Ningjun LIU, Zhihao CAI, Jiang ZHAO, Yingxun WANG

School of Automation Science and Electrical Engineering, Beihang University, Beijing 100083, China

KEYWORDS

Abstract An attempt is made to apply modern control technology to the roll and yaw control of a rudderless quad-tiltrotor Unmanned Aerial Vehicle(UAV)in the latter part of the flight mode transition, where aerodynamic forces on the tiltrotor’s wings start to take effect. A predictor-based adaptive roll and yaw controller is designed to compensate for system uncertainties and parameter changes.A dynamics model of the tiltrotor is built.A Radial-Basis Function(RBF)neural network and offline adaptation method are used to reduce flight controller workload and cope with the nonlinearities in the controls.Simulations are conducted to verify the reference model response tracking and yaw-roll control decoupling ability of the adaptive controller,as well as the validity of the offline adaptation method.Flight tests are conducted to confirm the ability of the adaptive controller to track different roll and yaw reference model responses.The decoupling of roll and yaw controls is also tested in flight via coordinated turn maneuvers with different rotor tilt angles.

1. Introduction

The tiltrotor is a relatively new type of Vertical Take-Off and Landing (VTOL) aircraft.1Existing models such as Bell-Boeing V-22, IAI Panther and the newly developed Bell V-280 have repeatedly demonstrated their excellent flight performance.2–4Despite the structural differences, all these tiltrotors possess both the VTOL abilities of helicopters and the high speed, range and endurance of fixed-wing aircrafts.This advantage comes from tiltrotor’s ability to transit its engine thrust vector from the vertical ‘‘helicopter” mode to the horizontal ‘‘fixed-wing” mode, and vice versa.1Consequently the challenges in tiltrotor aircraft design and development are mainly the rotors’ tilt transition mechanism and the corresponding transition control.5–9

Since the advent of fly-by-wire systems, modern control technologies have been widely used in aircraft flight control for both helicopters and fixed-wing aircrafts, and have achieved remarkable improvements in aircraft performance compared with the traditional Proportional-Integral-Deriva tive(PID) or state-feedback controllers.10However, the documentation and literature of the use of modern control technique in tiltrotors are relatively rare.1In this study, a modern control method, specifically predictor-based adaptive control,11,12is applied for the roll and yaw control of a rudderless quad-tiltrotor Unmanned Aerial Vehicle (UAV) during the latter part of the flight mode transition, where aerodynamic forces on the tiltrotor’s wings start to take effect. A Radial Basis Function (RBF) neural network13,14is also applied to improve controller performance.

In the work, a dynamics model of the tiltrotor is built. A predictor-based adaptive roll and yaw controller is designed to compensate for system uncertainties and parameter changes. An offline adaptation method is designed to reduce flight controller workload and cope with the nonlinearities in the controls.Simulations are conducted to verify the reference model response tracking and yaw-roll control decoupling ability of the adaptive controller,as well as the validity of the offline adaptation method. Finally, flight tests are conducted to confirm the ability of the adaptive controller to track different roll and yaw reference model responses.The decoupling of roll and yaw controls is also demonstrated via coordinated turn maneuvers with different rotor tilt angles.

2. Aircraft dynamics modelling and problem formulation

2.1. Quad-tiltrotor configuration

The studied rudderless quad-tiltrotor configuration is shown in Fig.1. The aircraft consists of a fuselage, fore and aft wings,left and right elevons, a vertical tail, and four tilt-able rotors at each wingtip. All four rotors are assumed to be rigid, without flapping or lead-lag blade motion.A single actuator is used to control all four rotors’ tilt angles, and make them always equal. There are seven controllable variables in all: deflections of the left and right elevons δeland δer,the thrust command of each of the four rotorsui（i=1，2，3，4）, and rotor tilt angle δt.When the aircraft is in the fixed-wing mode,δt=0,and in the quadrotor helicopter mode, δt=π/2.

Fig.1 Quad-tiltrotor configuration.

2.2. Dynamics model

The structure of the aircraft dynamics model is shown in Fig.2. It is assumed that the aircraft is near a trimmed‘‘straight and level” flight state. In the model, the aircraft is divided into three parts and modelled separately:

(1) Gravity force experienced by the aircraft;

(2) The fixed-wing part, which includes the fuselage, fore and aft wings, vertical tail, and the elevons. Forces and moments produced in this part can be estimated using wellestablished methods.15–17In the method, forces and moments produced by each aerodynamic component (e.g. wing or fuselage)are calculated separately.They are then transformed into the aircraft body axes and summed. For example, for the fore left wing, the local airspeed vector is

whereVis the aircraft’s airspeed vector, ω=[p，q，r]Tis the aircraft’s angular rate vector, andRlocalis the location vector of the fore left wing’s overall aerodynamic center in the body axes. The angle of attack is then deduced fromVlocal, with the wing’s incidence angleaiadded:

The lift, drag and moment are then calculated via

(3) The rotor part, which includes the tilt mechanism,motors and rotors. The force and moment produced in this part are from rotor thrust. An iterative algorithm18–20is used to calculate rotor thrust at different inflow angles.In this algorithm, the rotation speed calculated via:

whereuis the thrust command sent, andkΩis determined experimentally. Assuming the aircraft’s angle of attack is low, a solution for the rotor’s induced speedvi(the change in flow speed from far in front of the rotor to the rotor disc)is deduced as17,19

Fig.2 Dynamics model structure.

whereVis the aircraft’s airspeed,Tis the thrust, andAis the rotor disc area.viand Ω are then used in numerical blade element calculations to obtain a new solution forT. By applying suitable iterative solving methods, the resultant rotor thrust can be organized into a function of airspeed and tilt angle:

Each individual rotor’s thrust is calculated fromT（V，δt）,and then transformed forces and moments in the body axes.They are then summed up to form the output of the rotor model.

After the forces and moments from the fixed-wing and rotor models are calculated, they are summed and then fed into six-degrees-of-freedom equations of motion15to calculate the aircraft’s motion states, such as angular velocity, attitude angles, etc. These states are then looped back to calculate forces and moments, forming a closed-loop dynamics model.In a real tiltrotor there are also more complex aerodynamics phenomena, such as coupling between the fixed-wing part and rotor part. These are ignored in the model, and in real flight they are treated as uncertainties. For example, part of the wing in the rotor’s wake can reduce the rotor’s thrust; this is treated as an uncertainty in the rotor thrust produced.

2.3. Model simplification

The original dynamics model as obtained in Section 2.2 is quite complex, and needs to be simplified for controller design. The original roll and yaw dynamics equations in six-degrees-offreedom equations of motion15can be represented as:

whereLis the roll moment,Nis the yaw moment,Jxx,JzzandJzxare elements of the inertia tensor matrix. In this paper it is assumed that the absolute value ofJzxis much lower than bothJxxandJzz, and that the pitch rate of the tiltrotorqis close to zero. Eq. (7) can then be simplified as

These simplified equations are then used in controller design.

2.4. Roll and yaw control moment generation

During the latter part of ‘‘helicopter” to ‘‘fixed-wing” flight mode transition, the aircraft has gained some airspeed, and aerodynamic forces produced by the fixed-wing part become quite significant.Differential deflection of the elevons can then be used to produce roll control moments.This control input is set as

The linearized equation for the roll control moment produced is then

whereLδarepresents the roll moment generated per unit δa,i.e.dL/dδa.

The rotors’ tilt angle δtat this time is approximately in the range of 0.1–0.7 radians (5–40 degrees), and the horizontal component of rotor thrust (as shown in Fig.3) is quite large.This horizontal thrust component can then be utilized to produce yaw control moments, by creating a thrust difference between rotors on the left and right sides (Fig.4).

This control input is set as:

The linearized equation for the yaw control momentNc(Fig.4) is then

whereTuis the rotor thrust control gain, andbis the lateral distance between left and right rotors.

What needs to be noticed here is, when δaQis utilized, the vertical thrust components of rotors on the left and right sides are also different. And this produces a disturbance roll momentLd, as seen in Fig.4:

Besides elevon deflection and rotor thrust,rotor torque also produces roll and yaw moments, but are not used for control here. This is because the absolute values of the moments produced are usually very low,21and it is very difficult to obtain effective control by increasing the rotor torque without overloading or saturating the rotors’ command inputs.

2.5. Problems in control

Fig.3 Horizontal and vertical components of rotor thrust.

Fig.4 Left-right thrust difference produce moments.

Traditional control methods such as state feedback or PID are often used for well-defined systems. However if high amounts of model uncertainties that change with different flight conditions are encountered, the control accuracy and stability are often reduced.22

Some aircraft dynamics model parameters can be accurately obtained via wind tunnel tests, extensive CFD calculations, or comprehensive system identification. For small UAVs however, such activities are usually not conducted due to high costs. This leads to large amounts of uncertainties existing in the flight dynamics model,such as rotor thrust,elevon effectiveness and angular rate dampening coefficients.Besides that, because of the tiltrotor’s tilt transition mechanism, some model parameters can change drastically in different flight conditions(e.g.airspeed and rotor tilt angle).If only traditional control methods are employed for roll and yaw control, the consistency of control responses would be unsatisfactory.

To provide more consistent control performance under different flight conditions, a modified ‘‘PMRAC” (Predictorbased Model Reference Adaptive Control) controller11,12is designed for the tiltrotor’s roll and yaw control.The controller is supplemented by an RBF neural network that is used to‘‘memorize” the adaptive parameters and run the adaptation process off-line,so that the adaptation results in different flight conditions can be reused.

3.Predictor-based model reference adaptive roll and yaw control

3.1. Application strategy

In the study,the adaptive controller is inserted into the control loop, and drives the aircraft’s control responses to those of a pre-set reference dynamics model by modifying the servo control commands (δaand δaQ) based on aircraft system states(e.g.p,r,V, δtetc.). This creates a ‘‘virtual aircraft” with the reference model’s dynamics properties. A simple fixed-gain PID-style baseline controller is used for attitude and angular rate tracking. The entire system structure is shown in Fig.5.

The adaptive controller can also be designed such that it takes attitude or angular rate feedback as input directly to track their commanded values, without the inserted baseline controller. But this ‘‘a(chǎn)ll-adaptive” strategy is not used for the following reasons:

(1) The final form of an adaptive controller is often rather restrictive,e.g.it has to be in a state-feedback form in some situations. The baseline controller in Fig.5 provides more flexibility in design.

(2)As both the aircraft’s inherent properties and the desired attitude/angular rate tracking performance have to be taken into consideration simultaneously, it can be difficult to adjust and tune the ‘‘a(chǎn)ll-adaptive” controller properly.

In Fig.5,the function of the flight controller is divided into two parts, where the baseline controller controls attitude and angular rates,and the adaptive controller maintains consistent‘‘virtual aircraft” dynamics properties. Therefore the different functions can be tuned and adjusted separately.

3.2. Baseline controller

The baseline PID controller provides basic attitude stabilization and control for the simulations and flight tests.

The roll and pitch angular rate controllers are PI controllers supplemented with a feedforward component:

The yaw angular rate controller is a PD controller:

Fig.5 Structure of the controller-aircraft system.

In Eqs.(15),(17)and(18)kP·,kI·,kD·andkF·are controller parameters, which are designed and tuned using conventional methods in linear control system design.

3.3. Predictor-based model reference adaptive control

The adaptation algorithms for roll and yaw are based mainly on simplified scalar forms of the controller presented in Refs.(11) and (12). For roll control, the linearized equation for roll angular rate acceleration is

δacmdis the original commanded δa; without adaptive control we would have δa=δacmd. To avoid overstressing the adaptive controller,amandbmmust reflect fixed-wing flight dynamics that are not too far from the tiltrotor’s inherent properties.

The state predictor is

The final control structure is shown in Fig.6: In Fig.6,the function of the adaptive controller is to produce a ‘‘corrected input” that drives the output of the state predictor (^p)to that of the plant (p). The inverse of this ‘‘corrected input”is used as the system input, and causes both the plant’s and the state predictor’s outputs to converge to that of the reference system (Eq. (20)). As the reference system is fixed,the adaptive controller provides consistent roll control responses even if uncertainties and changes are present in the system parameters. In particular, the yaw-roll control decoupling problem (described in Section 2.5) is automatically solved, as the effect of δaQis already treated as a ‘‘disturbance” in the state predictor (Eq. (21)), and is compensated for in the adaptive controller (Eq. (23)). Therefore the roll disturbance caused by δaQis automatically cancelled out by the adaptive controller’s δaoutput.

For yaw control, a similar process is used to obtain an adaptive controller.Assuming the angle of sideslip β is negligible, the desired relationship between δaQcmdand yaw angular rateris set as

The state predictor is then

The adaptation process is simply

whereer=^r-ris the predictor tracking error. The control law is

Fig.6 Structure of the adaptive controller.

3.4. Application of RBF neural network

The RBF-NN input is set asVand δt.RBF functions of the form

In this way, at each （V，δt）, the adaptation parameter will be similar to what was created before in nearby （V，δt） points,thus creating a ‘‘memorization” effect.

3.5. Offline parameter adaptation method

In practice, problems can be encountered when the PMRAC algorithm supplemented with RBF-NN is run on an onboard flight controller in flight. The algorithm’s adaptation process increases the flight controller’s workload and at the same time, the adaptation gains (γδa, γpetc.) have to be adjusted so that parameter adaptation is neither too fast(which can lead to reduced robustness, oscillatory transient behavior or numerical instabilities23,24) nor too slow. In addition, the nonlinear nature of the adaptive controller reduces the predictability of controller behavior, and makes controller tuning complex and unintuitive.

To avoid these problems, an offline adaptation method is designed, and its application in roll control is described here in detail. In this method, the PMRAC structure in Fig.6 is divided into two sub-systems: the adaptation loop, as shown in Fig.7, and the control loop, as shown in Fig.8.

Fig.7 Adaptation process.

For yaw control,the same offline adaptation process can be used to create a GS ‘‘a(chǎn)daptive” yaw controller.

In summary, the function of the designed controller is that the tiltrotor’s roll and yaw outputs are driven to follow those of pre-set reference systems at different flight conditions, i.e.airspeed and tilt angle （V，δt）, and with uncertainties in roll/yaw system parameters.

It should be noted that the GS ‘‘a(chǎn)daptive” controller designed here cannot adapt for unexpected in-flight system changes,i.e.changes that were not accounted for in the offline adaptation process, such as disturbance moments at large aerodynamic angles, shift of the aircraft’s center of gravity,etc. Because the GS ‘‘a(chǎn)daptive” controller does not conduct parameter adaptation in-flight compared to a controller designed with online adaptation.

Fig.8 Final GS ‘‘a(chǎn)daptive” control structure.

4. Simulation work

4.1. Reference model response tracking

The purpose of the simulation here is to verify the ability of the PMRAC controller described in Section 3.3 to track different reference model responses at different flight conditions. The structure of the simulation model for roll and yaw are as in Figs. 9(a) and (b).

In the structure,the input to the control system is the servo command δacmd(for roll) or δaQcmd(for yaw). The reference model response roll ratepor yaw raterare calculated from these inputs. The input to the aircraft dynamics model can be switched to either the inputted servo command,or the output of the adaptive controller.In other words,adaptive control can be turned on/off to evaluate differences in plant (i.e. aircraft dynamics) output. To simulate different disturbances and flight conditions, δaQ(only for roll simulation),Vand δtare fed into the plant. The aircraft dynamics model used here is based on the V-44 tiltrotor used in flight tests,as described in Section 5.1.

In the simulations, random δacmd, δaQcmd,Vand δtare generated first and then fed into the systems in Fig.9. The generated random inputs are shown in Fig.10.

Four reference models are set for roll and yaw control respectively as seen in Table 1, to obtain different simulated reference model response tracking results.

Fig.9 Simulation model structure.

Fig.10 Random δacmd, δaQcmd, V and δt inputs.

Table 1 Reference models used in simulations.

In all simulations, the adaptation gains are set at 0.3, and final results are shown in Figs. 11 and 12. In the figures, the reference model responses are calculated as ‘‘ideal”porrand set as the standard to be matched(red line).The simulated aircraft’sporroutput responses with adaptive control turned on (blue line) and off (green line) are then plotted.

In Figs. 11 and 12, it can be seen clearly that the system control equipped with the PMRAC controller can drive the aircraft’s roll and yaw responses much closer to the ‘‘ideal”values obtained from the responses of reference models, compared with those in which adaptive control is turned off. This means that the system response tracking ability and response consistency are effectively improved.

4.2. Offline parameter adaptation

The simulation here is to test the effectiveness of the offline parameter adaptation method discussed in Section 3.5,by evaluating the reference model response tracking ability of GS‘‘a(chǎn)daptive” controllers. The reference models are the same as in the previous simulations (in Table 1).

To make GS ‘‘a(chǎn)daptive” controllers work, the ‘‘raw flight data” are needed in simulation. These data are obtained by simulating aircraft responses under random input servo commands and flight conditions, as shown in Fig.13.

In this step, random δacmd, δaQcmdcommands are given to the aircraft dynamics model directly;Vand δtare also set to random. The simulation is run for 2000 seconds. The outputpandrvalues are saved along with the aircraft dynamics model inputs δa, δaQ,Vand δt, to form raw flight data. After that, for each reference model in Table 1, the obtained raw data are fed into the adaptation loop. When the process is completed, the RBF weighting coefficients (^Θpi, ^Θδaietc.) are extracted and used to form GS ‘‘a(chǎn)daptive” roll and yaw controllers.

Fig.11 Tracking of reference roll model responses.

Fig.12 Tracking of reference yaw model responses.

Fig.13 Obtaining ‘‘flight data” for offline adaptation process.

Next, the model reference response tracking ability of the GS‘‘a(chǎn)daptive”roll and yaw controllers are tested.The simulation structure is similar to that in Fig.9,except that the‘‘a(chǎn)daptive controller” part in Fig.9 is replaced by GS ‘‘a(chǎn)daptive”controllers. The same random inputs and states used in Section 4.1 (Fig.10) are set for the controller-aircraft system in Fig.15. The final simulation results are as in Figs. 14 and 15. Again the reference model response (red line), the simulated aircraft’s response with GS ‘‘a(chǎn)daptive” control switched on (blue line) and off (green line) are plotted.

The simulation results in Figs.14 and 15 are very similar to those in Figs.11 and 12:the system control equipped with the GS ‘‘a(chǎn)daptive” controller can drive the aircraft’s roll and yaw responses closer to the ‘‘ideal” values obtained from the responses of reference models, just like PMRAC. This shows the GS‘‘a(chǎn)daptive”controllers retain the model reference tracking abilities, thus the effectiveness of the offline adaptation method is verified.

4.3. Decoupling of roll and yaw controls

To better demonstrate the adaptive controller’s ability to decouple yaw and roll controls, as discussed in Section 3.3, a coordinated turn maneuver is simulated in this part. This maneuver is chosen because in a coordinated turn,yaw control has to be used in conjunction with roll control to change the aircraft’s heading direction, and the ‘‘disturbance” effect of yaw control on roll control is quite prominent.

Fig.14 Tracking of reference roll model responses (using GS ‘‘a(chǎn)daptive” controller).

Fig.15 Tracking of reference yaw model responses (using GS ‘‘a(chǎn)daptive” controller).

Fig.16 Step roll command tracking.

Fig.17 V-44 quad-tiltrotor in flight tests.

A positive δaQis needed to maintain a positiverin the operation. However if δt＞0, this also produces a positive disturbance roll momentLd(explained in Section 2.5) and the more δtincreases,the largerLdis(as seen in Eq.(13)).If uncompensated, thisLdcauses the aircraft’s roll moment to be ‘‘larger than expected”, and leads φ to become much larger than the commanded value. When adaptive control is applied, the controller’s reference model response tracking ability allows it to adjust δaappropriately to cancel out theLd,so that the decoupling is realized.

The reason for choosing δt=0 rad and δt=0.5 rad is:when δt=0,theoretically there is zero roll disturbance caused by yaw, and when δt=0.5, the disturbance caused approximately reaches peak while a relatively constant airspeed is maintained.

The final simulated results with the adaptive controller turned off/on are shown in Figs. 16(a) and (b).

It can be seen in Fig.16(a),when adaptive control is turned off,the roll responses φ with δt=0.5 and δt=0 are quite different:the former exhibits an overshoot of about 0.2 rad,while the latter’s overshoot is zero. In contrast, in Fig.16(b), with the adaptive controller turned on, the system responses with δt=0.5 and δt=0 are nearly the same. This shows that unwanted roll moments are effectively cancelled out, and decoupling of yaw and roll controls is achieved.

5. Flight tests

5.1. Subject tiltrotor

The tiltrotor used in flight tests is a modified V-44(a.k.a.Ripmax Transition VTOL)RC model aircraft(Fig.17).The original flight controller in the V-44 is replaced by a Pixhawk flight controller and supplementary avionics. The entire flight control program is written using Pixhawk PX4 Support v2.1,provided by Mathworks. Aircraft parameters are shown in Table 2.

5.2. Flight data acquisition for offline parameter adaptation

Table 2 Aircraft parameters.

5.3. Reference model response tracking

Similar to the simulation work in Section 4.2, tests are conducted to verify the reference model tracking ability of the GS‘‘a(chǎn)daptive”controllers in real flight.And the same four reference models as shown in Table 1 are used for roll and yaw control.

One flight test is first conducted with the adaptive controller turned off to get the aircraft’s roll and yaw responses.After that,three flights are conducted with the adaptive control turned on,one for Reference model 1 (roll), one for Reference model 3(yaw)and one for Reference model 2(roll)plus Reference model 4(yaw).The purpose of using Reference model 2 and Reference model 4 together in one flight is to check system’s ability of handling roll and yaw control simultaneously.

Fig.18 Command input in flight tests.

Fig.19 Airspeed V and pitch angle θ in one of the tests.

Fig.20 Reference model and aircraft p and r responses with GS ‘‘a(chǎn)daptive” controller turned off.

The final results of flight tests are shown in Figs.20 and 21.Like the simulation work in Section 4.2, the reference model responses calculated from δacmdand δaQcmdare presented as‘‘idealpandr”and set as the standard to be matched(red line in Figs. 20 and 21). The aircraft’spandrresponses with the adaptive controller turned off (green line), and on (blue line)are shown in Figs. 20 and 21 separately.

In Figs. 20 and 21, it can be seen that when adaptive controller is switched off,the aircraft’spandrresponses are quite different from the‘‘ideal”pandrresponses.In contrast,when the adaptive controller is turned on, the aircraft’spandrresponses become very close to the ‘‘ideal” responses. This means roll and yaw reference model response tracking ability is achieved, and the simulation work in Section 4.2 is confirmed.

By comparing the roll and yaw responses of model 2 and model 4 with the responses in model 1 and model 3 in Fig.21, it can be seen that, in the situation of the roll and yaw control running simultaneously in the same flight, the adaptive controller can still perform effectively.

5.4. Coordinated turn maneuver

The work here is very similar to the simulations in Section 4.3,but the coordinated turn maneuvers are conducted in real flights to further test yaw-roll control decoupling ability of the adaptive controller.

The final results with the GS ‘‘a(chǎn)daptive” controller turned off/on are shown in Figs. 22 (a) and (b).

It can be seen in Fig.22(a) that, when the adaptive controller is turned off, the peak value difference in roll response φ responses with δt=0 and δt=0.5 is about 0.25 rad. Meanwhile in Fig.22(b),where the adaptive control is turned on,the peak value difference between the two response curves is about 0.12 rad, which is much smaller than in Fig.22(a). This confirms the work in the simulations in Section 4.3,and the effectiveness of yaw and roll control decoupling.

One phenomena should be noticed here. By comparing the flight test result in Fig.22(a) with the simulation work in Fig.16(a), it can be seen that the performance of coordinated turn in real flight is much worse than in simulation.This is reasonable,because real flight conditions are much more complex and harsh than simulated ones. However, even in these‘‘harsher” conditions, when the adaptive controller is turned on, the aircraft’s coordinated turn performance in part Fig.22(b) does not differ much from the simulated work in Fig.22(b).This further shows the effectiveness of the adaptive controller and the great potential of the application of modern control technology in flight control.

6. Conclusions and future work

Fig.21 Reference model and aircraft p and r responses with GS ‘‘a(chǎn)daptive” controller turned on.

Fig.22 In-flight step roll command tracking.

(1) An attempt is made to apply modern control technology to the roll and yaw control of a rudderless quad-tiltrotor UAV. A dynamics model of the aircraft is built, a predictor-based adaptive controller is designed to track different roll and yaw reference model responses. An RBF-NN and offline adaptation method and a GS‘‘a(chǎn)daptive” controller are used to cope with the nonlinearitiesinthecontrols,anddrivetheaircraft’sroll andyawresponsestothoseofpre-setreferencesystems.Simulationsareconductedtoverifythereferencemodel responsetrackingandyaw-rolldecouplingabilityofthe adaptivecontroller,aswellasthevalidityoftheoffline adaptationmethod.Finalflighttestsconfirmthatthe predictor-based adaptive controller can effectively improvetherollandyawcontrolqualityforthetested quad-tiltrotormodel.

(2)Fortiltrotors,becauseoftheircombinedfixed-wing androtorstructure,especiallythecomplextilttransitionmechanism,manychallengeswillbeencountered intheareaofflightcontrol.Torisetothechallenge,moderncontroltechnologycanplayanimportant role,andmoreeffortsshouldbeputintoitsapplication.