Adaptive Dynamic Inversion Control with Actuator Saturation Constraints Applied to Tracking Spacecraft Maneuvers

Monish D. Tandale and John Valasek Flight Simulation Laboratory, Aerospace Engineering Department Texas A&M University, USA

Abstract
This paper presents an adaptive control methodology for nonlinear plants, that prevents parameter drift arising from trajectory errors due to control saturation. The reference trajectory is modied on saturation, so that the modied trajectory approximates the original reference closely and can be tracked within saturation limits. The adaptive parameters are updated by the error between the plant trajectory and this modied reference. Asymptotic stability of the tracking errors between the plant trajectory and the modied reference, as well as bounded learning of the adaptive parameters, is guaranteed. A numerical example of attitude tracking for a rigid spacecraft is presented.

Introduction
Structured Adaptive Model Inversion (SAMI) [1] is based on the concepts of Feedback Linearization [2], Dynamic Inversion, and Structured Model Reference Adaptive Control (SMRAC) [3, 4]. In SAMI, dynamic inversion is used to linearize the nonlinear dynamics and solve for the control. The dynamic inversion is approximate, as the system parameters are not assumed to be modeled accurately. An adaptive control structure is wrapped around the dynamic inverter to account for the uncertainties in the system parameters [5, 6, 7]. This controller is designed to drive the error between the output of the actual plant and that the reference trajectories to zero, with prescribed error dynamics. Most dynamic systems can be represented as two sets of differential equations, an exactly known kinematic level part, and a momentum level part with uncertain system parameters. The adaptation included in this framework can be limited to only the uncertain momentum level equations. This restricts the adaptation only to a subset of the state-space, making it efcient. Adaptive control usually assumes full authority control, and lacks an adequate theoretical treatment for control in the presence of actuator saturation limits. Saturation becomes more critical for adaptive systems than non adaptive systems, since the adaptation is based on the tracking error. Assuming that the dynamics are modeled perfectly and only parametric uncertainties exist in the system, the tracking error has contributions due to the initial error conditions, parametric uncertainties, and saturation. The adaptation scheme adapts only the uncertain parameters, so the error driving

the adaptation scheme should not include the error due to saturation. Including the error component due to saturation will cause incorrect adaptation. Correct adaptation in the presence of saturation is ensured by using the concept of pseudo control hedging. The pseudo control hedging methodology has been successfully demonstrated by Johnson, E.N., Calise, A.J. et al. in a neural network based direct adaptive control law [8, 9, 10]. The difference between the calculated and the applied control effort due to saturation results in a lack of acceleration produced in the plant as compared to the demanded reference acceleration. This is called the hedging signal. If the hedge is removed from the reference, the resulting modied reference can be tracked within saturation limits. The tracking error seen will be only due to the initial error and the parametric uncertainty, and not due to saturation. Hence the controller will adapt correctly. This paper presents an application of Structured Adaptive Model inversion in alliance with Pseudo-Control Hedging, to the tracking of an attitude trajectory for a rigid spacecraft.

Mathematical formulation
Formulation of the mathematical model with the minimal parameterization of the inertia matrix The equations of motion of a rigid spacecraft can be cast into a structured form as an exactly known kinematic differential equation and a momentum level equation with uncertain parameters. The kinematic differential equation is, = T () 1 ((1 T )I33 + 2 + 2T ) T () = 4 (1) (2)

where R3 is a vector of Modied Rodrigues Parameters (MRPs) that represent the orientation of the spacecraft with respect to an inertial axis. The MRPs result in a minimal, non singular attitude description. R3 is the angular velocity vector, and is the skew symmetric vector cross product operator. 0 3 2 0 1 = 3 (3) 2 1 0 The momentum level equation is I + I = ua (4)

where I is the mass moment of inertia of the spacecraft and ua is the vector of the applied control torques. Equations1 and 4 can be manipulated to obtain the following form J () +C (, ) = PT ()u (5)

 where the matrices J (), C (,) and P() are dened as P() J () C (, ) T 1 () PT IP J T P + PT [P]IP (6) (7) (8)

The left hand side of Equation 5 can be linearly parameterized as follows J () +C (, ) = Y (, , ) (9)

where Y(,,) is a regression matrix and is the constant inertia parameter vector T I11 I22 I33 I12 I13 I23 . The construction of Y is adapted dened as from the reference [11]. It can be seen that the product of the inertia matrix and a vector can be written as I = (), where R36 is dened as 1 () 0 0 R3 0 3 2 (10)

0 2 0

0 0 3

2 1 0

3 0 1

(11)

The terms on the left hand side of Equation 5 can be written as J () = PT IP = PT (P) C (, ) = PT IPT P + PT [P]IP = PT {(PT P) + [P](P)} Combining Equations12 and 13 we have the linear minimal parameterization. J () +C (, ) T = P {(P) (PT P) + [P](P)} = Y (, , ) Formulation of the control law The attitude tracking problem can be formulated as follows. The control objective is to track an attitude trajectory in terms of the MRPs. The desired reference trajectory is assumed to be twice differentiable with respect to time. Let -r be the tracking error. Differentiating twice with respect to time and multiplying by J throughout J = J J r (15) (12) (13)

(14)

Adding (Cd +C ) + Kd on both sides, where Cd and Kd are the design matrices, J + (Cd +C (, )) + Kd = J J r + (Cd +C (, )) + Kd (16)

The RHS of Equation 16 can be written as (J +C (, )) (J r +C (, )r ) +Cd + Kd (17)

From Equation 5 and the construction of Y similar to Equation 14 the RHS of Equation 16 can be further written as PT ua Y (, , r , r ) +Cd + Kd (18)

where Y (,,r ,r ) = J ()r +C (,)r = Yr . So the control law can be now chosen as uc = PT {Yr Cd Kd } (19)

where uc is the calculated control. The above control law requires that the inertia parameters be known accurately, but they may not be known accurately in actual practice. So from the certainty equivalence principle, adaptive estimates for the iner tia parameters will be used for calculating the control. uc = PT {Yr Cd Kd } Enforcing actuator saturation limits and control hedging The calculated control is obtained from Equation 20, but it has position as well as rate saturation limits. Consider the saturation limit being enforced by a sign function. If the signal x is subjected to saturation, with maximum absolute value xm , the value of the saturated signal is given by (20)

S (x) = {

x xm sign(x)

i f |x| xm i f |x| > xm

(21)

If we use the sign function to limit the control position limits, the relation between the applied control and the calculated control has a sharp corner when | x |= xm . which leads to innite values for x. To smoothen this transition from unsaturated to saturated behavior, we use a saturation function using tanh. Tanh(x) 1 when x < 2, tanh(x) 1 when x > 2 and slope of tanh(x) = 1 at x = 0 and in the neighboring region. So, on appropriate scaling, tanh can be used to represent saturation behavior.

S (x) = tanh(

x )xm xm

(22)

To enforce the position and rate saturation limits together, the following actuator model is used. ua = S (K(S (uc ) ua )) (23)

Figure 1: Saturation functions : sign and tanh where K > 0 is an appropriate gain. Now, let be the difference between the calculated control and the applied control = uc ua From Equation 20 and Equation 24, ua = PT {Yr Cd Kd } (25) (24)

With the applied control presented in Equation 25 the closed loop attitude dynamics takes the following form J + (Cd +C ) + Kd = Yr PT (26)

where =- is the parameter estimation error. Dening e1 m and e2 m r . Here m denotes the MRP vector for the modied reference trajectory. Substituting = e1 + e2 in Equation 26 J ( 1 + e2 ) + (Cd +C )( 1 + e2 ) + Kd (e1 + e2 ) = Yr PT e e (27)

The term PT is absorbed into the reference dynamics to give the dynamics of the error between the modied reference and the original reference as e J e2 + (Cd +C ) 2 + Kd e2 = PT (28)

The hedge signal PT now acts as a disturbance input and causes the modied reference trajectory to deviate from the original reference. Tracking the original reference throughout the duration of the trajectory is impossible because of the control saturation limits. Demanding the system to track a desired reference for which the control saturates for the entire time duration is impractical. Hence for any feasible reference trajectory the control should be unsaturated for some period of time and the hedge signal should be zero. Whenever the control is not in saturation and the hedge is zero,

the original reference dynamics are recovered. If reference modication given by Equation 28 is implemented, the modied tracking error between the plant trajectory and the modied reference is e J e1 + (Cd +C ) 1 + Kd e1 = Yr (29)

Equation 28 cannot be implemented as it requires the knowledge of the true parame ters . If reference modication is implemented with the learning parameters , the modied tracking error becomes e J e1 + (Cd +C ) 1 + Kd e1 = (Yr +Ye2 ) e where Ye2 = Y (,, e2 , e2 ) = J () 2 +C (,) 2 . e (30)

Parameter update and stability analysis

where 1 is a symmetric positive denite gain matrix. Taking the derivative of the Lyapunov function along the closed loop trajectories given by Equation 30 V = 1 T 1 T 1 1 e1 J e1 + eT J e1 + eT Kd e1 + 1 2 2 (32)

which can be simplied as V 1 T 1 1 = eT [ J C ] 1 eT Cd e1 + ( T (Yr +Ye2 ) + 1 ) e e1 2 (33)

1 the rst term vanishes because [ 2 J C ] is skew symmetric. Setting the coefcient of to 0 and noting that the true parameter remains constant, the adaptive laws are obtained as follows.

= = (Yr +Ye2 )T e1 This update law renders the derivative of the lyapunov function V = T Cd e1 e1

(34)

(35)

Thus the derivative is negative semi denite. From Equation 31 and Equation 35 it can be concluded that e1 , e1 , , L , and e1 L2 . Using the standard procedure of application of Barbalats lemma, we conclude that e1 0 as t .

Parameter

Actual value 30 10 5 10 20 3 5 3 kg m2 15

Guessed values 28.5 9.5 4.75 9.5 19 2.85 4.75 2.85 kg m2 14.25

Inertia

States (t0 ) (t0 ) Control ua

Actual value [0.1, 0.09, 0.05]T [0.09, 0.07, 0.08]T Position limit [6, 6, 6]T Nm

Reference [0, 0, 0]T rad s1 [0.1, 0.1, 0.1]T rad s1 Rate limit [10, 10, 10]T N ms1

Table 1: Simulation parameters

220 200 180 |Adaptive Inertia Vector| 160 140 120 100 80 60 40 20 0 20 40 Time (sec) 60 80 100 True Parameter Learned with Hedging Learned without Hedging

Figure 2: Time histories of the norm of the inertia parameter

0.45 0.4 0.35 0.3 | err| 0.25 0.2 0.15 0.1 0.05 0 0 20 40 Time (sec) 60 80 100 Modified Trajectory With Hedging Without Hedging

Figure 3: Time histories of the norm of the errors in the MRPs with respect to the original desired reference

Numerical simulation of attitude trajectory tracking for a rigid spacecraft

This numerical example simulates tracking of an attitude trajectory for a rigid spacecraft. The spacecraft properties are obtained from [12]. The reference maneuver is selected as a sinusoidal trajectory in the MRPs along all three axes, with amplitude 0.5 and sinusoidal frequency 0.2 rad s1 . The simulation is done with small errors injected in the system parameters and initial condition, and large trajectory errors due to control saturation limits. Thus the advantages of control hedging are seen clearly. The initial conditions and control limits are listed in Table 1. From Figure 2, we see that the learned adaptive inertia vector does not drift when control hedging is applied. For the test case simulated it actually approaches the true inertia values, but this does not hold for any general reference trajectory. Also, when control hedging is not applied the adaptive parameter adapts to tracking errors due to saturation and drifts away. From Figure 3, we see that the modied reference trajectory remains bounded, as desired. The plant trajectory converges to the modied reference when control hedging is applied. In the absence of control hedging, the tracking performance deteriorates. This is because the dynamic inverter uses drifted inertia values to determine the control.

Conclusion and future work

Acknowledgment
The material is based upon work supported by NASA under award no. NCC-1-02038. Any opinions, ndings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reect the views of the National Aeronautics and Space Administration.

References
[1] Subbarao. K, Structured Adaptive Model Inversion: Theory and Applications to Trajectory Tracking for Non-Linear Dynamical Systems, PhD Dissertation, Texas A&M University, College Station, Tx, 2001. [2] Slotine, J. and Li, W., Applied Nonlinear Control, Prentice Hall, 1991. [3] Akella, M.R., Structured Adaptive Control: Theory and Applications to Trajectory Tracking in Aerospace Systems, PhD Dissertation, Texas A&M University, College Station, TX, 1999. [4] Schaub, H., Akella, M.R. and Junkins, J.L., Adaptive Realization of Linear Closed-Loop Tracking Dynamics in the Presence of Large System Model Errors, The Journal of Astronautical Sciences, Vol 48, No. 4. Oct-Dec 2000. pp 537-551 [5] Narendra, K.S. and Annaswamy, A., Stable Adaptive Systems, Prentice Hall, 1989. [6] Sastry, S. and Bodson, M., Adaptive Control: Stability, Convergence, and Robustness, Prentice-Hall, Upper Saddle River, NJ, 1989. [7] Iannou, P.A. and Sun, J., Stable and Robust Adaptive Control, Prentice Hall, Upper Saddle River, NJ, 1995.

[8] Johnson, E., Limited Authority Adaptive Flight Control, PhD dissertation, Georgia Institute of Technology, 2000. [9] Johnson, E.N., Calise, A.J. and Corban, J.E., A six degree of freedom Adaptive Flight Control Architecture for Trajectory Following, AIAA Guidance Navigation and Control Conference and Exhibit, Monterey, California, 5-8 August 2002. [10] Johnson, E.N. and Kannan, S.K., Adaptive Flight Control for an Autonomous Unmanned Helicopter, AIAA Guidance Navigation and Control Conference and Exhibit, Monterey, California, 5-8 August 2002. [11] Ahmed, J., Coppola, V.T. and Bernstein, D.S., Adaptive asymptotic tracking of spacecraft attitude motion with inertia matrix identication, AIAA Journal of Guidance Control and Dynamics, vol. 21, no. 5, Sept-Oct 1998, pp. 684-691. [12] Schaub, H., Akella, M.R. and Junkins, J.L., Adaptive Control of Nonlinear Attitude Motions Realizing Linear Closed-Loop Dynamics, Proceedings of the American Control Conference, San Diego, California, June, 1999.