Simple Adaptive Control for Aircraft Control Surface Failures Abderrazak I. Belkharraz, Member, IEEE LaGuardia Community College Kenneth Sobel, Senior Member, IEEE The City College of New York Keywords: adaptive control, fault tolerant control, aircraft flight control This work extends the so-called simple adaptive control approach to direct model reference adaptive control of multi-input multi-output systems to include loss of control effectiveness failures. It is proven that all signals are bounded for loss of control effectiveness failures during a bounded input disturbance. A state space approach is introduced for computing the feedforward compensator that is required by the stability result. The adaptive algorithm is applied to a three input model of the linearized lateral dynamics of the F/A-18 aircraft. Simulation results are obtained with single, double, and triple control effectiveness failures of 88% during the occurrence of a lateral gust. These results show that the adaptive controller exhibits improved model following as compared to a fixed gain eigenstructure assignment controller. An earlier version of this paper was presented at the 2004 Guidance, Navigation, and Control Conference. This work was partially supported by a grant from The City University of New York Collaborative Incentive Research Program and by a grant from The City University of New York PSC-CUNY Research Award Program. Corresponding author’s address: Prof. Kenneth Sobel, Department of Electrical Engineering, The City College of New York, New York, NY 10031. email: sobel@ccny.cuny.edu I Introduction Aircraft flight control systems are designed with extensive redundancy to ensure a low probability of failure. During recent years, however, several aircraft have experienced major control system failures. These have caused an increased interest in fault tolerant flight control systems. The objective of a fault tolerant flight control system is to control and safely land the aircraft in case of severely damaged or inoperable control surfaces. One of the approaches to fault tolerant control is active control. An active fault tolerant control system has to either reconfigure or adapt the controller in response to the failure. Typical design methods include multiple model, switching, and tuning designs; adaptive designs; and fault detection and diagnosis designs. Adaptive failure accomodation designs have simpler control structures and do not rely on knowledge of the actuator failures. Direct adaptive designs use the system response error to achieve desired performance. Boskovic and Mehra [1] propose an indirect adaptive scheme for loss of aircraft control effectiveness failures. However, the authors assume that the system state is available and that the number of control effectors exceeds the number of variables that can be affected by the controls. Bodson and Groszkiewicz [2] consider adaptive schemes for aircraft control effectiveness failure accomodation. However, the authors require the system state to be available and the aircraft to be minimum phase and relative degree one. Tao et. al. [3] use a direct adaptive scheme to accomodate locked in place actuator failures. However, the authors require the aircraft to be minimum phase. Sobel et. al. [5] proposed simple adaptive control algorithms for multiple input multiple output systems. Later, Barkana [6] inserted a feedforward compensator around the plant so that the augmented system is almost strictly positive real (ASPR). Kaufman et. al. [7] summarize stability results that show all 2 signals in the adaptive system are bounded and that the augmented error is asymptotically vanishing if the augmented plant is ASPR. Morse and Ossman [4] use simple adaptive control to accomodate failures that can modelled by a zero column in the linearized aircraft control derivative matrix. However, the authors choose the outputs to be linear combinations of the states instead of the usual aircraft sensor measurements. Furthermore, no proof of stability is shown. The present paper extends the work described in Kaufman et. al. [7] to include loss of control effectiveness failures. It is proven that all signals are bounded for loss of control effectiveness failures during a bounded input disturbance. The percentage loss of control effectiveness is unknown and may be arbitrarily close to a complete loss subject to the satisfaction of the sufficient conditions for stability. A state space approach is introduced for computing the feedforward compensator which guarantees that the augmented plant is ASPR by using the MATLAB R LMI and Optimization toolboxes. The adaptive algorithm is applied to a three input model of the linearized lateral dynamics of the F/A-18 aircraft. Simulation results are obtained with single, double,and triple control effectiveness failures of 88% during the occurrence of a lateral gust. The maximum failure is chosen to be 88% because this is the largest failure for which a feedforward compensator could be designed that ensures that the aircraft plus feedforward is ASPR. The simulation results show that the adaptive controller exhibits improved model following as compared to a fixed gain eigenstructure assignment controller. II Problem Statement Let (Ti , Ti+1 ), i = 0, 1, . . . , q0 , with q0 finite and T0 = 0, be the time intervals on which the control surface failure pattern is fixed. That is, control surfaces only fail at time Ti , i = 1, . . . , q0 . 3 Then, the plant on the interval (Ti , Ti+1 ), i = 0, 1, . . . , q0 is described by ẋ p (t) = A p x p (t) + Bip u p (t) + d(t) Pi : y p (t) = C p x p (t) (1) where A p ∈ ℜn p ×n p , B p ∈ ℜn p ×m , C p ∈ ℜq×n p , Bip = B p αi , B p ≡ B0p , and d(t) denotes a bounded input disturbance vector. The value of the bound on the disturbance is unknown and not required by the adaptive algorithm. αi = diag{αi1 , αi2 , . . . , αim }; i = 1, 2, . . . , q0 I; (2) i=0 0 < αik < 1, if kth control surface fails, k = 1, . . . , m αi = 1, k if kth surface does not fail, k = 1, . . . , m Here the failure times are Ti , i = 1, . . . , q0 ; which control surfaces fail at Ti , i = 1, . . . , q0 is unknown; the amount of the loss of effectiveness at Ti given by αi , where αik ∈ (0, 1) is unknown. Furthermore, once a control surface fails it may fail again later with a different amount of loss of effectiveness. The plant is augmented with a single fixed feedforward compensator of the form H −1 (s) = K −1 f (3) 1 + τs A state space realization of H −1 (s) is given by ṡ p (t) = As s p (t) + Bs u p (t) (4) h p (t) = Cs s p (t) (5) The augmented plant is then given by ẋap (t) = Aap xap (t) + Bap αi u p (t) + D(t) a Pi : yap (t) = Cap xap (t) 4 (6) xap where d(t) x p a A p 0 a i B p αi a ; C p = ( C C ); D(t) = ; B p α = ; A p = = p s 0 Bs 0 As sp and where xap (t) ∈ ℜn ; u p (t) ∈ ℜm ; and yap (t) ∈ ℜq . Then the augmented output to be controlled is yap (t) = y p (t) + h p (t) (7) where h p (t) is the inverse Laplace transform of H −1 (s). The control objective is to design an adaptive control signal u p (t) such that all signals in the closed loop system are bounded and the augmented plant output yap (t) tracks the output of a reference model given by ẋm (t) = Am xm (t) + Bm um (t) (8) ym (t) = Cm xm (t) (9) where Am ∈ ℜnm ×nm , Bm ∈ ℜnm ×mm , Cm ∈ ℜq×nm . Remark 1 We remark that the order of the plant may be much greater than the order of the reference model. III The General Tracking Problem We summarize the general tracking problem for a known plant. These results are shown in Kaufman et al. [7]. Let the input command um (t) be the output of an unknown command generating system of the form v̇m (t) = Av vm (t) (10) um (t) = Cv vm (t) (11) 5 Thus, the allowable reference model commands um (t) are those commands that are the solution to some unknown set of linear ordinary differential equations. This includes steps, ramps, polynomials, exponentials, and sinusoids. Define the ideal trajectories (xap )∗ (t), such that, if the augmented plant could reach and move along them, its output would perfectly track the output of the reference model. That is, the ideal trajectories are targets that the augmented plant tries to reach or at least be close to, in order to have bounded tracking errors. Mathematically, (yap )∗ = Cap (xap )∗ = Cm xm = ym (12) The ideal trajectories are defined as (xap )∗ (t) = X11 xm (t) + X12 um (t) (13) and the ideal control is defined as u∗p (t) = K̃x xm (t) + K̃u um (t) (14) Substituting (xap )∗ (t) from (13) into (12) gives a condition for existence of the ideal target trajectories. Cap X11 xm (t) +Cap X12 um (t) = Cap X11 xm +Cap X12Cv vm = Cm xm (15) (16) or Cap X11 = Cm (17) Cap X12Cv = 0 Solutions for X11 and X12 in (17) exist, in general, because this system consists of more variables than equations. This implies the existence of some bounded trajectories in the x ap space that the plant needs to attain for perfect tracking. 6 IV Adaptive Control A Algorithm The adaptive control law is given by u p (t) = Kx xm (t) + Ku um (t) + Ke [ym (t) − yap (t)] (18) where Kx , Ku , and Ke are adaptive gains. The adaptive gains are concatenated into matrix K(t) defined as K(t) = [Ke (t), Kx (t), Ku (t)] (19) The concatenated gain K(t) is defined as the sum of a proportional gain K p (t) and an integral gain KI (t), each of which is adapted as follows: K(t) = K p (t) + KI (t) (20) K p (t) = eay (t)rT (t)T¯ (21) K̇I (t) = eay (t)rT (t)T − σKI (22) T (t), uTm (t)] rT (t) = [(eay (t))T , xm (23) eay (t) = ym (t) − yap (t) (24) where T , T¯are time invariant weighting matrices. The σ term in Eq. (22) was originally proposed by Ioannou and Kokotovic [8] and it is used to avoid divergence of the integral gains in the presence of disturbances. 7 Define ex = (xap )∗ − xap . The error equation is derived in appendix A and is given by ėx = Ac ex − Bap αi [K − K̃i ]r − F(t) (25) where Ac = Aap − Bap αi K̃eiCap (26) K̃i = [K̃ei , K̃xi , K̃ui ] (27) F(t) = (Aap X11 − X11 Am + Bap αi K̃xi )xm + (Aap X12Cv − X11 BmCv + Bap αi K̃uiCv − X12Cv Av )vm + D V (28) Stability Analysis In this section, we present two theorems describing the stability of the adaptive controller. Theorem 1 shows that positive definite quadratic Lyapunov functions exist such that their derivatives are negative definite whenever ex or KI is sufficiently large. Theorem 2 describes a sufficient condition for the boundedness of the Lyapunov functions at the failure instants. Theorem 1 If 1. the augmented plant Pia is ASPR , i.e., there exist positive definite matrices Pi , Qi , and a gain matrix K̃ei such that: Pi (Aap − Bap αi K̃eiCap ) + (Aap − Bap αi K̃eiCap )T Pi = −Qi < 0 (29) (Cap )T = Pi Bap αi (30) 8 2. dim[yap (t)] ≤ dim[u p (t)] and T and T¯positive definite symmetric 3. the disturbances are bounded then a positive definite quadratic Lyapunov function Vi (ex , KI ) can be selected such that its derivative V̇i is negative definite whenever ex or KI is sufficiently large for t ∈ (Ti , Ti+1 ), i = 0, 1, 2, . . . , q0 . Proof: see appendix B Remark 2 We present an overview of the proof of Theorem 1. Choose the positive definite quadratic Lyapunov function as Vi (ex , KI ) = eTx Pi ex + tr[(KI − K̃i )T −1 (KI − K̃i )T ] (31) then, the derivative of (31) along the trajectories of Eq. (25) is a a T a T T V̇i = −eTx Qi ex − 2(eay )T eay (eay )T T¯ e ey − 2(ey ) ey [xm T¯ x xm + um T¯ u um ] −2σtr[(KI − K̃i )T −1 (KI − K̃i )T ] −2σtr[(KI − K̃i )T −1 K̃iT ] − 2eTx Pi F(t) (32) Hence, we have V̇i < −β1 kex k2 − β2 keay k4 − β3 keay k2 (β4 kxm k2 + β5 kum k2 ) − β6 kKI − K̃i k2 + β7 kKI − K̃i k + β8 kex k (33) where β1 , β2 , β3 , β4 , β5 , β6 , β7 and β8 are positive constants. We conclude that whenever kex k or kKI − K̃i k increase beyond some bound, then the negative quadratic and quartic terms in (33) will become dominant, and thus V̇i (ex , KI ) becomes negative. The quadratic form of Vi (ex , KI ) then guarantees that ex and KI are bounded. 9 Theorem 2 Let Vi (t) ≡ Vi (ex (t), KI (t)), i = 1, 2, . . . , q0 . The Lyapunov functions at the failure instants given by Vi (Ti ), i = 1, 2, . . . , q0 are bounded. Proof: see appendix C VI Feedforward Design We propose a method which uses the MATLAB R LMI toolbox [9] and the MATLAB R Optimization toolbox [10] to design a robust feedforward compensator. Each augmented LTI plant P ia , i = 0, 1, . . . , q0 must be ASPR using the same feedforward. However, we do not know in advance which failures will occur. Suppose there are j0 possible failure configurations P j , j = 1, . . . , j0 . Each possible failure configuration P j corresponds to one surface failure, or two simultaneous surface failures, or three simultaneous surface failures, and so on. The P j ’s are parameterized by a range of possible control effectiveness loss. Thus, the P j ’s are different from the Pi ’s in (1) because the Pi ’s uniquely define a failure whereas the P j ’s define a possible parameterized failure configuration. Also, let (A p , B0p ,C p ) correspond to the unfailed plant. Consider an affine parameter dependent representation of P j given by ẋ p = A p x p + B pj (σ)u p P j (σ) : y p = Cpx p 10 (34) k j where B p (σ) = B0p − ∑k=1 σk Bk , j = 1, . . . , j0 and where k j is the number of surfaces of P j (σ) which j j j j j simultaneously fail. Here Bk describes the failure of one surface and σk describes the corresponding range of control effectiveness loss. The following two lemmas due to Barkana [6] show that the feedforward compensator for a strictly proper plant has a simple form. Lemma 1 Let G(s) be any m × m strictly proper transfer matrix of arbitrary McMillan degree. G(s) is not necessarily stable or minimum phase. Let K f be a nonsingular constant output feedback matrix such that the closed loop transfer matrix Gcl (s) = [I + G(s)K f ]−1 G(s) (35) is asymptotically stable. Then, the augmented open loop transfer matrix Ga (s) = G(s) + K −1 f (36) is ASPR. Lemma 2 Let G(s) be any m × m strictly proper transfer matrix of arbitrary McMillan degree. G(s) is not necessarily stable or minimum phase. Let H(s) = K f (1 + τs) be a stabilizing closed loop compensator of G(s). Then, the augmented plant Ga (s) = G(s) + H −1 (s) = G(s) + K −1 f (37) 1 + τs is ASPR. 11 We want to minimize the norm of K −1 f so that yap (t) = y p (t) + h p (t) ≈ y p (t) (38) while ensuring that K f robustly stabilizes every P j (σ), j = 1, . . . , j0 Let P̃ j (σ) be P j (σ) with feedback K f , j = 1, . . . , j0 Mathematically min k K −1 kF s.t P̃ j (σ) is robustly stable; j = 1, . . . , j0 f Kf (39) We ensure that the P̃ j (σ)’s are robustly stable by using program “pdlstab” from the MATLAB R LMI toolbox [9] and program “fmincon” from the MATLAB R Optimization toolbox [10]. At each iteration of the optimization, we call “pdlstab” and require that the quantity t min be negative. Here tmin is computed by “pdlstab” where tmin < 0 ensures the existence of a parameter dependent Lyapunov function for P̃ j (σ). With tmin < 0, every LTI plant which belongs to P̃ j (σ), j = 1, . . . , j0 is stable. We remark that on each interval t ∈ (Ti , Ti+1 ) we have an LTI plant Pi , i = 1, . . . , q0 . Let P̃i be the LTI plant Pi with feedback K. Then, every LTI plant P̃i is one of the P̃ j (σ) for some values of j and σ. Thus, every LTI plant P̃i is stable. Then, it follows that every augmented plant Pia , i = 1, . . . , q0 is ASPR. A flowchart of the design process is shown in Figure 1. First, “fmincon” and “pdlstab” are used to find the maximum failure such that K f is a robust stabilizing gain that minimizes k K −1 f k. However, it is possible that K f may be robust stabilizing for a larger maximum failure. This is because the function to be minimized and the constraints are quite complicated. Also, there is no guarantee of a global rather than a local minimum. Therefore, the gain K f from the optimization is tested for larger maximum failures by using “pdlstab” without optimization. This is done repeatedly until the maximum failure is found for 12 which “pdlstab” can ensure robust stability when using the gain K f . Then, the feedforward gain is K −1 f . Remark 3 The system in (34) with feedback H(s) = K f (1 + τs) is given by j ẋ p = Acl (σ)x p (40) where j Acl (σ) = A p − B pj (σ)[I + τK f C p B pj (σ)]−1 K f C p (I + τA p ) (41) j which we observe is not affine in B p (σ) as required by program “pdlstab”. However, the closed loop system matrix with constant feedback K f is given by j Acl (σ) = A p − B pj (σ)K f C p (42) j which is affine in B p (σ). This is the reason that we use “pdlstab” to compute a stabilizing constant feedback K f instead of the dynamic feedback K f (1 + τs). Remark 4 The implementation of a constant feedforward K −1 f yields an algebraic loop which causes a problem in the computer simulation. Therefore, we first compute a constant gain K f which stabilizes j Acl (σ) in (42) for all j and σ by using “pdlstab” and “fmincon”. Then, we implement the feedforward H −1 (s) = K −1 f /(1 + τs) where τ is sufficiently small. We show in Theorem 3 that such a τ exists so that j H(s) = K f (1 + τs) stabilizes Acl (σ) in (41) for all j and σ. Therefore, H −1 (s) = K −1 f /(1 + τs) satisfies lemma 2. j Theorem 3 Suppose Acl (σ) in (42) is stable with a non-defective modal matrix for all j and σ. Then, j there exists τ sufficiently small such that Acl (σ) in (41) is stable for all j and σ. Proof: see appendix D 13 VII Example A Aircraft and Reference Model Consider the linearized lateral dynamics of the F/A-18A aircraft. The rigid body states are lateral velocity (v), yaw rate (r), roll rate (p), and bank angle (φ). The control surface deflections are asymmetric trailing edge flaps (δte ), ailerons (δa ), and rudder (δr ). The measurements are sideslip angle (β), washed out yaw rate (rwo ), and roll rate(p). The unfailed aircraft is described by the triple (A p , B0p ,C p ) where the matrices A p , B0p , and C p are shown in appendix E. The block diagram of the adaptive control system is shown in Figure 2. Let T j , j = 0, 1, . . . , m0 , be the instants at which a bounded input disturbance in the form of a lateral gust vg (t) occurs, where vg (t) is defined as shown in MIL-F-8785-C [11] t − Tj < 0 0 π(t−T j ) vm vg (t) = 1 − cos 0 ≤ t − Tj ≤ T 2 T vm t − Tj > T (43) The lateral gust magnitude is given by vm . The gust length T is chosen to be the inverse of the natural frequency of the closed loop complex eigenvalue pair of the unfailed aircraft. Here the dutch roll eigenvalues are λdr = −2 ± j2 so that T = 1/ωn = 0.3536 sec. The state equations for the lateral dynamics are given by [12] v̇ = Yv v − rU0 + gφ +Yδa δa +Yδr δr −Yv vg ṙ = (44) Nv0 v + Nr0 r + N p0 p + Nδ0 a δa + Nδ0 r δr − Nv0 vg − ṗ = Lv0 v + Lr0 r + L0p p + Lδ0 a δa + Lδ0 r δr − Lv0 vg − 14 Nr0g U0 Lr0 g U0 v̇g v̇g (45) (46) where U0 is the trim velocity. Nr0g and Lr0 g are given by [12] Nr0g Lr0 g Nr 2 −1 Ixz = Nr 1 − Ix Iz 2 −1 Ixz Ixz = Nr 1 − Ix Ix Iz Ixz = Nr0 − Lr0 Iz (47) (48) (49) The inertias Ix , Iz , and Ixz were not available for the F/A-18 aircraft. Therefore, the values for the F16 aircraft [13] are used which are given by Ix = 9496, Iz = 63100, and Ixz = 982. The trim velocity is 646.9 f t/sec. Using these values, the gust derivative coefficients in (47)-(48) are Nr0g = −0.2578 and Lr0 g = −0.0266. The open loop aircraft has a lightly damped dutch roll mode with damping 0.13, a roll subsidence mode at -2.76, and a spiral mode at -0.0002. An output feedback gain matrix Keig is designed using eigenstructure assignment for the unfailed aircraft by assigning the dutch roll mode to have a damping of 0.707 and a natural frequency of 2.83 rad/sec. The roll subsidence mode is assigned to -4. The desired eigenstructure and the feedback gain matrix Keig are shown in Table 1. The closed loop spiral mode is unstable at 0.009. However, this instability is within the limits of MIL-F-8785C [11] specifications which specifies that the spiral minimum time to double amplitude is 12 seconds. Here the default value for TOL is used when computing the achievable eigenvectors using MATLAB R function pinv. Only rudder and aileron are used because these two control surfaces are sufficient to achieve desired handling qualities for the unfailed aircraft. However, the adaptive controller will use all three control surfaces in order to accommodate failures. The closed loop time responses of the reference model to a one degree initial sideslip (not shown) exhibit both desirable damping and sideslip to bank decoupling. 15 The unfailed aircraft with feedback Keig is not ASPR because there are two transmission zeros at z1 = 0.2028 × 10−13 ; and z2 = −0.0002 × 10−13 one of which is non-minimum phase. The feedforward which satisfies the ASPR condition will be designed for the unfailed aircraft with feedback Keig . This closed loop unfailed aircraft is described by the triple (A p − B0p KeigC p , B0p ,C p ). The reference model is chosen to be the same triple so that Am = A p − B0p KeigC p , Bm = B0p , and Cm = C p . Thus, when there are no failures the reference model is the aircraft with eigenstructure assignment feedback Keig . Next, it is shown how the reference model input um is computed. um = −Keig (ym − yc ) = −Keig β rwo p − pc (50) (51) K (1, 1) K (1, 2) K (1, 3) eig eig eig = − Keig (2, 1) Keig (2, 2) Keig (2, 3) β rwo p − pc Keig (1, 3) pc = −Keig ym + Keig (2, 3) (52) (53) The reference model in Figure 2 is implemented using (53) where p c is the pilot roll rate command. B Feedforward Computation A robust feedforward is designed so that the aircraft with feedback Keig and feedforward K −1 f is ASPR for all possible failures. The affine parameter dependent representation of the aircraft with feedback gain Keig 16 j j is described by the triple (A p − B p (σ)KeigC p , B p (σ),C p ). It can be shown that the closed loop system for j j (A p − B p (σ)KeigC p , B p (σ),C p ) with feedback gain K f is given by j ẋ p = Acl (σ)x p (54) where j Acl (σ) = A p − B pj (σ)[Keig + K f ]C p (55) The feedback gain K f is computed so that (54) is stable for all possible failures. The optimization in (39) is performed by using program “fmincon” from the MATLAB R Optimization Toolbox [10] and program “pdlstab” from the MATLAB R LMI toolbox [9]. The Frobenius norm of K −1 f is minimized by using “fmincon” as shown in (39). The possible failures are chosen to be the loss of control effectiveness of any one, two, or three surfaces. However, it is not known which control surfaces will fail nor the amount of control effectiveness loss. Experience indicates that the optimization should be initialized with a nonsingular stabilizing gain. The initial value for K f , denoted by Kinitial , is computed by using eigenstructure assignment for (A p − B0p KeigC p , B0p ,C p ). The initial gain Kinitial uses all three control surfaces because the adaptive controller will use all three control surfaces. The desired eigenvalues and eigenvectors are the same that were used to compute Keig . An attempt was made to obtain Kinitial using eigenstructure assignment, but the unassigned eigenvalues were unstable. The (5 × 3) control distribution matrix B0p has singular values 35.3966, 18.6783, and 0.2101. The small singular value, which corresponds to a direction of weak control, causes some of the feedback gains to be on the order of 10 4 . The physical explanation is that even with three control surfaces we can still only generate yawing moment and rolling moment. To alleviate this design problem, the pseudo-control strategy of Sobel and Lallman 17 [14] is used. The columns of the new control distribution matrix B̃0p are chosen to be the two singular vectors corresponding to the two larger singular values and eigenstructure assignment is used to compute a (2 × 3) feedback gain matrix. Then, this (2 × 3) feedback gain is mapped back to the original matrix B0p to obtain the (3 × 3) gain Kinitial . The designer must be careful when computing the achievable dutch roll eigenvectors because the pseudo-inverse of an ill-conditioned matrix is required. This ill-conditioned matrix represents the basis vectors for the achievable dutch roll eigenvector subspace modified by deleting the rows corresponding to unspecified entries in the desired dutch roll eigenvectors. Its pseudoinverse is used to calculate the least squares projection of the desired dutch roll eigenvectors onto the achievable dutch roll eigenvector subspace [15]. The ill-conditioning is caused by the deletion of the rows corresponding to unspecified entries in the desired dutch roll eigenvectors. Prior to the row deletion the four singular values were all unity indicating that the basis was orthonormal. The ill-conditioned matrix has singular values given by 1.0000, 1.0000, 0.9998, and 0.0028. The achievable eigenvectors were computed using MATLAB R function “pinv” with TOL=0.01. This has the effect of discarding the singular value at 0.0028 in the pseudo-inverse computation. Press et. al. [16] state that it is very often better to obtain the least squares solution by using a singular value decomposition and zeroing the small singular values. If the singular value at 0.0028 is not discarded, then the initial gain matrix has a norm on the order of 10 −10 . This would not be a useful initial gain for an optimization whose objective is to minimize the norm of the inverse gain. The initial gain Kinitial is shown in Table 2. A feedforward is designed which allows up to 50% loss of control effectiveness in any one, two, or three control surfaces. The optimization is over K f as shown in (39). The optimization converges with kK −1 f kF = 0.34415. The optimization is restarted with the final gain as the new initial gain. Then, the 18 optimization exceeds the maximum allowable function evaluations with kK −1 f kF = 0.0099724. The optimization is again restarted with this last gain as the new initial gain. The optimization converges with the optimal robust stabilizing gain Koptimal shown in Table 2. The Frobenius norm has been reduced from −1 −1 k Kinitial kF = 1.2 × 105 to k Koptimal kF = 0.000629 and robust stability is ensured because tmin is negative. Next, “pdlstab” is called using the optimal gain Koptimal with increasing values of loss of control effectiveness. Here there is no optimization. It was found that “pdlstab” could ensure robust stability for single, double, or triple failures of up to 88% loss of control effectiveness using the optimal gain Koptimal . Hence, −1 Theorem 1 ensures that the aircraft with feedback Keig and feedforward Koptimal is ASPR for all single, double, and triple control surface failures of up to 88% loss of control effectiveness. However, dynamics are added so that the plant is augmented with a dynamic feedforward compensator H −1 (s) = −1 Koptimal 1+τs where τ = 10−16 . This eliminates the algebraic loop associated with the use of a constant feedforward compensator. The maximum failure is chosen to be 88% because ”pdlstab” could not ensure robust stability for a maximum failure of 89%. Finally, it is noted that the example used version 2.2 of the Optimization toolbox and version 1.0.8 of the LMI Toolbox on a SUN Ultra 10 workstation. Solutions which use earlier versions of the toolboxes or which use 32 bit computing may produce different answers for Koptimal . This just indicates that roundoff error determines to which local minimum the optimization converges. It is not important whether a global minimum is obtained because only a robust stabilizing gain whose inverse has a small norm is needed. C Simulation Results The pilot command pc is chosen to be a roll rate step of 10 deg/s. The initial conditions on both the aircraft and reference model are zero. A lateral gust with magnitude of 15 f t/sec is applied at T1 = 2 sec 19 for t ∈ [2, 15]. A control surface failure occurs at t = 6 sec. Therefore, the control surface failure occurs during a gust condition. A non-adaptive simulation of the aircraft is performed with the fixed eigenstructure assignment feedback gain Keig for a single 88% failure. This corresponds to the block diagram in Figure 2 with Kx = 0, Ke = 0, Ku = I, and H −1 (s) omitted. The aircraft and reference model responses (not shown) exhibit poor model following in sideslip angle, yaw rate, and roll rate. The roll rate and yaw rate exhibit model following errors of 400% and 300%, respectively at 40 seconds when an aileron failure occurs at 6 seconds. The sideslip angle is negative for a positive roll rate which is an undesirable response. The aircraft should always sideslip into the turn. Next, simulations are performed of the adaptive system of Figure 2. The adaptive gains are initialized to be Kx = 0, Ke = 0, and Ku = I. This choice of initial gains causes the reference model and the unfailed aircraft to be identical with identical inputs. Thus, the time responses of the reference model and the unfailed aircraft will be identical. The weighting matrices T and T¯are chosen to be T = T¯= 1000I and σ = 0.01. The aircraft responses (not shown) exhibit perfect tracking in sideslip angle, yaw rate, and roll rate. Simulations are performed for an 88% simultaneous failure of any two control surfaces. The sideslip angle, yaw rate, and roll rate for the non-adaptive and adaptive controllers are shown in Figure 3. The unfailed aircraft responses (solid line) and reference model responses (solid line) are identical. The nonadaptive controller exhibits an aircraft roll rate that is approximately 20% of the reference model roll rate at 40 seconds when a simultaneous aileron and rudder failure (dotted line) occurs. The aircraft roll rate is approximately 25% of the reference model roll rate at 40 seconds when a simultaneous aileron and trailing edge flap failure (dashed line) occurs. Furthermore, the sideslip is negative for positive roll rate which is 20 an undesirable response. In both of these failure cases, the reduced roll rate causes the yaw rate to be reduced by approximately 67%. Finally, the aircraft tracking errors in roll rate and yaw rate are small for a simultaneous rudder and trailing edge flap failure (dashdot line). However, the tracking errors are increasing with time. Furthermore, the aircraft sideslip angle is twice that of the reference model at 40 seconds. This indicates degraded performance because the sideslip angle should be small for a coordinated turn. Observe from Figure 3 that the adaptive controller achieves perfect tracking in sideslip angle, yaw rate, and roll rate. Simulations are performed for an 88% simultaneous failure of all three control surfaces. The time responses using the non-adaptive controller (not shown) are similar to the simultaneous aileron and rudder failure given by the dotted line in Figure 3. The time responses using the adaptive controller (not shown) achieve perfect tracking in sideslip angle, yaw rate, and roll rate. VIII Conclusions A model reference adaptive controller for multi-input multi-output systems has been extended to include loss of control effectiveness failures. It has been proven that all signals in the adaptive system are bounded during the occurrence of a bounded disturbance. A state space approach was introduced for computing the feedforward compensator that is required by the stability result. The adaptive controller was applied to the linearized lateral dynamics of the F/A-18 aircraft when one, two or three control surfaces fail with an 88% loss of control effectiveness during a lateral gust condition. Simulations show that the adaptive controller exhibits significantly improved model following as compared to a fixed gain eigenstructure assignment controller. 21 Acknowledgment The authors thank Dr. I. Barkana of Kulicke & Soffa Industries and Dr. J.E. Piou of MIT Lincoln Laboratory for their contributions to theorems 2 and 3, respectively. Appendix A: Error Equation Define ex = (xap )∗ − xap . Then, ėx = (ẋap )∗ − ẋap (A.1) Add and substract Aap (xap )∗ ėx = (ẋap )∗ − ẋap + Aap (xap )∗ − Aap (xap )∗ (A.2) ėx = X11 (ẋm ) + X12 (u̇m ) − Aap xap − Bap αi u p − D + Aap (xap )∗ − Aap (xap )∗ (A.3) ėx = X11 Am xm + X11 Bm um + X12Cv Av vm − Aap xap − Bap αi u p − D + Aap (xap )∗ − Aap (xap )∗ (A.4) Recall u p = Kr, K = KI + K p and obtain ėx = Aap ex − Bap αi Kr − Aap (xap )∗ + X11 Am xm + X11 Bm um + X12Cv Av vm − D (A.5) ėx = Aap ex − Bap αi Kr − Aap (X11 xm + X12 um ) + X11 Am xm + X11 Bm um + X12Cv Av vm − D (A.6) ėx = Aap ex − Bap αi Kr − (Aap X11 − X11 Am )xm − (Aap X12 − X11 Bm )um + X12Cv Av vm − D (A.7) Add and substract Bap αi K̃i r ėx = Aap ex − Bap αi Kr + Bap αi K̃i r − Aap (X11 − X11 Am )xm − (Aap X12 − X11 Bm )um + X12Cv Av vm − Bap αi K̃i r − D (A.8) 22 Use K̃i from Eq. (27) to obtain ėx = Aap ex − Bap αi [K − K̃i ]r − (Aap X11 − X11 Am + Bap αi K̃xi )xm − (Aap X12 − X11 Bm + Bap αi K̃ui )um + X12Cv Av vm − Bap αi K̃ei eay − D (A.9) Use um = Cv vm and eay = ym − yap = (yap )∗ − yap = Cap (xap )∗ −Cap xap = Cap ex . ėx = Aap ex − Bap αi [K − K̃i ]r − (Aap X11 − X11 Am + Bap αi K̃xi )xm − (Aap X12 − X11 Bm + Bap αi K̃ui )Cv vm + X12Cv Av vm − Bap αi K̃eiCap ex − D (A.10) ėx = (Aap − Bap αi K̃eiCap )ex − Bap αi [K − K̃i ]r − (Aap X11 − X11 Am + Bap αi K̃xi )xm − (Aap X12Cv − X11 BmCv + Bap αi K̃uiCv − X12Cv Av )vm − D (A.11) Finally ėx = Ac ex − Bap αi [K − K̃i ]r − F(t) (A.12) where Ac = Aap − Bap αi K̃eiCap (A.13) F(t) = (Aap X11 − X11 Am + Bap αi K̃xi )xm + (Aap X12Cv − X11 BmCv + Bap αi K̃uiCv − X12Cv Av )vm + D (A.14) Appendix B: Proof of Theorem 1 Consider the Lyapunov function: Vi (ex , KI ) = eTx Pi ex + tr[(KI − K̃i )T −1 (KI − K̃i )T ] (B.1) where Pi is an n × n postive definite symmetric matrix, K̃i is an unspecified matrix, T −1 is a positive definite matrix, and, “tr” denotes the trace. Note that Vi is positive definite in the state variables of the 23 adaptive system, ex (t) and KI (t). The time derivative of Vi is V̇i = ėTx Pi ex + eTx Pi ėx + tr[K̇I T −1 (KI − K̃i )T + (KI − K̃i )T −1 K̇IT ] (B.2) V̇i = ėTx Pi ex + eTx Pi ėx + 2tr[(KI − K̃i )T −1 K̇IT ] (B.3) Use Eqs. (22) and (25) to obtain V̇i = [Ac ex − Bap αi [K − K̃i ]r − F(t)]T Pi ex + eTx Pi [Ac ex − Bap αi [K − K̃i ]r − F(t)] + 2tr[(KI − K̃i )T −1 Tr(eay )T − σ(KI − K̃)T −1 KIT ] (B.4) V̇i = eTx ATc Pi ex + eTx Pi Ac ex − [Bap αi [K − K̃i ]r]T Pi ex − eTx Pi Bap αi [K − K̃i ]r − F T (t)Pi ex − eTx Pi F(t) + 2(eay )T (KI − K̃i )r − 2σtr[(KI − K̃i )T −1 KIT ] (B.5) Use KIT = KIT + K̃iT − K̃iT , eay = Cap ex , and combine terms to obtain V̇i = eTx (ATc Pi + Pi Ac )ex − 2eTx Pi Bap αi [K − K̃i ]r − 2eTx Pi F(t) + 2eTx (Cap )T KI r − 2eTx (Cap )T K̃i r − 2σtr[(KI − K̃i )T −1 (KI − K̃i )T ] − 2σtr[(KI − K̃i )T −1 K̃iT ] (B.6) V̇i = eTx (ATc Pi + Pi Ac )ex − 2eTx Pi Bap αi (KI + K p )r + 2eTx Pi Bap αi K̃i r − 2eTx Pi F(t) + 2eTx (Cap )T KI r − 2eTx (Cap )T K̃i r − 2σtr[(KI − K̃i )T −1 K̃iT ] − 2σtr[(KI − K̃i )T −1 (KI − K̃i )T ] (B.7) V̇i = eTx (ATc Pi + Pi Ac )ex − 2eTx Pi Bap αi KI r − 2eTx Pi Bap αi K p r + 2eTx Pi Bap αi K̃i r − 2eTx Pi F(t) + 2eTx (Cap )T KI r − 2eTx (Cap )T K̃i r − 2σtr[(KI − K̃i )T −1 (KI − K̃i )T ] − 2σtr[(KI − K̃i )T −1 K̃iT ] 24 (B.8) Use Eq. (30) to obtain V̇i = eTx (ATc Pi + Pi Ac )ex − 2eTx (Cap )T K p r − 2σtr[(KI − K̃i )T −1 (KI − K̃i )T ] − 2σtr[(KI − K̃i )T −1 K̃iT ] − 2eTx Pi F(t) (B.9) Use K p = eay (t)rT T¯and eay = Cap ex to obtain V̇i = eTx (ATc Pi + Pi Ac )ex − 2(eay )T eay rT T¯r − 2σtr[(KI − K̃i )T −1 (KI − K̃i )T ] −2σtr[(KI − K̃i )T −1 K̃iT ] − 2eTx Pi F(t) (B.10) T x + uT T¯u to obtain a Finally, use Eq. (29) and r T T¯r = (eay )T T¯ x m e ey + xm T¯ m u m a a T a T T V̇i = −eTx Qi ex − 2(eay )T eay (eay )T T¯ e ey − 2(ey ) ey [xm T¯ x xm + um T¯ u um ] −2σtr[(KI − K̃i )T −1 (KI − K̃i )T ] − 2σtr[(KI − K̃i )T −1 K̃iT ] − 2eTx Pi F(t) (B.11) Hence, we have V̇i < −β1 kex k2 − β2 keay k4 − β3 keay k2 (β4 kxm k2 + β5 kum k2 ) − β6 kKI − K̃i k2 + β7 kKI − K̃i k + β8 kex k (B.12) where β1 , β2 , β3 , β4 , β5 , β6 , β7 and β8 are positive constants. We conclude that whenever kex k or kKI − K̃i k increase beyond some bound, then the negative quadratic and quartic terms in (B.12) will become dominant, and thus V̇i (ex , KI ) becomes negative. The quadratic form of Vi (ex , KI ) then guarantees that ex and KI are bounded. Appendix C: Proof of Theorem 2 The plant before failure is given by ẋap = Aap xap + Bap u p + D(t) (C.1) After the finite change in Bap at t = T1 the plant becomes ẋap = Aap xap + Bap α1 u p + D(t) (C.2) 25 So at t = T1 , α1 is bounded; hence, ẋap bounded. Then, it follows that xap (T1− ) = xap (T1+ ). Since, ex (T1 ) = (xap )∗ (T1 )−xap (T1 ) it follows that ex (T1− ) = ex (T1+ ). And the adaptive gains KI = Rt 0 vr T T dt, K p = vrT T¯are bounded at t = T1 . Thus V (T1 ) = eTx P1 ex + tr[(KI − K̃1 )T −1 (KI − K̃1 )T ] is bounded. The above argument is valid for any finite number of failures. Appendix D: Proof of Theorem 3 The system in (34) with feedback H(s) = K f (1 + τs) is given by j ẋ = Acl (σ)x (D.1) where j Acl (σ) = A p − B pj (σ)[I + τK f C p B pj (σ)]−1 K f C p (I + τA p ) (D.2) A necessary and sufficient condition for the existence of the inverse in (D.2) is that the eigenvalues λ i satisfy 1 + τλi {K f C p B pj (σ)} 6= 0 (D.3) Use the matrix inversion lemma to obtain j Acl (σ) = A p − B pj (σ)[I − τK f C p (I + τB pj (σ)K f C p )−1 B pj (σ)]K f C p (I + τA p ) (D.4) Rearrange (D.4) to obtain Acl (σ) = (A p − B pj (σ)K f C p ) + [τB pj (σ)K f C p (I + τB pj (σ)K f C p )−1 B pj (σ)K f C p (I + τA p ) − B pj (σ)K f C p τA p(]D.5) j j where K f is computed so that (A p − B p (σ)K f C p ) is stable for all j and σ. Use the result from Sobel et. al. [17] that the system in (D.1) is stable if α > K2 (M)β (D.6) 26 where α = −max Re[λi (A p − B pj (σ)K f C p )] (D.7) j K2 (M) is the condition number of the modal matrix of (A p − B p (σ)K f C p ). and k τB pj (σ)K f C p (I + τB pj (σ)K f C p )−1 B pj (σ)K f C p (I + τA p ) − B pj (σ)K f C p τA p k≤ β (D.8) It is clear that β can be made small by choosing τ sufficiently small. Let αmin be the minimum of α over j, σ; and let K2max be the maximum of K2 (M) over j, σ. Then, we require β < αmin /K2max (D.9) j Hence, there exists τ sufficiently small such that (D.9) is satisfied which implies that A cl (σ) in (D.4) is a stability matrix for all j, σ. Appendix E: Aircraft State Space Model Ap 0 −0.245 −646.9 0.0285 32.189 0.00849 −0.2460 0.112 0 0 = 0.73 −2.83 0 0 −0.0256 0 0 1 0 0 0 0.5 0 0 −0.5 27 B0p Cp 0 −2.915 34.909 −0.835 −0.896 −3.26 = 13.06 13.14 4.4 0 0 0 0 0 0 1/650 0 0 0 0 = 1 0 0 −1 0 0 0 1 0 0 References [1] Boskovic, J.D. and Mehra, R.K., ”An Adaptive Scheme for Compensation of Loss of Effectiveness of Flight Control Effectors”, 40th IEEE Conference on Decision and Control, Orlando, Florida, December 2001, 2448-2453. [2] Bodson, M. and Groszkiewicz, ”Multivariable Adaptive Algorithms for Reconfigurable Flight Control”, IEEE Tran. on Control Systems Technology, Vol. 5., No. 2, 217-229, March 1997. [3] Tao, G., Chen S., Tang, X., and Joshi, S.M., ”Adaptive Control of Systems with Actuator Failures”, Springer-Verlag London, 2004. [4] Morse, W.D., and Ossman, K.A., ”Model following reconfigurable flight control system for the AFTI/F-16”, Journal of Guidance ,Control, and Dynamics, pp. 969-976, 1990. [5] Sobel, K., Kaufman, H., and Mabius, L., ”Adaptive Control for a Class of MIMO Systems”, IEEE Transactions on Aerospace, Vol. 18, 1982, 576-590. 28 [6] Barkana, I., “Parallel Feedforward and Simplified Adaptive Control” International Journal of Adaptive Control and Signal Processing, Vol. 1, No. 2, pp. 95-102, 1987. [7] Kaufman, H., Barkana, I., and Sobel, K.M., “Direct Adaptive Control Algorithms Theory and Applications”, Second Edition, Springer-Verlag New York, Inc., 1998. [8] Ioannou, P., Kokotovic, P., “Adaptive Systems with Reduced Models”, Springer-Verlag, 1983. [9] Pascal, G., Arkadi, N., Alan, L., Mahmoud, C., “LMI Control Toolbox For Use with Matlab” The Mathworks, 1995. [10] Branch, M.A., and Grace, A., “MATLAB R Optimization Toolbox User’s Guide”,Version 1.5,, The Mathworks, 1996. [11] “Military Specifications-Flying Qualities of Piloted Airplanes”, MIL-F-8785C, ASD/ENESS, Wright-Patterson AFB, Ohio. [12] McRuer, D., Ashkenas, I., Graham, D., “Robust Adaptive Control”, Princeton University Press, 1973, pp. 541-544. [13] Brian, L.S., Frank, L.L, “Aircraft Control and Simulation” John Wiley & Sons, Inc. , pp. 584, 1992. [14] Sobel, K.M., and Lallman, J.F., “Eigenstructure Assignment for the Control of Highly Augmented Aircraft” J. of Guidance ,Control, and Dynamics, Vol. 12, No. 3, May-June 1989. [15] Andry, A.N., Jr., Shapiro, E.Y., and Chung, J.C., ”Eigenstructure Assignment for Linear Systems”, IEEE Transactions on Aerospace and Electronic Systems, Vol. AES-19, No. 5, pp.711-729, September 1983. 29 [16] Press, W.H., Flannery, B.P., Teukolsky, S.A., and Vetterling, W.T., ”Numerical Recipes: The Art of Numerical Computing”, Cambridge University Press, 1986. [17] Sobel, K.M., Banda, S.S., and Yeh, H.-H., 1989, Robust control for linear systems with structured state space uncertainty, International Journal of Control, 50, 1991-2004. 30 Table 1: Desired Eigenstructure and Feedback Gain Matrix Keig Dutch Roll Roll subsidence λdr = −2 ± 2 j λroll = −4 1 + jx 0 x + j1 Keig β rwo p v -1.8768 0.4583 0.1111 δac 0 r 1.7234 -1.2 -0.0656 δrc 0 + j0 1 p 0 + j0 x φ x + jx 0 x5 Table 2: Initial and Optimal Gain Matrices Kinitial −1 kKoptimal k Koptimal β rwo p β rwo p 1.6456 -0.6527 -0.0067 2242.3 -407.34 882.47 δte 1.7811 -0.6997 -0.0072 601.74 -2184.6 8356.2 δa -0.9257 0.2858 0.003 49.048 -1792.4 -3891.6 δr 31 6.3×10−4 Use eigenstructure assignment to compute initial gain Kinitial Choose maximum failure to be any single, double, or triple failure up to maximum failure percent Call function fmincon to minimize kK −1 f k subject to tmin < 0 Reduce maximum failure percent Call function pdlstab to compute tmin No Converge? Yes Call function pdlstab with increasing values of maximum failure percent tmin < 0? Yes Increase maximum failure percent by 1% No Maximum allowable failure for K f is current value minus 1% Feedforward gain is K −1 f Figure 1: Flowchart for design of feedforward gain using MATLAB LMI and Optimization toolboxes 32 u p (t) - +i - 6 - H −1 (s) h p (t) - Aircraft y p (t) yap (t) Keig pc Keig (1, 3) Keig (2, 3) Reference Model - +i - Unfailed Aircraft Model 6 ym (t)- -? i eay (t) Keig +i Kx Ke ? Ku i - +? ? 6 Figure 2: Block Diagram of Adaptive Control System 33 xm (t) Non−Adaptive Controller Adaptive Controller Sideslip Angle 0.6 0.4 0.2 0 Yaw Rate −0.2 0 10 20 30 0.4 0.2 0 −0.2 40 8 8 6 6 Yaw Rate Sideslip Angle 0.6 4 2 0 0 10 20 30 4 2 0 20 30 40 0 10 20 30 40 0 10 20 Time(sec) 30 40 2 6 Roll Rate Roll Rate 6 10 4 0 40 0 0 10 20 Time(sec) 30 4 2 0 40 Figure 3: Sideslip angle, yaw rate, and roll rate using the non-adaptive and adaptive controllers for reference model (solid), aircraft with 88% aileron and trailing edge flap failures (dashed), aircraft with 88% rudder and trailing edge flap failures (dashdot),and aircraft with 88% aileron and rudder failures (dotted). 34