An Autonomous Control Technique for Launching Ship Based Unmanned Air Vehicles (Uavs) in Extreme Conditions

ICAS 2002 CONGRESS

AN AUTONOMOUS CONTROL TECHNIQUE FOR LAUNCHING SHIP BASED UNMANNED AIR VEHICLES (UAVS) IN EXTREME CONDITIONS

M.R. Crump , C. Bil Sir Lawrence Wackett Center for Aerospace Design Technology, RMIT University GPO Box 2476V Melbourne 3001, Vic. Australia.

Keywords: UAV, launch, control, N4SID

Abstract intensive and prone to both technical and human factor complications which can result in This paper presents a solution to the au- the complete loss of the aircraft and full mis- tonomous control problem for launching un- sion failure. This situation is exacerbated manned air vehicles from maritime platforms by extreme environmental conditions (both in extreme conditions. A controller is de- oceanic and atmospheric). Autonomous con- veloped utilising both linear robust modern trol of the UAV during the launch procedure control theory (H∞ methods) and a new ap- has been identified as a key technological re- proach involving trajectory optimisation and quirement for ship based fixed wing UAV op- system identification. The resulting controller erations. is found to allow the aircraft to fly to its phys- The major problem identified when ical limits. This controller has been found to launching a small UAV from a ship deck have very good robustness properties through results when a large tailwind is present. This extensive simulation. large tailwind reduces the aircraft airspeed during the launch process, resulting in both 1 Introduction loss of controllability and aircraft stall, either of which can be catastrophic during launch. The ability to launch ship based Unmanned As opposed to launch from conventional Air Vehicles (UAVs) in extreme sea conditions runways, where multidirectional launch capa- would greatly extend the usefulness of these bility is generally present, a relatively small aircraft for both military and civilian mar- ship deck presents limited launch directions itime purposes. Currently the majority of ship and it is undesirable to manoeuvre the ship based UAVs are either VTOL aircraft or fixed to allow a feasible launch direction. Thus, wing aircraft remotely launched under limited launch in these tailwind conditions becomes conditions. The advantages of fixed wing air- necessary in some situations. craft over VTOL aircraft in terms of flight Other problems identified include large speed, range and endurance are well known, amounts of turbulence through the ship air- and can be of great benefit to various missions. wake presenting large disturbances to the air- Thus there exists a need for fixed wing UAV craft which a controller (or human pilot) must operation off maritime platforms. be able to cope with. Remote launch procedures involving an ex- This paper presents a control and guid- ternal human pilot are generally very labour

562.1 M.R. CRUMP , C. BIL ance strategy that allows autonomous launch located on the twin vertical tails. The simula- of a fixed wing UAV under a large set of en- tion model has been slightly modified to give vironmental conditions, including conditions a slight increase in aerodynamic performance exhibiting the turbulence and tailwind prob- around the high angle of attack region. The lems. It is found that the limiting conditions modified model includes hypothetical aircraft for launch when implementing this controller stall and post-stall effects, nonlinear actuator are physical limits that are set by the air- effects (including rate and positional satura- craft dynamics and not those set by the con- tion) and sensor limitations. troller. Thus, the controller allows aircraft The ship model is based upon seakeeping launch through the entire physically possible theory [10] in conjunction with measured data aircraft launch envelope. [6] and some results of simulations based upon This controller is developed using a com- a generic frigate type vessel. The simulation bination of both modern linear robust con- allows prerecorded (or artificially generated) trol techniques (H∞ ) and a newly proposed motion data to be used, or simulates generic control method. The robust linear controller ship motion using a six degree of freedom sim- controls the aircraft through the majority of ple harmonic motion model. The model takes the launch phase, guiding the aircraft along account of the ship heading with respect to a nominal trajectory whilst rejecting external wave heading, sea state, ship speed and ba- disturbances in the form of turbulence from sic ship characteristics (such as length). The the ship airwake. The controller has been de- ship model is that of a generic frigate, typical signed using a variation of the H∞ mixed sen- of the Australian designed ANZAC Frigate or sitivity approach. the Oliver Hazard Perry Class FFG’s in ser- Control for high tailwind launches is pro- vice with the Australian Navy. vided initially by a Trajectory Optimisation The launcher model uses linear dynamics with Inverse System Identification (TOISI) to accelerate the aircraft while constraining controller. This controller guides the aircraft aircraft rotation in all dimensions and motion through a minimum altitude loss trajectory, in the ramp normal directions. This simulates whilst maintaining a nominated climb angle. an attachment between the launcher and air- Once a safe climb regime has been obtained, craft, that is released at the desired launch the controller switches to the linear robust point. The model can also simulate a ’free’ controller, which guides the aircraft to a safe launcher, in which the aircraft is free to lift off altitude where a mission controller may oper- the ramp when it has attained sufficient speed. ate. The simulation model also has the capability This technique has been tested comprehen- to simulate rocket assisted takeoff, removing sively in nonlinear simulations of the aircraft the need for a launch ramp. and has been found to follow the physically The atmospheric model contains steady optimal launch trajectory in the presence of state wind components, turbulence spectra both uncertainty in the simulation models and (based upon Dryden spectra shown in MIL-F- external disturbances due to turbulence. 8785C [1]), gust components and windshear el- ements which are based upon the effects of the 2 Dynamic Models. ship airwake and ocean surface effects. These approximations are included as there is lit- The aircraft used for development of the con- tle published data on ship airwakes (however troller is based upon the University of Sydney there are current studies [5, 7]). UAV Ariel [9]. This aircraft is a conventional Further details of all the models used may fixed wing UAV with four main control inputs, be found in Crump [4, 2, 3]. throttle, elevators, ailerons and twin rudders

562.2 An Autonomous Control Technique for Launching Ship Based Unmanned Air Vehicles (UAVs) in Extreme Conditions 3 Control Objectives needs to be controlled. the main objective is to maintain a positive altitude. This implies When launching any aircraft, several objec- that altitude needs to be controlled, however tives arise. The first and main objective is the accuracy of various altitude measuring de- to maintain a positive altitude. Several other vices is somewhat questionable, so controlling sub-objectives may also be specified. The first altitude using these devices may not be fea- is to gain altitude as rapidly as is safely possible. Instead, the altitude can be indirectly sible. The second is to maximise the distance controlled by maintaining a positive rate of between the aircraft and the ship. climb. Whilst it is possible that the UAV mission Whilst maintaining a climb angle, it is also controller is capable of launching the aircraft important to maintain the aircraft’s airspeed, autonomously, it is unlikely that it will suc- and as such a control on the aircraft’s velocity cessfully fulfill the launch objectives in an op- is also important. To simplify the flight condi- timal manner. Considering the fact that the tion, it is also desired that the aircraft is flying majority of flight control in today’s aircraft is in the conventional upright position with lit- digital, the addition of another controller to tle bank angle and sideforce. To this end, a specifically control the aircraft during launch simple bank angle control will also be imple- should not prove too cumbersome in terms of mented with a sideforce regulator (if desired physical implementation. these two controllers may be changed into a The first step in the design of a launch con- heading hold controller). troller is to determine the desired launch tra- The control of a UAV during catapult jectory that will be followed in the nominal launch is a problem involving rapidly chang- case (this trajectory will change according to ing dynamics over a short period of time, ambient conditions). This trajectory should corresponding to the first few seconds after maintain a positive altitude at all times, with ramp departure, followed by a longer period as high a rate of climb (ROC) as possible dur- of slowly changing dynamics during the air- ing the early stages of the flight. Maintain- craft climb. This problem suggests itself to a ing this high rate of climb will require a high dual control solution. angle of attack, so in order to prevent stall A highly robust nominal controller de- occurring regularly (due to gusts and turbu- signed using linear robust modern control the- lence), this rate of climb should be reduced ory is implemented for use when the aircraft is slightly to give a margin of safety. This rate flying well within the linear flight regime. This of climb should also be a sustainable rate of controller should be capable of guiding the air- climb, corresponding to a trimmed condition craft through the launch phase in nominal or of the aircraft. near nominal conditions, however it is unrea- For the case of the UAV Ariel, a launch sonable to expect this controller to be capable speed of 30 m/s with a rate of climb of climb of of handling the rapidly changing nonlinear be- 6.2m/s is chosen for the launch trajectory, giv- haviour when the aircraft is well out of its lin- ing a climb angle of 12◦. To impart this initial ear flight envelope (such as when the aircraft climb angle, the aircraft is launched along an is experiencing strong head or tail winds). inclined ramp (at 12◦), with a ramp exit speed For control through this region, a novel of 30m/s. For control design, this model is lin- nonlinear control design technique has been earised and then reduced, removing the three implemented to control the aircraft. This tech- location states and the yaw angle state. This nique involves an optimisation of a desired tra- reduction of states results in the controllers jectory based on the nonlinear model and then generated having lower order. the identification of an appropriate controller. The other important consideration is what

562.3 M.R. CRUMP , C. BIL

4 Nominal Controller Design in nominal launch conditions, with low steady wind conditions. In the case of mid strength The mixed sensitivity H∞ approach is an easy winds, the linear controller fails and launch to apply formulation of the H∞ control design is unsuccessful. Further results for the linear problem. This method has been used success- controller performance can be found in [2]. fully for many different applications (including aircraft applications) and whilst a relatively 5 Trajectory Optimisation with In- simple member of the H∞ control family, still verse System Identification provides reasonable performance and robustness with a simple design procedure. The As stated in the previous section, the lin- mixed sensitivity H∞ problem is one that in- ear controller performs very well in nominal volves the minimisation of a combination of launch conditions. However, the presence of different weighted transfer functions, usually high head or tail winds during the launch pro- involving the sensitivity function S along with cess causes problems. High head winds cause others such as the complementary sensitivity a large overspeed launch condition. The con- function T and KS. This is technique involves troller attempts to reduce speed, and in ex- adjusting the shape and magnitude of the vari- treme conditions, the climb performance of the ous weights to obtain satisfactory performance aircraft is reduced. This problem is quite easy of the system. to solve using nonlinear error weighting, de- The combination selected for use is the tails of which can be found in [4]. S/KS/T mixed sensitivity approach with gust The problem of tail wind launches however rejection properties, which includes weights is more complicated. In the presence of large upon T for tracking purposes and a weight tailwinds, the aircraft airspeed is dramatically upon KS to limit control activity. Using this reduced during the launch process. This prob- approach, it is possible to limit the control lem leads to aircraft stall, with the resulting signal, whilst maintaining acceptable perfor- aircraft motion more than likely resulting in mance in the presence of gust disturbances. catastrophic launch failure. This problem gives the H∞ optimisation prob- This problem has been overcome using a lem to find the controller which minimises: new control technique termed Trajectory Op- timisation with Inverse System Identification W1SW3 −W1SGdW4 ° ° (TOISI). This technique can be logically bro- ° W2KSW3 −W2KSGdW4 ° (1) ° ° ken into two distinct components, the trajec- ° T W3 SGdW4 ° ° °∞ tory optimisation component and the inverse ° ° system identification component. The weights use integration in W1 to en- sure no steady state error and consistent low 5.1 Trajectory Optimisation frequency behaviour. W2 is used to minimise the high frequency control activity, whilst not The first step is to determine certain time do- penalising low frequency behaviour. W3 is main constraints that must be met. In this used to scale the inputs to give desired be- case, this is accomplished by the following con- haviour, with a large emphasis placed upon straints climb angle tracking and W4 is used to bal- ance the emphasis between performance and • Minimal altitude loss over first seconds gust rejection. of flight This controller exhibits very good linear • Minimum error in climb angle tracking performance with small rise times, little over- shoot and very good gust rejection qualities. The first and most important of these con- The controller also performs extremely well straints is to minimise any loss in altitude.

562.4 An Autonomous Control Technique for Launching Ship Based Unmanned Air Vehicles (UAVs) in Extreme Conditions The aerodynamic nature of aircraft means This optimisation is performed on the full that in low speed situations it may be neces- nonlinear model, and on a Intel P3-700 com- sary to reduce altitude to allow the velocity to puter running the MATLAB optimiser on a increase to amounts necessary for a sustained compiled C++ simulation model will converge climb. The ﬁrst constraint emphasises that satisfactorily in about 10 hours for a 100Hz little altitude may be lost in the launch situa- sampling rate. However by ﬁrst utilising lower tion. The second constraint similarly ensures sampling rate estimations (10Hz and 50Hz) as that the aircraft is climbing as rapidly as is starting points, this time can be reduced to physically possible. about 5 hours. These constraints can be expressed math- Several sample trajectories and elevator in- ematically as a cost function, giving puts are shown in Figure 1. These trajectories

t=5 2 2 Cost = 10000 (hr − hmin) +10 (γopt − γ (t)) t X 15 t=tl (2) 16m/s tailwind 10 13m/s tailwind where hr represents the height of the launch 10m/s/ tailwind ramp, hmin is the minimum altitude for a 5 flight, tl is the launch time, γopt is the opti- 0 mum climb angle and γ is the time varying Altitude (m) climb angle over the duration of the simula- −5 tion (5 seconds in this case). The altitude loss term is highly weighted and the climb angle −10 term is time weighted. 20 This cost function can then be minimised 10 using a multivariable optimisation procedure. 0 The purpose of this optimisation and hence the optimisation variables are the necessary −10 control inputs to minimise this cost. Intu- −20 itively this situation will require full throttle, −30 thus removing this control from the optimisa- Elevator Deflection (deg) −40 tion, and as the problem is mainly a longitu- 0 1 2 3 4 5 dinal one, the lateral nominal controller can Time (secs) be used to provide appropriate lateral control Fig. 1 Optimal launch trajectories in tail- during the flight. This leaves simply the ele- winds vator deflection to be optimised. Differing sample times may be utilised for the elevator deflection commands, resulting in show the physical limitations of the aircraft different accuracies in the trajectory. It has (as indicated by the mathematical model) and been found that a sampling frequency of 100Hz indicate the best achievable trajectory. This is provides the best results, with higher frequen- what the controller should achieve. A 10m/s cies giving no improvement. The 5 second sim- tailwind results in no loss of altitude, a 13m/s ulation results in 4.3 seconds of flight (once tailwind results in no loss of altitude, however the launch has occurred), resulting in 430 el- a short period of horizontal flight is necessary. evator deflection variables. An optimisation Finally a 16m/s tailwind requires a small loss of these deflections can be performed, ensur- in altitude. ing that any physical constraints such as positional and rate saturations are included.

562.5 M.R. CRUMP , C. BIL

5.2 Inverse System Identification case, good linear performance does not nec- essarily imply good nonlinear performance. In Now given the optimal control inputs and fact, a controller that almost perfectly gener- the corresponding system outputs, a controller ates the optimal elevator deflection can have can be produced that generates the optimal extremely poor and even unstable nonlinear control inputs from the system outputs. In performance. On the other hand, a con- this format, the problem lends itself to a troller with satisfactory linear performance differing use of existing system identification (having small errors between the identified techniques. These techniques generally have case and the nominal case) can have excel- known system inputs which generate measured lent nonlinear performance. Due to this lin- system outputs, the combination of which can ear/nonlinear mismatch between controllers, be used in various algorithms to predict dy- many controllers can be generated, which can namic models of the system. then be tested and either kept or discarded de- A similar case exists here where the input pending upon their linear and nonlinear per- to the controller is known (the system mea- formance. Due to the nature of the process, surements) and the output of the controller this can be achieved quite rapidly. is known (the optimal control inputs). Thus, these system identification algorithms should 6 Controller Performance be able to identify the controller that generates the optimal control signals given the mea- Using the method of the previous sections, a sured system outputs. controller has been generated that gives very The system identification algorithm used good nonlinear performance. This controller for this is a modified version of the N4SID will be used for nonlinear simulation tests of technique of Van Overschee and De Moor [11] the nominal aircraft launch system, the dis- which utilises subspace signal processing tech- turbed aircraft launch system and also for ro- niques to identify a dynamic model from in- bustness tests where the aircraft dynamics are put/output data. perturbed. The modifications to this algorithm in- Using the three tailwind launch cases of clude a stability modification suggested by the previous section in the original high fi- Maciejowski [8] and also a multiple data set at delity nonlinear aircraft model with the fully once (MDSAO) modification to give a single implemented nominal and inverse optimisa- best fit controller over several input/output tion identification controller, the simulation data sets. This modification is achieved results shown in Figure 2 are produced. These through a matrix augmentation of the in- results show that the controlled trajectory is put/output data within the algorithm. Due very close to the physically optimal trajectory to space limitations, this modification will not and that the controller performs well in the be explained in detail. Further detail can be nonlinear system. Note that this case includes found in [2, 3]. the effects of sensor noise, but contains no at- Using this technique, many different con- mospheric turbulence or gusts. trollers can be generated by using different More realistic simulation results are shown combinations of system output signals as con- in Figures 3, 4 and 5 where the aircraft is sub- troller inputs and by using various combina- jected to a severe turbulence level in a gusty tions of signal delays, initial conditions and environment with a 10m/s tailwind, headwind noise estimates. These controllers will have and crosswind respectively. A computational clearly stated linear performance, based upon time delay of 0.05 seconds is also included. how well they generate the optimal elevator These results show successful launches for all control signals, however as is generally the three cases, however the tailwind launch case

562.6 An Autonomous Control Technique for Launching Ship Based Unmanned Air Vehicles (UAVs) in Extreme Conditions

15 40 35 10 30 (deg) (m/s) α 25T 5 20 15 Tailwind TailwindHeadwind 0 10 HeadwindCrosswind Total Velocity V Angle of attack 5 Crosswind 5 0 0 1 2 3 4 5 0 1 2 3 4 5 15 20 10 5 15 (deg) (deg) 0 γ φ 10 -5 -10 5 -15 -20 Bank Angle Climb Angle 0 -25 -30 -5 0 1 2 3 4 5 0 1 2 3 4 5 Time (secs) Time (secs)

Fig. 3 Full Simulation Results

15 This diﬀerence is due to the turbulence and 10m/s Tailwind Controlled 10m/s Tailwind Optimal gust inﬂuence, where in this case (as shown in 13m/s Tailwind Controlled 10 13m/s Tailwind Optimal Figure 5) the aircraft experiences a tailwind 16m/s Tailwind Controlled 16m/s Tailwind Optimal gust during the launch process. The aircraft

5 also experiences vertical turbulence which has a small negative impact. The other graphs show other aircraft pa- Altitude (m) 0 rameters during the launch. The tailwind launch requires a high angle of attack. The −5 stall angle of the Ariel is approximately 12 ◦, thus, the aircraft actually stalls very briefly −10 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 before the controller rectifies the situation. Time (secs) The ramp exit velocity shows the difference between headwind and tailwind case, with the Fig. 2 Altitude aircraft in the headwind case leaving the ramp with an airspeed of about 40m/s. The aircraft shows greater difficulty in maintaining altitude in the tailwind case leaves the ramp with an than for the previous 10m/s tailwind case. airspeed of only 20m/s. The climb angle is

562.7 M.R. CRUMP , C. BIL

Fig. 4 Full Simulation Results maintained well for the headwind and cross- example, the lift curve slope may be increased wind cases and only drops below zero for a (by a reasonable amount) with no performance short period for the tailwind launch. The lat- problems. Similarly, it may be reduced with eral controller works well, even in the pres- no control problems, however due to the low- ence of large crosswinds. Finally, in the head- ering of the lift force, the launch trajectory wind case, little throttle is needed due to the is compromised. Similar processes may be large launch velocity. A significant amount of used for the other coefficients. The only prob- rudder control is used to control the sideforce lems for control exist when the moment deriva- present from the crosswind. tives change excessively, resulting in instabil- The other influence that must be tested is ity. This excessive change however is unlikely the robustness of the controller to unmodelled to occur if due care is taken when developing and incorrectly modelled dynamics. This can the mathematical models. be partly achieved by perturbing some of the These robustness tests have been carried aircraft aerodynamic coefficients (by changing out over thousands of different simulation slopes, adding constants etc). This process cases involving varying ship positions and has been utilised, modifying all aircraft dy- wind influences. Nearly 100% of these simula- namics by varying amounts, resulting in ex- tions show successful launches, with the few tremely good robustness of the controller. For failures explained by extremely large devia-

562.8 An Autonomous Control Technique for Launching Ship Based Unmanned Air Vehicles (UAVs) in Extreme Conditions

15 15

10 10 5

0 5

−5 East wind (m/s)

North wind (m/s) 0 −10

−15 −5 0 1 2 3 4 5 0 1 2 3 4 5 3 30 25 Tailwind 2 Headwind 20 1 Crosswind 15 0 10 5

−1 Altitude (m) 0 −2Vertical wind (m/s) −5 −3 −10 0 1 2 3 4 5 0 1 2 3 4 5 Time (secs) Time (secs)

Fig. 5 Full Simulation Results tions in aircraft parameters, excessively strong The methods used here are directly appli- tailwinds or vertical gust and turbulence prob- cable for controller design for launch of con- lems on ramp exit. ventional fixed wind UAVs and they should be easily applicable for other control situations requiring short term control of systems requir- 7 Conclusions ing quite stringent time domain constraints. It is the belief of the authors the the meth- A controller has been developed using H ∞ ods presented here could easily be utilised for modern control techniques and a new method launch and recovery control of VTOL UAVs in termed Trajectory Optimisation with Inverse similar shipboard situations. System Identification for launch control of UAVs in extreme conditions. This controller References is found to allow the aircraft to fly extremely close to its physical limits, meaning that the [1] Anon. Flying qualities of piloted airplanes. launch envelope is only limited by the aircraft Technical Report MIL-F-8785C, 1985. and not the controller. The controller has been [2] Crump M. R and Bil C. Control techniques found to be very robust to external distur- for autonomous launching of ship based un- bances and model uncertainty through simu- manned air vehicles (uavs). Proc Bristol lation tests. UAV Conference, Bristol, 2001.

562.9 M.R. CRUMP , C. BIL

[3] Crump M. The Dynamics and Control of Catapult Launching Unmanned Air Vehicles from Moving Platforms. PhD Thesis, RMIT, 2002. [4] Crump M, Riseborough P, Bil C, and Hill R. Dynamic control aspects of the shipboard launch of unmanned air vehicles. Proc ICAS 2000, Harrogate UK, 2000. [5] Healey J. Establishing a database for ﬂight in the wakes of structures. Journal of Air- craft, Vol. 29, No 4, pp 559–564, 1992. [6] Hope K. Angular Motions of an Australian FFG in rough Seas SS6-7 - Statistics and Spectra. Australian Defence Force Academy, 1996. [7] Liu J and Long L. N. Higher order ac- curate ship airwake predictions for the he- licopter/ship interface problem. Proc An- nual Forum Proceedings - American Heli- copter Society, Vol. 1, 1998. [8] Maciejowski J. Guaranteed stability with subspace methods. Systems & Control Let- ters, Vol. 26, No 2, pp 153–156, 1995. [9] Newman D and Wong K. Six Degree of Freedom Flight Dynamic and Performance Simulation of a Remotely-Piloted Vehicle. Aero.Note 9301, The University of Sydney, 1993. [10] Rawson K and Tupper E. Basic Ship Theory Volume 2. Longman, 1983. [11] Van Overschee P and De Moor B. N4sid : Subspace algorithms for the identiﬁcation of combined deterministic-stochastic systems. Automatica, Vol. 30, No 1, pp 75–93, 1994.

562.10