1 Introduction
This paper studies the control design for a novel application, an automatic crosswind stabilization system for a buoyantlyassisted aerial vehicle as shown in Figure 1. This novel aerial vehicle design has several advantages over conventional rotarywing and fixedwing aircraft, which may be summarized as follows. First, such a vehicle has the ability to takeoff and land in very short distances based on the lift generated from the buoyant gas. Second, the vehicle has the ability to carry heavy payloads compared to its own weight; a major part of the lift comes from buoyancy and a small part from aerodynamics. Third, while helicopters depend on fossil fuel, the weight of which can reach 50% of the helicopter’s payload, this novel vehicle design is powered by electric engines. The vehicle’s electric batteries can be recharged in different ways, including solar panels mounted on the wing. These advantages, among others, make the novel vehicle design particularly suitable to serve remote areas, most of which have no roads or runways.
However, despite of its advantages, this novel design results in large surface areas, which make the air vehicle more susceptible to gusts and crosswinds, especially when operating at low speeds during takeoff, landing, and taxiing on the ground. In particular, the crosswind causes a roll motion of the wing, which prohibits safe takeoff, landing, and taxiing. To solve this problem, an embedded crosswind stabilization system is developed in hardware, which estimates the roll angle of the vehicle’s wing using inertial measurement units (IMUs), and then operates electric motors at the vehicle’s wingtips to produce a torque counteracting the roll motion caused by the wind. The initial, basic design of the automatic crosswind stabilization system involved the use of a proportionalintegralderivative (PID) controller; however, the response of the system under a welltuned PID controller is too slow, which prohibits the system from counteracting wind gusts. This sets the stage for the research carried out in this paper.
The objective of this paper is to design and implement an advanced control strategy for the automatic crosswind stabilization of such novel hybrid aerial vehicles that satisfies the following important properties:

fast stabilization of the roll motion of the aerial vehicle under crosswinds and wind gusts, as compared to the basic PID controller;

minimal changes of the current hardware of the automatic crosswind stabilization system;

implementation on an embedded, microcontrollerbased device requiring a computationally efficient control algorithm.
In this paper, we provide a computationally feasible, estimationbased model predictive control strategy for this novel application, which reduces the response time of a welltuned PID strategy by around , without requiring any hardware changes of the automatic crosswind stabilization system.
We first show that a good approach to achieve objective (i) is to measure the torque caused by the wind, and then use a feedforward compensation to directly compensate for the wind disturbance. However, it is difficult to mount highaccuracy wind sensors on the vehicle’s wing and to satisfy objective (ii), we replace wind sensors with a wind torque estimator that relies on inertial measurements. We provide two methods for designing the wind torque estimator: the poleplacement technique and Kalman filter survey_Kalman . We then combine the estimator with an improved control algorithm. In particular, we utilize a model predictive control (MPC) strategy predictive1 scenario_mpc , which optimizes the system’s future behavior while accounting (directly or indirectly) for system and actuation constraints.
This paper adds to the wide range of applications of estimators and MPC by utilizing them in this interesting new application. While most of the papers in the literature focus on wind disturbance estimation and rejection for smallsize aerial vehicles, as discussed in detail in the next section, we consider a largesize hybrid aerial vehicle and carry out our experimental study on an elevenmeter wing. Moreover, we present and compare, for the considered application, two MPC strategies: a constrained MPC strategy, which requires solving a quadratic program at each sampling instant, and an unconstrained MPC strategy, which reduces to an explicit control law evaluated at each sampling instant. The two MPC strategies are computationally feasible, satisfying objective (iii) above. Finally, experimental results show the effectiveness of the proposed control strategy.
This paper is organized as follows. Section 2 summarizes the related work. In Section 3, we present the dynamic model of the system and discuss how the parameters of the model change depending on the operating conditions. In Section 4, we present the basic PID control strategy and illustrate its drawbacks. In Section 5, we present our proposed, estimationbased, predictive control strategy. In Section 6, we present experimental results that verify the effectiveness of the proposed strategy and compare the performance of the different control approaches, including unconstrained MPC and constrained MPC. In Section 7, we conclude the paper.
2 Related Work
In this paper, we study the controller design problem for an automatic crosswind stabilization system for a novel hybrid aerial vehicle. Multiple different approaches have been proposed for dealing with this problem in varying contexts. In turbulence_mitigation_MAV , a turbulence mitigation system is developed for microaerial vehicles (MAVs) using a biomimetically birdinspired design. Using multiple pressure probes mounted on the wings of the vehicle, the authors of turbulence_mitigation_MAV anticipated the properties of the oncoming wind, and were able to use this data in a simple feedforward controller to mitigate the effect of turbulence on the MAV. Although this method of anticipating the wind over the vehicles’ wings using pressure probes shows promise for smallsize aerial vehicles, it does not extend well to large hybrid aerial vehicles. In particular, there are many difficulties associated with mounting reliable pressure probes on our large wing. For instance, these pressure probes do not work well for the low speeds that our vehicle is taxiing at, and thus, the calibration of the sensor must be very accurate and its resolution must be high. As another example, the boundary layer affecting the air flow around the hybrid aerial vehicle is considerably big, which significantly influences the accuracy of these probes when they are mounted on the wing. Thus, we replace the wind sensors in our study with a wind torque estimator. In roll_control , active wing morphing, together with a simple gain feedback controller, was used for a teninch MAV to create a washout effect to aid in damping the windinduced roll motion. While this method uses a unique control surface to achieve stability rather than anticipating the effect of the disturbances, this technique cannot be efficiently implemented on a large hybrid aerial vehicle due to the prohibitive size and rotational inertia of the airframe.
Kalman filter techniques have been used for wind estimation for other types of vehicles, as well as for crop protection and building wind load design; see multiple_model_filters real_time_muav_imu . In these papers, Kalman filters together with various data mining techniques were used to estimate the wind model, model parameters, and the current state of the wind. Nevertheless, our approach is fundamentally simpler in the sense that we directly estimate the wind torque affecting the roll motion, and rely on only inertial measurements from interoceptive sensors. Also, unlike multiple_model_filters and crosswind_rejection which contain only simulation results, we provide experimental results on an elevenmeter hybrid aerial vehicle. In wind_speed_turbines , a Kalman filter is used to estimate the wind torque on the rotor of a wind turbine for control purposes. Similar to wind_speed_turbines , we estimate the wind torque on the hybrid airship, but what is unique, compared to wind_speed_turbines , is how we use this for roll stabilization and humanintheloop control of the vehicle during ground taxiing, landing and takeoff. Kalman filtering and model predictive control have been also used for attitude control and trajectory tracking under aerodynamic disturbances as investigated in real_time_muav_imu , coac_wind_disturbances small_uav_flight_in_wind . These studies are limited though in that they mainly deal with small systems such as quadrotors or MAVs, and it is not known how these techniques would scale to larger systems with higher rotational inertias and larger wind disturbances.
While most of the available results in the literature are for inflight crosswind stabilization of smallsize aerial vehicles, our main contribution is to design and implement an efficient automatic crosswind stabilization system for largesize hybrid aerial vehicles during taxiing, and verify its effectiveness experimentally on an elevenmeter hybrid aerial vehicle.
3 System Model
We consider the hybrid aerial vehicle in Figure 1, and develop an automatic crosswind stabilization system for the scenario when the vehicle is taxiing on the ground. The vehicle considered in this paper is a manned vehicle. During taxiing, the pilot of the vehicle has complete control over the yaw and pitch motions of the aerial vehicle. Through experimental observations, we found that this is not the case for the roll motion under crosswinds. Hence, in this paper, we assume that the pilot is able to keep the pitch angle at zero during taxiing and study the roll motion of the wing. In particular, we design and implement an automatic crosswind stabilization system that detects the roll angle of the wing and then operates motors at the wingtips to counteract the wind torque.
The objective of this section is to derive a simplified model of the roll motion of the wing when the wing is subject to crosswinds or wind gusts and under the assumption of zero pitch angle. This simplified model will be utilized in our simulations and in Section 5 for the design of the estimator and the model predictive controller.
Although the wing of the considered aerial vehicle is nonrigid, we observed through experiments that the wind torque affecting the wing is not high enough to cause a noticable buckling of the inflated wing. Also, for practical reasons, we seek a simple, finitedimensional model upon which we can build our estimator and controller. To that end, we assume that the inflated wing is rigid, and utilize Newton’s second law for rotational motions to derive an approximate dynamic model of the roll motion of the wing. Let be the roll angle in radians, be the total torque of the motors, and be the wind torque affecting the roll motion. It is evident that
(1) 
where is the time index, is the inertia of the wing, is its stiffness (spring constant), and is its damping coefficient. To counteract the wind torque , two DC motors are mounted on the tips of the wing to produce the counteracting torque . Both motors are kept at an idle speed and then based on the direction of the wind, one of the motors is sped up to provide a thrust and a counteracting torque. In our proposed setup, each motor is operated in only one direction to avoid the delay associated with reversing the rotational direction of the motor. Hence, we can write
(2) 
where are the thrusts of the two motors, and is the wing span. The thrust of the motor is given by Kumar :
(3) 
where is a constant, and is the th motor speed. The dynamics of the motors themselves can be represented by dorf :
(4) 
where is the motor’s current, is its input voltage, is its torque constant, , are frictional constants, is its armature windings resistance, and is its armature windings inductance. The term represents the back electromotive force (emf). In practice, the inputs to the above dynamical system are the motor commands sent from a microcontroller to the motor controllers through communication cables, and so it is reasonable to assume that
(5) 
where is the voltage command sent to to the motor , and is a communication delay. To sum up, the dynamics of the roll motion of the wing with the motors on the wingtips can be represented by equations (1)(5). Having two states representing the roll angle dynamics (1) and four states representing the two motors, the above dynamics represent a sixthorder nonlinear dynamic system with a time delay.
The dynamics of the above system can be further simplified as follows. Since the inertia of the two motors is much lower than the inertia of the wing, we neglect the dynamics of the motors in (3); instead, we assume a longer time delay , , between sending the motor commands and the change of the motor torque to the desired value. Since (2) and (3) are algebraic equations that can be simply accounted for in the microcontroller code, we assume without loss of generality (w.l.o.g.) that we directly control the motor torque, and hence the dynamics of the system can be represented by:
(6) 
where is the commanded, total torque of the two motors. Given the physical limits of the motors, the motor torque must satisfy
(7) 
where is the maximum torque that can be provided by the motors, and since the two motors in our setup are identical. Equations (6) and (7) represent a simplified dynamical model of the system: a linear model with time delay and input constraints. Note that, however, the parameters , , in (6) are, in general, timevarying parameters since they depend on the operating conditions of the aircraft such as the ambient temperature and pressure, and the pressure of the wing gas. Since these parameters are changing slowly, we assume that these parameters are fixed during the taxiing, and identify them through simple experimental tests just prior to the start of the vehicle’s operation.
4 Basic Control Design
The objective of the automatic crosswind stabilization system is to keep the wing at the horizontal position despite crosswinds and wind gusts; that is, to keep the roll angle of the wing close to zero under the wind torque disturbance . At first glance, the problem looks simple. Hence, in this section, we first explore via simulations on MATLAB Simulink the use of a standard PID controller. By analyzing the drawbacks of the standard design, we then propose in the next section an advanced control strategy that significantly improves the control performance.
For our simulations in this paper, we assume kg., N.m/rad, N.m.s/rad and s, which are nominal values experimentally identified for our elevenmeter wing prototype with air filling. We also use a wind torque model that relates the wind speed and direction to the resulting wind torque affecting the roll motion, and this model is identified based on computational fluid dynamics (CFD) simulations on the elevenmeter wing.
For the basic PID control algorithm in our simulations, we select the sampling rate to be Hz, and the parameters of the PID controller to be , , and . The controller gains were selected based on wellknown insights into the role of each term of the PID controller and based on extensive trialanderror. We also add a simple, firstorder, lowpass filter before the Dterm to avoid noise amplification. Moreover, we use an antiwindup mechanism for the Iterm to avoid the windup problem caused by this term if the motor torque is saturated, see (7). In particular, we saturate the output of the Iterm if it hits the actuation limits. Figure 2 shows the displacement of the wingtip, given by where is the wingspan, for the PID controller under a steady wind of .
Using a standard PID controller, the system is slow and has large oscillations; the settling time is higher than . For sudden changes in the wind speed, emulating wind gusts, the controller is not able to return the wing to the horizontal position before the wind changes again. For the wind speed profile in Figure 3(a), the corresponding system response is shown in Figure 3(b).
The inadequate performance of the PID controller is due to the input time delay, as well as the relatively large inertia of the system, causing a slow system response against disturbances. This highlights the need for an advanced control strategy that can sense the wind disturbance and correct for it in a feedforward fashion, and that can take into account the time delay and the actuation limits.
5 Proposed Control Design
In the previous section, we have shown that standard feedback control strategies lead to a slow, oscillatory response against constant crosswinds and wind gusts. When a wind disturbance is applied, the controller first detects the deviation of the angle from the horizontal position, and only then starts changing the motor torque to counteract the wind torque, further delayed by the inherent time delay. The response of the system can be made significantly faster if the wind torque is measured, and directly counteracted by changing the motor torque. This can be achieved by using a wind sensor and a feedforward compensation technique. The crosswind speed can be measured by standard wind sensors such as multihole pilot tubes or normal pilot tubes with small vanes multi_hole_tubes ; pilot_tubes2 .
However, implementing a reliable wind sensor for the large hybrid aerial vehicle has several difficulties. First, the wind sensors do not work well for the low speeds that our vehicle is taxiing at, and thus, the calibration of the sensor must be very accurate and its resolution must be high. Second, the boundary layer affecting the air flow around the hybrid aerial vehicle is considerably bigger than the one of a samesize, fixedwing aerial vehicle. This large effect on the airflow influences the accuracy of the measurements of the wind sensors. Third, the cost of accurate wind sensors is high. Therefore, in our work, we replace the wind sensor with a wind torque estimator.
Based on the above discussion, in this paper, we use an estimationbased control strategy as shown in Figure 4. The estimator receives the roll angle measurement and the motor commands, and outputs an estimation of the wind torque that is then subtracted from the controller output to directly compensate for the wind torque disturbance. In the next subsection, we show how to systematically design the wind torque estimator, and then, in Section 5.2, we present a model predictive controller that takes into account the input time delay and the system’s actuation limits.
5.1 Estimator Design
Consider the discretetime state space model (3) of the system, and suppose that the system is affected by an unknown, constant wind torque . For the short time scale of our stabilization problem, the constant wind assumption in the prediction model is standard and called the ‘persistence technique’ for shortterm wind forecasting, see wind_forecast1 . The objective of this subsection is to design an estimator that can successfully estimate the unknown wind torque from the roll angle measurements and the process model. Since by assumption, then (3) can be rewritten in the following augmented form:
(9) 
where , and the matrices of the augmented system are:
We have verified that for the nominal system parameters the pair is observable. Hence, we can design an observer that successfully estimates the states of the augmented system (5.1) from the output measurements. In particular, we can successfully estimate the unknown value of from the output measurements. To that end, let be the estimated value of the augmented state in (5.1). We use the standard estimator equation astrom
(10) 
where is the estimator gain to be chosen. From (5.1) and (10), the dynamics of the estimation error are:
(11) 
where . In the following part, we discuss two methods for the proper selection of .
First, a simple approach is to use pole placement to assign the eigenvalues of
to lie anywhere in the open unit disc, and this is always possible since is observable. This ensures that as , as desired. In our simulations and experiments, we selected the locations of the desired eigenvalues to be , , .Second, we use the Kalman filter equations to design survey_Kalman ; kalman_ss2 . Consider again the augmented system model (5.1) and suppose that, in practice, noise signals are added to both the state and output equations of (5.1), resulting in:
(12) 
where and
are zero mean white noise vectors with covariance matrices
and , respectively. Note that for our problem, is a positive semidefinite matrix, while is a positive scalar. Since the parameters of the wing system are slowlychanging as discussed in Section 3, we assume that the matrices , , are constant during operation, and use a steadystate Kalman filter kalman_ss2 kalman_ss3 . In particular, we first solve the algebraic Riccati Equation (ARE) to find the steadystate value of the state covariance matrix :(13) 
Then, we use the obtained matrix to calculate the Kalman filter gain as follows:
(14) 
The obtained gain can be then used in the observer equation (10) (). The Kalman filter is optimal in the sense that if the noise in (5.1) has the distribution as assumed, then the Kalman filter minimizes the mean squared error of the estimated state vector. Even if the noise is not Gaussian, the Kalman filter is the best linear estimator for minimizing the mean squared error of the estimated state vector survey_Kalman ; kalman_ss1 . In practice, we typically do not know the covariance matrices and . Hence, and are tuned to get a good performance of the estimator. For our aircraft, we select to be a diagonal matrix with the following elements on the diagonal: , , and , respectively, and .
5.2 Model Predictive Controller
As shown in Figure 4, the estimated wind torque in the previous subsection is utilized to directly compensate for the effect of the wind torque through a standard feedforward compensation technique. Assuming perfect cancellation of the wind torque using the feedforward compensation, the discretetime dynamics of the system become:
(15) 
The objective of this subsection is to present a model predictive control (MPC) strategy for (5.2) to optimize the system’s behavior over a prediction horizon of samples. We start by deriving a prediction model for the system (5.2). Note that (5.2) has an input time delay, and consequently, one should start the prediction at a time step beyond the value of the time delay since these are the outputs that can be affected by the current and future inputs. To that end, it is straightforward to show:
At the current sampling instant , we suppose that we have , , , available and, consequently, we can compute . Let . Then, it can be shown that for all ,
(16) 
Collecting these equations in a matrix form, we have
(17) 
where is the predicted output vector, is the vector of motor control torques to be selected, and are known matrices.
After calculating the matrices and at a sampling instant , the constrained MPC problem is
(18) 
where , , , are constant vectors, is a positive semidefinite matrix, and is a positive definite matrix. The optimization problem is equivalent to:
(19) 
where and . The cost function of (5.2) is quadratic and the constraints are linear in , and so (5.2) is a quadratic program that can be efficiently solved online at each sampling instant boyd . The first element of the obtained is applied to the system, and the same procedure is repeated at each sampling instant.
The tuning parameters of the MPC are the matrices and the prediction horizon . We select as diagonal matrices with positive entries on the diagonal, representing the weight of each output/input in the cost function. The higher the weight, the higher priority is given to minimizing the corresponding variable. We select the last element of (and hence the terminal cost) higher than the other elements of to impose to be close to as desired. As a good initial guess, we select , where is a reasonablyselected, desired number of samples within which we need the system to reach the desired output. In particular, recalling that the sampling interval s, we have , and we select .
Figure 6 shows the simulated system response under the proposed estimationbased, constrained MPC strategy for the wind speed profile in Figure 3(a). Each time the wind speed suddenly changes, the proposed strategy succeeds to stabilize the wing to the horizontal position in about , which is around times faster than the PID control strategy in Figure 3(b). The proposed strategy also has lower oscillations than the PID controller. On the other hand, the constrained MPC strategy requires solving an online optimization problem at each sampling instant, and so it is more computationally demanding than the PID controller.
Hence, we next present an estimationbased, unconstrained MPC strategy that still achieves significant improvements over the basic PID design, and does not require solving online optimization problems. This approach is useful if only limited online computational resources are available in the automatic crosswind stabilization system. To that end, we ignore the hard inequality constraints on and in the constrained MPC problem, and indirectly account for these constraints by proper selection of the cost matrices and . By ignoring the constraints of (5.2), it is evident that the solution of the optimization problem (5.2) is a closedform control law given by:
(20) 
Note that is always invertible since is positive definite and is positive semidefinite by selection. The matrix is typically of high dimension; however, it is only a function of the system matrices and the cost matrices, and consequently, its inverse can be precomputed offline to further reduce the online computational burden. Similar to the constrained MPC approach, only the first element of is applied to the system, and then the computation of in (20) is repeated at each sampling instant.
6 Experiments
In this section, we briefly describe the setup of the experiments, and then present experimental results that verify the effectiveness of the proposed estimationbased MPC approach for the considered application. Finally, we provide a detailed discussion of the results.
6.1 Experimental Setup and Results
We present experimental results on an meter prototype of a hybrid aerial vehicle. The experiments were carried out at the MarsDome of the University of Toronto Institute for Aerospace Studies (UTIAS). In these experiments, the constant wind speeds were simulated with weights attached with a rope at the vehicle’s wingtips, while wind gusts were simulated with large, sudden changes in the weight disturbance that can be achieved by putting on and off weights during the experiment. Two inertial measurement units (IMUs) are placed close to the vehicle’s wingtips to measure the roll angle of the wing, and we emphasize that these are the only available sensors in our experiments, see Figure 7.
We start with verifying the effectiveness of our proposed wind torque estimator. To that end, we carried our experiments in which we put on and off known weights at the left wingtip, and then detected the corresponding torques affecting the wing using the proposed estimator. Figure 8(a) shows the wind torque estimation for putting on and off lb of weight, which corresponds to torque changes of approximately N.m. For this experiment, the estimator gain is designed based on the pole placement technique discussed in Section 5.1. One can see that the estimator successfully detects the fast changes in the weights, and it estimates the disturbance torque value almost accurately. However, one can see that the estimated torque is very noisy due to the noise in the measured angle, and consequently, we always use a filtered value of the estimated torque (red curve). Figure 8(b) shows the wind torque estimation for putting on and off lb of weight, causing disturbance torque changes of approximately N.m, where for this experiment a Kalman filter is applied with the same parameters used in our simulations in Section 5.1. The estimator successfully detects the fast changes in the weight and the filtered, estimated torque is close to accurate.
After verifying the effectiveness of the proposed wind torque estimator, we next test the proposed estimationbased, constrained MPC strategy, and compare it to a basic PID control strategy
where is the output of the PID controller at sampling instant , is the error in the roll angle at , is a filtered value of the roll angle measurement , and is the integral term given by
where is the sampling interval, and we saturate beyond the actuation limits to avoid the windup problem of the Iterm. For the Dterm, we select to be the average of the Euler approximation of the error derivative over three samples to reduce the effect of noise. That is, we use , where . Figure 9 shows the displacement of the wingtip under the two control strategies and for a constant weight of 15 lb added at time s. Based on crosswind experiments on the meter wing, the torque caused by the lb is equivalent to the effect of km/hr crosswind speed. The blue trajectory shows the system response under the basic PID controller, with the parameters , and , which are the same parameters used in the simulation in Figure 2. From Figure 9, one can see that the response with the PID controller is very oscillatory and, for around s, the wing does not settle back to zero displacement, i.e., the PID control strategy fails to stabilize the wing to the horizontal position. On the other hand, the magenta trajectory is the result of our proposed control strategy, and one can see that it has much less oscillations. It settles in around 6 s and has a significantly lower peak value. A demo video is available at http://tiny.cc/EMPCHybridAircraft.
We second test the ability of the counter gust system to respond fast to wind gusts, which are mainly characterized by fast changes in the wind speed. To simulate this in the MarsDome, we put on and off the lb weight approximately every 25 s, and the counter gust system must handle this change in weight. Figures 10 and 11 show the wingtip displacement and the motor commands, respectively, for the same two controllers as in Figure 9, particularly the basic PID controller (blue) and the proposed estimationbased, constrained MPC (magenta).
One can see from Figure 10 that the proposed estimationbased strategy was able to stabilize the wing to zero displacement quickly (in around s) every time the weight changes, and the response has significantly less oscillations and lower peak values. Also, from Figure 11, the motor commands in the proposed estimationbased MPC strategy reach suitable values faster every time the weight changes, and they oscillate less resulting in a smoother response.
We next test the unconstrained MPC strategy, and compare it to the constrained MPC strategy. Figures 12 and 13 compare the estimationbased, constrained MPC response with the estimationbased, unconstrained MPC response for constant and square weight disturbances, respectively. The proposed estimationbased, unconstrained MPC strategy results in a significantly better performance than the PID controller in Figures 9, 10 and in a slightly worse performance than the estimationbased, constrained MPC strategy. On the other hand, the unconstrained MPC strategy does not require solving online optimization problems and, consequently, is less computationally demanding than the constrained MPC strategy.
6.2 Discussion of the Results
In this paper, we have presented several experimental results that show the effectiveness of our proposed estimationbased MPC strategy and the significant improvement it achieves over the basic PID control strategy. The results also verify the effectiveness of the different controller/estimator design approaches discussed in this paper and provide comparisons between these approaches. In particular, we have shown the following important points through our experimental results:

Our proposed estimationbased MPC strategy speeds up the response of the basic PID controller around ten times, as shown in Figure 9, without the need for adding any sensors. Hence, unlike the basic PID control strategy, the estimationbased MPC strategy enables the counter gust system to handle fastchanging disturbances (wind gusts) as shown in Figure 10. This faster response is expected since, in our approach, we estimate the wind torque and directly counteract its effect through the feedforward compensation instead of waiting until the wind torque affects the roll angle to start counteracting its effect using the feedback controller.

For the considered application, both the pole placement technique and the Kalman filter technique can be efficiently used to design the estimator gain, and both techniques achieve good estimation results as shown in Figure 8.

For the considered application, the unconstrained MPC strategy achieves an excellent balance between optimality in performance and computational efficiency. In particular, it achieves lower peak values, less oscillations, and a significantly smaller settling time compared to the basic PID controller and a slightly worse performance than the constrained MPC strategy, as shown in Figures 12 and 13. On the other hand, the unconstrained MPC strategy is much less computationally demanding than the constrained MPC strategy since it does not require solving online optimization problems.
For future research, we first plan to carry out extensive field tests, in which the vehicle is taxiing subject to crosswinds and wind gusts, instead of using weights attached to the wingtips for emulating wind torques. This will require buying more powerful motors that can counteract high wind speeds. Second, we are currently working on the control of a meter, fully autonomous version of the hybrid aerial vehicle. Third, for the current estimator design, we assume that the wind torque is constant during our short prediction horizon (around 3 s), and then counteract this estimated, constant wind torque using feedforward compensation. It may be possible to create a wind prediction model that observes the past and current wind estimates from the Kalman filter as well as other environmental variables (e.g., ambient temperature and pressure), and then uses these values to predict the wind behavior over the prediction horizon. The predicted wind disturbance over the prediction horizon should be incorporated in the MPC for more accurate predictions and hence better overall response. Since the future wind predictions are typically probabilistic, stochastic MPC stochastic_mpc or scenariobased MPC scenario_mpc may be needed.
7 Conclusions
In this paper, we designed and implemented an automatic crosswind stabilization system for a novel hybrid aerial vehicle design, which enables takeoff and landing in short distances without the need for runways. Despite its advantages, the vehicle’s wing has a large surface area, which makes it very susceptible to crosswinds and wind gusts. The proposed automatic crosswind stabilization system detects the roll angle deviation, and then controls motors mounted at the vehicle’s wingtips to counteract the wind effect. We have shown that standard control strategies such as a PID controller do not achieve fast stabilization of the wing and, consequently, cannot counteract wind gusts. Given the difficulties of mounting reliable wind sensors on the considered wing, we used instead a wind torque estimator and provided two methods for the design of the estimator, pole placement and a Kalman filter. Then, we proposed an estimationbased predictive control strategy and showed through experimental results that the proposed strategy reduces the response time of the system by around  compared to the PID controller without adding sensors or changing the hardware of the crosswind stabilization system.
References
 (1) D. Simon. Kalman filtering with state constraints: a survey of linear and nonlinear algorithms. IET Control Theory and Applications, vol. 4(8), pp. 13031318, 2010.
 (2) A. Aswani, H. Gonzalez, S. S. Sastry, C. Tomlin. Provably safe and robust learningbased model predictive control. Automatica, vol. 49, pp. 12161226, 2013.
 (3) J. B. Rawlings, D. Q. Mayne. Model Predictive Control: Theory and Design, Nob Hill Publishing, LLC, 2009.
 (4) M. Cannon, Q. Cheng, B. Kouvaritakis, S. V. Raković. Stochastic tube MPC with state estimation. Automatica, vol. 48(3), pp. 536541, 2012.
 (5) M. Maiworm, T. Bäthge, R. Findeisen. Scenariobased model predictive control: recursive feasibility and stability. IFACPapersOnLine, vol. 48(8), pp. 5056, 2015.
 (6) A. Mohamed, M. Abdulrahim, S. Watkins, R. Clothier. Development and flight testing of a turbulence mitigation system for micro air vehicles. Journal of Field Robotics, vol. 33(28), pp. 639660, 2015.
 (7) H. Garcia, M. Abdulrahim, R. Lind. Roll control for a micro air vehicle using active wing morphing. AIAA Guidance, Navigation, and Control Conference and Exhibit, 2003.
 (8) A. Sharifi, H. Nobahari. Multiple model filters applied to wind model estimation for a fixed wing UAV. International Conference on Mechanical and Aerospace Engineering, 2016, pp. 109115.
 (9) S. Shanmuganathan, P. Sallis. Data mining methods to generate severe wind gust models. Atmosphere, vol. 5(1), pp. 6080, 2014.
 (10) N. Smith. Optimal lateral guidance for automatic landing of a lightweight high altitude long endurance unmanned aerial system with crosswind rejection. Master’s thesis, University of Kansas, 2016.
 (11) P. Neumann, M. Bartholmai. Realtime wind estimation on a micro unmanned aerial vehicle using its inertial measurement unit. Sensors and Actuators A: Physical, vol. 235, pp. 300310, 2015.
 (12) D. Bourlis, J.A.M. Bleijs. A wind speed estimation method using adaptive Kalman filtering for a variable speed stall regulated wind turbine. IEEE International Conference on Probabilistic Methods Applied to Power Systems, 2010, pp. 8994.
 (13) K. Alexis, G. Nikolakopoulos, A. Tzes. Experimental constrained optimal attitude control of a quadrotor subject to wind disturbances. International Journal of Control, Automation, and Systems, vol. 12(6), pp. 12891302, 2014.
 (14) T. Cheviron, F. Plestan, A. Chriette. A robust guidance and control scheme of an autonomous scale helicopter in presence of wind gusts. International Journal of Control, vol. 82, pp. 22062220, 2014.
 (15) K. Alexis, G. Nikolakopoulos, A. Tzes. On trajectory tracking model predictive control of an unmanned quadrotor helicopter subject to aerodynamic disturbances. Asian Journal of Control, vol. 16, pp. 209224, 2014.
 (16) D. Weibel, D. Lawrence, S. Palo. Small unmanned aerial system attitude estimation for flight in wind. Journal of Guidance, Control, and Dynamics, vol. 38(7), pp. 13001305, 2015.
 (17) V. Kumar, N. Michael. Opportunities and challenges with autonomous micro aerial vehicles. The International Journal of Robotics Research, vol. 31(11), pp. 12791291, 2012.
 (18) R. C. Dorf, R. H. Bishop. Modern Control Systems, 11th edition, Pearson Education, Inc., 2008.
 (19) H. Lee, I. Shinder, J. Wright, M. R. Moldover. Application of ANFIS method to the nonnulling calibration of multihole pitot tube, Procedia Engineering, vol. 79, pp. 125132, 2014.
 (20) A. Barlas. Active aerodynamic load control on wind turbines: aeroservoelastic modeling and wind tunnel experiments. Doctoral Thesis, TU Delft, 2011.
 (21) W.Y. Chang. A Literature review of wind forecasting methods. Journal of Power and Energy Engineering, vol. 2, pp. 161168, 2014.
 (22) K. J. Åström, B. Wittenmark. ComputerControlled Systems: Theory and Design, 3rd edition, Dover Publications, 2013.
 (23) D. Banfield, Don, A. P. Inofrsoll, C. L. Keppenne. A steadystate Kalman filter for assimilating data from a single polar orbiting satellite. Journal of the Atmospheric Sciences, vol. 52(6), pp. 737753, 1995.
 (24) L. Kleeman. Understanding and applying Kalman filtering. The Second Workshop on Perceptive Systems, 1996.
 (25) M. S. Persson, I. Sharf. Steadystate invariant Kalman filter for attitude and imbalance estimation of a neutrallybuoyant airship. AIAA Guidance, Navigation, and Control Conference, 2013.
 (26) S. Boyd, L. Vandenberghe. Convex Optimization, Cambridge University Press, 2004.
Comments
There are no comments yet.