , uk , 0) Pk k Pk k 1 Pk k 1CdT Cd Pk k 1CdT Ek Ecov EkT (4.19) C P 1 d k k 1 The system in 3.44 was discretized considering the unknown parameter Kc as a variable of the system and using the approximation X X k 1 X k T (4.20) where T is the sampling period For T small enough, this discretization method is almost equivalent to the zero-order-hold method. The Jacobian matrices of this discretized system are 36 Section 4. Observers and Controllers 1 TK vf / I Iy Tsign( x1k 1 k 1 ) / I Iy 0 1 Ak 0 0 2 0 T r g11 0 0 0 0 g 0 0 22 Wk 0 0 g33 0 0 0 g 44 0 0 0 1 T T 2 r 1 T r2 0 0 (4.21) Ck 0 0 1 0 Ek 1 The parameters g11, g22, g33, g44 will be used in the design process to tune how the noise w affect each of the states of the system. Using the matrices from equation 4.21 and the equations 4.184.19, an S-function was programmed in MATLAB to perform the Extended Kalman Filter algorithm. The next simulation was the result of applying the EKF with identical conditions as in the first simulation performed in the previous subsection. Because we are not using any steady-gain, the filter is a time-varying system, thus the computational load is much higher than when using the steady Kalman gain. The results are plotted in figure 4.11 Figure 4.11 Real vs. Estimated CFT using EKF Amplifying the image for one of the periods shows clearly the enormous improvement in the estimation of the CFT over the previous algorithm, with a relative error less than 0.05%. 37 Section 4. Observers and Controllers Figure 4.12 Zoom over real vs. estimated CFT using EKF It is very important to point out that the right choice of the different design-parameters of the EKF, i.e. matrix W, initial covariance matrix P0 and noise covariance matrices Edcov and Wdcov, is fundamental for the correct functioning of the filter and usually can only be found empirically after several tuning of each of the parameters mentioned. Even after increasing the measurement noise up to the same level as in the second simulation of the previous subsection the relative estimation error was kept below 0.1%, with some noticeable increase in the estimation noise. A magnification for this case is shown in figure 4.13 Figure 4.13 Zoom over real vs. estimated CFT using EKF for high MN 38 Section 4. Observers and Controllers Finally a third simulation was performed with a MN variance six orders higher than before (Ecov = 0.01). Even under these extremely noisy measurements, the estimator was able to keep the error in acceptable margins (less than 19%). Figure 4.14 Zoom over real vs. estimated CFT using EKF for extremely high MN For a better idea of what kind of noisy measurement we are referring, a plot with the real angular velocity and the measured by the gyroscope are shown in figure 4.15. Not only the angular velocity measured is extremely noisy but there is also a substantial phase-shift between both signals. Figure 4.15 Sensor output vs. ωIy The reason for this exceedingly performance of the extended Kalman filter resides in three main factors that interact with each other: 39 Section 4. Observers and Controllers There is no approximation done, so the complete system dynamics are used to estimate the variables. Because the nonlinear behavior of the CFT is taken into account, the convergence process does not restart every time there is a switch in the direction of the CFT (or equivalently in the sign of OI ). In the previous algorithm the high Kalman gain caused an extreme amplification of the MN, in this case the Kalman feedback gain is considerably smaller than before thus this magnification does not occur. The drawbacks of using the EKF are the slower convergence rate at the beginning of process and the high computational cost of implementing a time-varying observer. The estimation of the angular v e l o c i t yωIy also exceeds the performance achieved with the 1order approximation. For high values of the parameter Kcf this difference is even more considerable. Ag r a p h i cp l o t t i ngt hee s t i ma t e da n da c t u a lωIy is presented in figure 4.16. Figure 4.16 Re alv s .e s t i ma t e dωIy Zooming figure 4.16 shows the smoother behavior of the estimated signal compared with the 1order approximation and the extremely small misalignment with the r e a lωIy. 40 Section 4. Observers and Controllers Figure 4.17 Zoom over real vs. estimated CFT ωIy 4.2 Controller for the inner gimbal The controller is in charge of driving the necessary signal to the motor to control and cancel the angular velocity around the inner gimbal y-axis and to allow the target pointing and tracking. The motor dynamics should also be included in the controller equations, but because the relation power-torque of the motor is a static equation a complete analysis can be done without including this relation until the implementations phase. The control algorithm for disturbance rejection and tracking is split in two parts. Ck K p Iy Iydesired CFT where Iydesired is the desired angular velocity necessary for target tracking. The second term of the right-hand side is for canceling the nonlinear CFT. If a perfect cancellation is achieved, the real system (without including sensor dynamics) will became equivalent to a first order LTI system, hence a P controller should be suitable for the controller algorithm. KP represents the proportional gain. Although theoretically the system should remain stable even for extremely high values of KP, due to uncertainties in the system, unavailability of the actual variables ωIy and Kc, the necessity to implement a discretization that will limit the 41 Section 4. Observers and Controllers bandwidth, and other factors that will be described in the next section, increasing too much the value of Kp tends to be unstabilize the system. The impossibility to use the real variables imposes a slight change in the controller approach, forcing instead the use of the estimated variables. Thus Ck K p Iy Iydesired CKF (4.23) Simulations using three different approaches were performed. In the first case the controller includes only the proportional term and the output of the gyroscope is taken as the estimation of ωIy. The second approach includes the observer (in this case the Extended Kalman Filter) used t oe s t i ma t e dωIy and again only the proportional term. Finally, the third approach included the observer us e dt oe s t i ma t et hebot h,ωIy and CFT, the proportional term and the CFT-cancellation term. To obtain more realistic results, the observer used in subsection 4.1 to estimate the di s t ur b a nc eωOy was neglected and no assumptions were made about the disturbance frequency. Figure 4.17 a) displays the results of simulating the first controller. For a better idea of the effect of the controller in rejecting the disturbance the first five seconds of the simulations were run without any control being applied while after that the controller was activated. Figure 4.18 b) zooms the response after the controller is applied. Figure 4.18 a) Angular velocity, before and after control signal is applied 42 Section 4. Observers and Controllers Figure 4.18 b) Zoom over angular velocity, before and after control signal is applied Measurement noise standard deviation =1-4 Proportional gain Kp = 310 Disturbance ωOy = 0.8 sin(2πt)[rad/sec] Sampling period = 0.001 [sec] or equivalently 1000Hz Although a significant disturbance rejection was obtained, the cancellation attained can not be considered exceptional. It is important to take into consideration that the disturbance used for the simulation was rather high, with peaks up to 46 [degrees/sec] and that the measurement noise was four orders higher than the one used in the second simulation in subsection 4.1. For very low frequency disturbances and small MN this method come close to the behavior of the second method shown next. Simulation results for the second controller are plotted in figure 4.19 Figure 4.19 Angular velocity before and after control signal is applied (2º controller) 43 Section 4. Observers and Controllers Measurement noise standard deviation =1-4 Proportional gain Kp = 1500 Disturbance ωOy = 0.8 sin(2πt)[rad/sec] It is fairly evident the great improvement achieved by the second controller contrasted with the first. The variance reduction was above 92%.A more detailed perception of the cancellation achieved is seen in figure 4.20. Figure 4.20 Zoom over ωIyusing 2º controller It is interesting to notice the peaks observed in the angular velocity after the controller is applied. These peaks take place at the instants when there is a change in the sign of OI and hence in the direction of the CFT. For a consistent comparison, the same magnification was done after simulating the third controller. Figure 4.21 displays the results Figure 4.21 Zo omov e rωIy using 3º controller 44 Section 4. Observers and Controllers Clearly, a noteworthy improvement was achieved for the third controller (about a 26% smaller variance over the previous one). Simulations also showed that when the proportional gain Kp is reduced, the differences between both controllers are even more noticeable. We can see that although not crucial, to include the term for canceling the CFT was important for improving the overall performance. It important to remark that even when CFT is canceled the peaks mentioned above still appear in the response. The explanation is simple: even though the estimation of the parameter K cf stays constant after the convergence period, the estimation in the sign of OI depends on the estimation of ωIy a ndωOy. Therefore, the phase-misalignment and estimation error of these variables cause that for a short period of time, when the sign of OI changes, the control torque that is supposed to be canceling the CFT is being added to it. 4.2.1 Constraints for the controller There are four main constraints for the controller when only the inner gimbal is taken into account. Although all of them are related to each other it is useful to try to analyze the role of each one in the overall performance. Quantization: Different tests were made simulating the A/D converter quantization process and studying its effect on the overall performance. It is important to notice that for higher values of K p stronger are the requirements on the quantization step size for the A/D converter. There is an easy explanation for this: for high values of Kp,t h ea ng ul a rv e l oc i t y ωIy became very small, consequently, unless the quantization levels are small enough a lot of information is lost in the process. To analyze properly the effect of the quantization we need to do it jointly with the proportional gain because for different values of Kp we will obtain different values of suitable quantization levels. Tests exhibited quite an interesting phenomenon. Even for a relatively big quantization step size, using to convert the continuous signal coming from the sensor that is measuring ωIy to a 45 Section 4. Observers and Controllers discrete one, the controller performance was barely influenced. This important property can be visualized clearly in the next example. As before, the proportional gain Kp used in the controller algorithm was equal to 1500 and the MN standard deviation 1-4. Figure 4.22 shows the result of simulating the same controller under the same conditions with the only difference that in the first case no quantization was made for the measured signal used by the controller and for the second case an A/D converter was placed after the sensors with a quantization step size of 1.3-3 (assuming a range from 0.65 to -0.65 [rad/sec] this is equivalent to a 10 bits A/D converter) Figure 4.22 System response with and without quantization Disturbance ωOy = 0.8 sin(2πt)[rad/sec] Even for this large quantization step there is only a slight degradation in the controller performance when including the A/D converter. For a better visualization an observer was pl a c e da f t e rt hes e ns orme a s ur i ngωOy so the peak sobs e r v e dbe f or edon ’ ta pp e a r( t hepe r i o d where OI and its estimation does not coincide is too small to influence the response). The results di dn’ td i f f e rwh e nt h eobs e r v e rwa sn oti nc l ude d . The extreme quantization of the measur e dωIy is manifest in figure 4.23. The figure plots the out pu toft h eA/ Dc onv e r t e rp l a c e da f t e rt heg y r os c opeus e dt os e ns o rωIy. 46 Section 4. Observers and Controllers Figure 4.23 Sensor output after quantization After amplifying figure 4.23, is evident that the measured angular velocity ωIy, after the transient period when the observer is still converging, only takes three different values, -1.3e-3, zero and, 1.3e-3. For smaller values of Kp even bigger quantization step sizes can be used with the same results. Figure 4.24 Zoom over sensor output after quantization Even for this extreme distortion caused by the converter the controller capacity to cancel disturbances is barely affected, as it was shown in figure 4.22. This is a great advantage because it allows us to be loose in the requirements of the converter, and consequently in the microprocessor in charge of dealing with the calculations. 47 Section 4. Observers and Controllers Saturation : Another practical limitation when implementing the controller is the limitation in the torque that the motor is able to induce. If the control signal exceeds certain maximum levels the motor will enter in saturation condition. Because of the difficulties in predicting the response of the system when the motor torque is continuously entering in saturation mode, it is important to analyze the system behavior when this situation occurs. For the control law proposed before, the parameter Kp will have a direct effect on the control signal magnitude. For this reason, if the condition of no saturation is imposed, the maximum allow levels will determine a practical limitation for the proportional gain Kp. For a better understating the of the consequences of reaching saturation levels a simulation was done, where the maximum torque that can be produced by the motor is assumed to be 11% below the maximum theoretical torque that should be produced according to the control signal. Figure 4.25 plots both motor torques driven by the control signal, with and without saturation restrictions. Figure 4.25 Control signal with and without saturation From the figure above is possible to visualize the small difference between both signals. Next two figures will show how, even for these slight differences in the torque produced by the motor, the performance discrepancies are exceptionally large. In figure 4.26, we can see a 48 Section 4. Observers and Controllers magnification of the steady-state response of ωIy when the torque, the motor is able to exert, has no limitation. Figure 4.26 Zoom over angular velocity ωIy without saturation For a consistent comparison, the same magnification is done for figure 4.27 where the motor maximum torque is limited by the saturation levels showed in figure 4.25. Figure 4.27 Zo omov e ra ng ul arv e l oc i t yωIy with saturation The peak value of ωIy for the controller with saturation limitation was 21 times higher than when no limits for the torque the motor can produced was imposed and the variance was 275 times higher. It proved to be critical for the control signal to not reach the saturation levels. This requirement will have to be taken into account when designing the controller algorithm. Unless 49 Section 4. Observers and Controllers a severe degradation in the performance is consider acceptable for the application, in case of saturation condition a change of the motor or of the control law will be required. Sampling period: The sampling period used for the observer and controller algorithms are of utmost importance in defining the system behavior. The two main reasons are: 1. It defines the computational cost of the algorithm 2. It limits the achievable bandwidth of the system One of the main drawbacks of using a high frequency sampling is that a faster microcontroller, able to perform all the necessary calculations, will be required for the implementation. The influence of the sampling period in the system bandwidth will determine another practical limitation for the proportional gain, hence for the disturbance cancellation effectiveness. The increase in bandwidth will allow not only a more effective disturbance cancellation but also will make it possible to cancel a broader range of disturbance frequencies. The sampling period (SP) is not the only limitation on the range of disturbance frequencies feasible to be cancelled; an ot h e ri mpor t a ntl i mi t a t i o ni st heba n dwi d t hoft h es e ns o r ’ st r a ns f e rf u nc t i on.Ont heot he r hand for very low sampling frequencies the dynamics of the sensor can not be modeled, therefore the EKF will not include its dynamics, adding another restriction when using low frequency sampling. Numerous tests were made to analyze the computational cost, cancellation rates and range of disturbance frequencies possible of being rejected, for different sampling periods. A special focus was set on the behavior when using a 50Hz sampling frequency, since this is the projected sampling frequency (SF) to be implemented in the real application. This frequency is too low to include any sensor dynamics in the observer algorithm. For a 50Hz SF and a 50Hz sensor bandwidth, disturbance frequencies up to 2.5Hz were viable to be cancelled efficiently. To study the major influence of sensor bandwidth on the controller capability to broad the frequency range for disturbance rejection several simulations using different sensor transfer functions were 50 Section 4. Observers and Controllers performed. As an example, when the bandwidth of the sensor was increase to 120Hz the 2.5Hz limit for effective disturbance cancellation was increased to 10Hz, four times higher than before. The optimal value of KP for a sampling frequency of 50Hz was equal to 80. Above this value the system performance started to degrade until becoming completely unstable for values of KP above 105. Next figure shows the results of applying a 50Hz sampling, the optimal KP and a sinusoidal disturbance of with frequency of 2Hz and peak amplitude of 8 degrees/sec. The EKF us e dd i dn ’ ti nc l udet hes e n s ordy na mi c s . Figure 4.28 Angular velocity ωIy for 50Hz sampling Disturbance ωOy = 0.13 sin(4πt)[rad/sec] Proportional gain Kp = 80 The results show that although the convergence time is relatively slow a considerable attenuation was achieved. The steady-state standard variance obtained was 7.32-7. Increasing the sampling frequency to 300Hz moves the optimal KP to 420. The results after running the simulation for the same disturbance are shown in figure 4.29. Even for this sampling frequency it was not convenient to include the sensor dynamics in the observer. 51 Section 4. Observers and Controllers Figure 4.29 An gul arv e l oc i t yωIy for 300Hz sampling Disturbance ωOy = 0.13 sin(4πt)[rad/sec] Proportional gain Kp = 420 It can be seen that the convergence time was drastically reduced and the steady-state variance obtained was 2.93-7, 2.5 times smaller the in the previous case. Finally the results after simulating the system with a sampling frequency of 1 kHz are displayed in figure 4.30. All the sensor dynamics were included in the observer algorithm. The convergence rate was not influenced but the steady-state variance reached was 5.9-8, almost 5 times better than before. The optimal Kp was found to be equal to1500. Figure 4.30 An gul arv e l oc i t yωIy for 1000Hz sampling Disturbance ωOy = 0.13 sin(4πt)[rad/sec] Proportional gain Kp = 1500 52 Section 4. Observers and Controllers Estimation error: Because the controller algorithm includes the estimated variable, an error in the estimation will affect directly the effectiveness of the controller. The estimation error is directly related to the noise measurement; therefore this will have a direct impact in the controller design process limiting the upper limit of the proportional gain. Fortunately simulations showed that MN of the sensors has small influence over the whole system. The MN of the sensor measuring ωIy was almost completely suppressed by the used of the EKF. Now,be c a u s et h ee s t i ma t i onofωOy is only used by the controller when estimating the sign of OI the MN has only influence when facing a change in the sign, thus for a very short period of time. This influence proved to be less significant than the one produced by the phaseshift caused by the sensor dynamics. Consequently unless the MN of the sensors is particularly high, the degradation over the performance has not much consequence. 4.2.2 Alternative controller for the inner gimbal The main goal underneath disturbance rejection is to maintain the aim of the different instruments, mounted on the gimbal, fixed at the target. This means to keep the angle over the inner axes (in this case only the y-axis is considered while in the complete system both y and zaxis need to be considered) fix with respect to the inertial frame. This angle will be identified as θ Iy; what is consistent with the fact that is time behavior is the integral of the angular velocity ωIy.Cl e a r l y ,a t t e nua t i ngωIy will reduce the angle variance but this will not assure minimization of this value nor guarantee that it will have zero mean. Moreover, all simulations using the controller developed in section 4.2.1 showed a steady-state shift in the angle response. This shift can be visualized in figure 4.31. 53 Section 4. Observers and Controllers Figure 4.31 Angle θ Iy for 50Hz sampling disturbance : Oy 0.13sin 2 pi / 9 proportional gain : K P 80 These results prompted the idea of augment the observer-s y s t e mt oi nc l ud et h ea ng l eθ Iy. If an estimation of t hea ng l eθ Iy could be obtained many options for the controller method would be available. Thedi f f i c u l t yi ne s t i ma t i ngθ Iy lays in the fact that to simply integrate Iy is not acceptable. Initial errors in Iy and posterior misalignments between Iy and Iy would make t hee s t i ma t e dθ image Iy completely useless. To overcome this obstacle we will make use of the “ processor sensor”used by the system for target recognition and therefore angle measurement. Although this sensor work at a much lower frequency compared with the gyroscopes the new EKF will use the information provided by this sensor in every update to correct the current angle estimation. Simulations showed that even for angle update frequencies as low as 1Hz, or even below, the new hybrid-sampled EKF work fairly well. A complete block diagram of this system is represented in figure 4.32. 54 Section 4. Observers and Controllers Figure 4.32 Complete system diagram for inner gimbal stabilization The Angle tracker processor bl oc kt ha ta p p e a r si nf i g ur e4. 32r e p r e s e nt st he“ image processor s e ns o r ” mentioned above. Using this augmented system, a discrete linear-quadratic (LQ) controller was designed. Because the system is to be operated in continuous mode, the stationary feedback gain L was utilized. The LQ controller was design to minimize the following cost function J X kT QX k Ruk (4.24) k 0 The right selection of the matrices Q and R are crucial aspects of the controller design process. The matrix Q will define the weight of the different states of the model while the matrix R represent the cost of the control signal and will be used to avoid saturation conditions. The control law minimizing the cost function stated in equation 4.24 come from solving the following discrete Riccati algebraic equation S Q AT S SB R BT SB BT S A 1 (4.25) The feedback control law will be u Lx (4.26) where 1 L R BT S R R B SB T 55 (4.27) Section 4. Observers and Controllers To show the effectiveness of this new controller, a simulation under the same conditions as the ones described for figure 4.31 was run. Then angle response can be seen in figure 4.33. Figure 4.33 Angle θ Iy for 50Hz sampling and LQ controller Controller feedback gain : L 80 500 We can see that not only the variance was significantly reduced but also the shift observed in the previous case was completely cancelled. The variance when using the discrete LQ controller was 53 smaller than when using the P controller (9.8-7 vs. 1.84-8). In addition, not only the angle response was dramatically improved but also the angular velocity response exceeded the cancellation rates achieved with the previous controller. To illustrate this, next two figures show the time response of Iy for both controllers. Figure 4.34 An gul arv e l oc i t yωIy for proportional controller 56 Section 4. Observers and Controllers Figure 4.35 An gul arv e l oc i t yωIy for LQ controller For the angular velocity the variance of the discrete LQ controller was 18 times smaller than for the P controller (4.46-7 vs. 2.49-8). I ti si nt e r e s t i ngt onot et h epe a k sobs e r v e di nωIy at the beginning of the simulation. These pe a k sc o i nc i dee xa c t l ywi t ht het i me si ns t a n t swh e nt h eda t af r om t he“ image processor sensor” is being used to correct the angle estimation. A more clear visualization is shown in figure 4.36. Figure 4.36 Zoom over ang ul arv e l oc i t yωIy for LQ controller After a number of corrections these peaks start to vanish. In figure 4.37, plotting the real and estimated angle for the first seconds of the simulations, it can be visualized how the estimated error convergence almost to zero after approximately six corrections. The image processing sensor was assumed to be working at a frequency of 1Hz. 57 Section 4. Observers and Controllers Figure 4.37 Zoom over real angle vs. estimated angle One of the limitations for this LQ-controller comes from the hybrid-sampled Extended Kalman filter. The use of this modified algorithm for the observer reduced the range of disturbances frequencies feasible to be canceled to frequencies up to 2Hz. For frequencies above this value t hee s t i ma t i one r r oro fθ Iy, generated by the EKF were too large to be used by the controller. Again, it is important to remark that increasing the sensor bandwidth will increase this frequency limit. The degradation of the performance when approaching this frequency is shown in the next four figures. The first two compare t het i mer e s pons eo fθ Iy for both controllers when a disturbance of 1.2Hz and the same amplitude as before is exerted on the system. Figure 4.38 Angle θ Iy for 50Hz sampling and proportional controller 58 Section 4. Observers and Controllers Figure 4.39 Angle θ Iy for 50Hz sampling and LQ controller Even when the LQ-controller is not as effective as for lower disturbance frequencies is still surpass the performance of the proportional controller regarding variance and steady-state shift. The steady-state variance for the LQ-controller was still 6 times smaller compared to the proportional controller (7.26-9 vs. 1.2-9) and the steady-state shift was almost completely annulled (more than 3000 times smaller). Finally, last two figures show the time response of Iy for both controllers. Figure 4.40 An gul arv e l oc i t yωIy for 50Hz sampling and proportional controller 59 Section 4. Observers and Controllers Figure 4.41 An gul arv e l oc i t yωIy for 50Hz sampling and LQ controller Controller feedback gain: L=[80 1400] As with the angle response, the efficacy of the LQ-controller in canceling the angular velocity Iy exceeds the one of the proportional controller; achieving a steady-state variance reduction greater than 6 (4.12-7 vs. 6.7-8). Another fundamental advantage of this new controller is the simplicity to apply angle tracking. Si mpl er e p l a c i ngt h es t a t eθ desired will cause the system to move Iy Iy of the control law by the angle to the desired position. Figure 4.39 plot the results of applying the tracking control with a desired angle of seven degrees with exactly the same parameters used in the previous simulation. Figure 4.42 Angle θ Iy for LQ controller in tracking mode 60 Section 4. Observers and Controllers The convergence time although not extremely rapid is quite fast for most practical purposes and the steady-state response shows the same characteristics as before. Finally it is important to remark that due to the static relation between the control signal and motor torque (see equation 2.2), the signal calculated with the control law has to be multiplied by an adjustment factor as expressed in the following equation when implementing the controller. 2 controller output Power signal driving the motor Km The power signal makes reference to the voltage to be applied to the motor. 61 (4.28) Section 5. Conclusions 5 Conclusions 5.1 General Conclusions Although interesting research has been done in the field of LOS stabilization and many papers has been published, most of that work has a tendency to offer idealistic assumptions related the real implementation of the system. One of the major contributions of this work is precisely it focus in analyzing the influences of practical constrains faced when implementing the LOS stabilization system. Along with this analysis a novel Extended Kalman filter was developed and an optimal LQ-controller was designed for stabilizing the elevation axis. The Extended Kalman filter proved to be remarkably accurate in estimating the nonlinear CFT compared to the approach described in [8]. Using this estimation to cancel the nonlinear torque showed a great improvement in the cancellation rate achieved. Modifying the Extended Kalman filter f ore s t i ma t i ngt hea n g l eθ Iy allowed the design of an LQ-controller. The LQ-controller proved to be much more effective than the proportional controller; achieving a significant r e duc t i onofωIy a ndθ Iy variances along with a complete cancellation of the θ Iy steady-state shift observed when using the proportional controller. It is interesting to notice the small influence of the quantization step size in the overall performance. This will allow a considerable reduction in the computational load and the use of cheaper A/D converters. Theus eo fa no bs e r v e rt oe s t i ma t eωIy made the system highly insensitive to the transfer function –when the sampling frequency was high enough to model it- and measurement noise oft h eg y r o s c opeus e dt os e ns eωIy. The measurement noise o ft heg y r os c opeu s e dt os e n s eωOy did not affect appreciably the results except for large noise variance. The disturbance-bandwidth rejection range (DBRR) attested to depend essentially on two factors: sampling frequency and sensor bandwidth (ofg y r os c op es e n s i ngωOy). 62 For high Section 5. Conclusions frequency sampling the DBRR was exclusively limited by the sensor bandwidth. Even when using a 50Hz sampling frequency the sensor bandwidth was the major limitation for the DBRR for cut-off frequencies below 150Hz; above that frequency the main constraint was imposed by the sampling frequency. This is of key importance when selecting the gyroscope optimal characteristics. If the system will work at 50Hz it is pointless to buy gyroscopes with bandwidths higher than 150Hz. Finally, simulations showed the extreme degradation produced when saturation conditions were reached; therefore it will be imperative to avoid these conditions modifying the controller law or changing the motor. 5.2 Future Work The main goal of any future work should be to complete the work done in section four, with the inclusion of the outer gimbal equations to design a controller for the whole system. It would be also valuable to develop and analyze the possible use of alternative controller designs, e.g. robust controllers, adaptive controllers, neural networks, and to perform a detailed analysis to define the best cost-benefit arrangement for the sensors. 63 Section 6. References 6 References [1] Peter J. Kennedy, Rhonda L. Kennedy, Direct versus indirect line of sight (LOS) stabilization, IEEE Trans. Controls System Tech. Vol.11, Nº1, 2003, 3-15. [2] Piezoelectric Vibrating Gyroscopes (GYROSTAR), Murata Manufacturing Co., Ltd, Kyoto, Japan, November 2002. [3] Brush Type DC Motors Handbook, Axsys Technologies, 2005. [4] Marcel J. Sidi, Spacecraft Dynamics and Control: A Practical Engineering Approach, Cambridge Aerospace Series, Cambridge University Press, 1997. [5] H. Ambrose, Z. Qu, R. Johnson, Nonlinear robust control for a passive line-of-sight stabilization system, Proc IEEE, Conference on Control Applications, Sept 2001, 942-947. [6] T. H. Lee, E. K. Koh, M.K.Loh, Stable adaptive control of multivariable servomechanisms, with application to a passive line-of-sight stabilization system, IEEE Trans. Industrial Electr. Vol 43, N°1, 1996, 98-105. [7] Marcelo C. Algrain, James Quinn, Accelerometer Based Line-of-Sight Stabilization Approach for Pointing and Tracking Systems, IEE Conference on Control Applications, Sept 1993, 159-163. [8] Bo Li, David Hullender, Mike DiRenzo, Nonlinear Induced Disturbance Rejection in Inertial Stabilization Systems, IEEE Trans. Control Systems Tech. Vol.3, Nº3, 1998, 421-427. [9] B. Li, Identification and compensation for Coulomb friction in stochastic systems, Ph.D. dissertation, Univ. Texas Arlington, Dec. 1994. [10] D. A. Hullender and B. Li, Application of advanced control techniques to line-of-sight stabilization systems, Texas Instruments, Inc., Dallas, TX, Final Rep., Feb. 1994. [11] Greg Welch and Gary Bishop, An Introduction to the Kalman Filter, University of North Carolina at Chapel Hill, 2001. 64