Research Journal of Applied Sciences, Engineering and Technology 4(20): 3930-3936, 2012 ISSN: 2040-7467 © Maxwell Scientific Organization, 2012 Submitted: December 20, 2011 Accepted: April 23, 2012 Published: October 15, 2012 Discrete Neural Altitude Control for Hypersonic Vehicle Via Flight Path Angle Tracking Shixing Wang, Bin Xu, Fuchun Sun and Yifu Jiang Department of Computer Science and Technology, Tsinghua University, Beijing 100084, China Abstract: In this study, the altitude control is analyzed for the longitudinal dynamics of a generic Hypersonic Flight Vehicle (HFV). By transforming altitude command into the tracking of flight path angle with fast dynamics, the system design is focusing on the control of the attitude subsystem. The virtual control is designed with nominal feedback and Neural Network (NN) approximation via back-stepping. Under the proposed controller, the Semiglobal Uniform Ultimate Boundedness (SGUUB) stability is guaranteed. The slow dynamics are transformed into the parameter estimation problem and the update law is designed. The simulation is presented to show the effectiveness of the proposed control approach. Keywords: Discrete-time, hypersonic flight vehicle, neural network INTRODUCTION Hypersonic flight vehicles are intended to present a reliable and more cost efficient way to access space. The longitudinal model of the dynamics are known to be unstable, non-minimum phase with respect to the regulated output and affected by significant model uncertainty. The sliding mode control (Xu et al., 2004) is based on the input-output linearization using Lie derivative notation. Robust control (Wang and Stengel, 2000) took advantage of genetic algorithm to track the parameter of inverse nonlinear system. The adaptive control (Gibson et al., 2009) and robust control (Hu et al., 2010) are investigated by linearizing the model at the trim state. Intelligent modeling is also applied on HFV control. The basic idea is to formulate the HFV model with fuzzy TS model (Hu et al., 2011) or characteristic modeling (Luo and Li, 2011). The great convenience of this way is that the complicated dynamics are synthesized by several linear systems. The sequential loop closure controller design (Fiorentini et al., 2009) is based on the decomposition of the equations into functional subsystems. The method followed the approach that combined robust adaptive dynamic inversion with backstepping arguments to obtain control architecture (Kokotovic, 1991). In Xu et al. (2011), one high gain observer based controller is proposed for HFV control with only one NN to compensate the lumped uncertainty. In Gao and Sun (2011), the attitude states are considered as fast dynamics where the altitude control is transformed into the flight path angle tracking. In this study, we analyze the discrete flight control of hypersonic vehicles. For the modeling, the discrete HFV model is obtained by Taylor expansion in this study. By proper assumptions the discrete model is transformed into the strict-feedback form where some theoretical results have been studied in Yang et al. (2008). For the nonlinearity and the coupling, the nominal part of the dynamics should be considered for the feedback control design to provide good performance. In GAO et al. (2009), the back-stepping design of HFV is proposed. The upper bound of control gain is taken to avoid the circular design with new stability analysis (Xu et al., 2011). In this study, we took the fast and slow dynamics for controller design. Continued with the work in Xu et al. (2011), the altitude tracking is done with flight path angle tracking. The results are related with system states and reference command performance. In this study, we transform the altitude control of HFV into the flight path angle tracking. In addition, considering the characteristics of HFV, the nominal part of the nonlinearity during each step is eliminated and the NN is taken to approximate the system uncertainty. Moreover, parameter estimation is designed for the linearized slow dynamics. Simulation result shows the effectiveness of this method. METHODOLOGY Hypersonic vehicle modeling: The control-oriented model of the longitudinal dynamics of a generic hypersonic aircraft considered in this study is given by Wang and Stengel (2000). This model is comprised of five state variables X = [V, h, ", (, q]T and two control inputs Uc = [*e, $]T, where V is the velocity, ( is the flight path angle, h is the altitude, " is the attack angle, q is the Corresponding Author: Shixing Wang, Department of Computer Science and Technology, Tsinghua University, Beijing 100084, China 3930 Res. J. Appl. Sci. Eng. Technol., 4(20): 3930-3936, 2012 pitch rate, *c is elevator deflection and $ is the throttle setting: V& = T cos α − D µ sin γ − m r2 x&1 = f 1 ( x1 ) + g1 ( x1 ) x 2 x& 2 = f 2 ( x1 , x 2 ) + g 2 ( x1 , x 2 ) x 3 x& 3 = f 3 ( x1 , x 2 , x 3 ) + g 3 ( x1 , x 2 , x 3 )u y = x1 (1) h& = V sin γ (2) L + T sin α µ − V r cos γ − mV Vr 2 The definition of fi and gi, i = 1, 2, 3 V, can be found in Xu et al. (2011). The control objective of system (8) is to design an adaptive controller, which makes (6(d , further h6hd and all the signals involved are bounded. 2 γ& = α& = q − γ& q& = M yy (3) (4) Assumption 2: fi and gi are unknown smooth functions. fi can be decomposed into the nominal part fiN and the unknown part )f. There exist constants gi and gi such that gi ≥ gi ≥ gi > 0 , i = 1, 3, V. (5) I yy where, T, D, L and Myy represent thrust, drag, lift-force and pitching moment respectively. m, Iyy and : represent the mass of aircraft, moment of inertia about pitch axis and gravity constant. r is the radial distance from center of the earth and r = h+RE. More information about this model can be found in Xu et al. (2004) and Wang and Stengel (2000). Given the tracking reference Vd and hd , we design the velocity and altitude controller separately in Section IV. System transformation: Strict-feedback formulation: Assumption 1: Since r is quite small, we take sin ( = ( in (2) for simplification. The thrust term .T sin ". in (3) is neglected because it is generally much smaller than L. C Discrete-time model: By first-order Taylor expansion with sample time Ts, systems (6) and (8) can be approximated as: Slow dynamics: The velocity subsystem (1) can be rewritten as follows: uV = β yV = V V(k+1) = V(k)+Ts[fV(k)+gV(k)uV(k)] (9) x1(k+1) = x1(k)+Ts[f1(k)+g1(k)x2(k)] x2(k+1) = x2(k)+Ts[f2(k)+g2(k)x3(k)] x3(k+1) = x3(k)+Ts[f3(k)+g3(k)u(k)] (10) Discrete control design: Adaptive NN control for fast dynamics: The errors are defined as: V& = f V + gV uV C (8) z1(k) = x2(k)!x1d(k) (11) z2(k) = x2(k)!x2d(k) (12) z3(k) = x3(k)!x3d(k) (13) (6) where, x2d(k), x3d(k) are the virtual control inputs to be designed. Fast dynamics: The tracking error of the altitude is ~ defined as h = h − hd and the flight path command is chosen as: ⎡ − k (h − h ) − k (h − h )dt + h& ⎤ h d I ∫ d d ⎥ V ⎢ ⎥ ⎣ ⎦ γ d = arcsin ⎢ Step 1: From (11): z1(k+1) = x1(k)+Ts[f1(k)+g1(k)x2(k)]!x1d(k+1) (14) where, x1d(k+1) is acquired from (7): (7) Define X1(k) = [V(k), x1(k), x1d(k+1)]T. The uncertainty is defined as: if kh>0 and kI>0 are chosen and the flight-path angle is controlled to follow (d, the altitude tracking error is regulated to zero exponentially. Define X = [x1, x2, x3]T, x1 = (, x2 = 2D, x3 = q, where 2D = "+(. Then the strict-feedback form equations of the altitude subsystem (2)-(5) are written as: U 1 (k ) = 3931 g1 [ − x1 ( k ) − Ts f 1 ( k ) + x1d ( k + 1)] + x1 ( k ) + Ts f 1 N ( k ) g1 ( k ) = ω *1T θ1 ( X 1 ( k )) + ε1 ( X 1 ( k )) (15) Res. J. Appl. Sci. Eng. Technol., 4(20): 3930-3936, 2012 where, fIN(k) is the nominal part of f1(k), T*1 is the optimal parameters for NN to approximate U1(k) and g1(X1) is the NN reconstruction error. Take x2(k) in (14) as the virtual control input and design its desired value as: x2d ( k ) = − x1 ( k ) − Ts f 1 N ( k ) + c1 z1 ( k ) + ω$ 1T ( k )θ1 ( X 1 ( k )) The update law for the NN weights is: ω$ 2 ( k + 1) = ω$ 2 ( k ) − λ2 z2 ( k + 1)θ2 ( X 2 ( k )) − δ 2ω$ 2 Z3(k+1) = x3(k)+Ts[f3(k)+g3(k)u(k)]-x3d(k+1) where, 0<c1<1 and ω$ 1 is the estimation of T*1. Combining (12), (14) and (16), the following equation can be obtained: ] U 3 (k ) = + Ts g1 ( k ) z2 ( k ) Step 2: From (12): u( k ) = z2(k+1) = x2(k)+Ts[f2(k)+g2(k)x3(k)]!x2d(k+1) (19) Define X2(k) = [V(k), x1(k), x2(k), x1d(k+1), x1d(k+2)]. From (16), x2d(k+1) involves x1(k+1), f1(k+1), z1(k+1) and x1d(k+1). It can be concluded that x2d(k+1) is the function of X2(k). The uncertainty U2(k) is defined and can be approximated by NN as: c3 z3 ( k ) + ω$ 3T ( k )θ3 ( X 3 ( k )) − x 3 ( k ) − Ts f 3 N ( k ) z3 ( k + 1) = (20) where, g3 ( k ) cz g3 3 3 [ g3 ( k ) ~ T ω 3 ( k )θ3 ( X 3 ( k )) − ε 3 ( X 3 ( k )) g3 2 2 2 (27) ω~3 ( k ) = ω$3 ( k ) − ω *3 : The update law for the NN weights is: Theorem 1: Considering system (10) with the controller (26) and the update law (18), (23) and (28), all the signals involved are semiglobal uniform ultimate bounded. Proof: The following Lyapunov function is considered: L(k) = L1(k)+L2(k)+L3(k) 2 ] ω$ 3 ( k + 1) = ω$ 3 ( k ) − λ3 z3 ( k + 1)θ3 ( X 3 ( k )) − δ 3ω$ 3 ( k ) (28) (21) (22) (26) Tsg 3 ( k ) + where, 0<c2<1 and ω$ 2 is the estimation of *2. The result of (20) and (21) is due to the fact that f2 = 0 and g2 = 1. Combining (13), (19) and (21), the following equation can be obtained: z2 ( k + 1) = c2 z2 ( k ) + Ts g 2 ( k ) z3 ( k ) + [ω~ T ( k )θ ( X ) − ε ( X )] (25) where, 0<c3<1 and ω$ 3 is the estimation of T*3. Then: where, T*2 is the optimal parameters and g2(X2(k)) is the NN reconstruction error. Take x3(k) in (19) as the virtual control input and design its desired value as: X3d(k) = [Tsg2(k)-1][-x2(k)+c2z2(k)+ ω$ T2(k)22(X2)] g3 ( k ) + x 3 ( k ) + Ts f 3 N ( k ) where, T*3 is the optimal parameters for NN to approximate U3(k) and g3(X3(k)) is the NN reconstruction error. The actual control input is designed as: ω$ 1 ( k + 1) = ω$ 1 ( k ) − λ1 z1 ( k + 1)θ1 ( X 1 ( k )) − δ1ω$ 1 (k) (18) U2(k) = x2d(k+1) = T* 2 (X2(k))+ ω$ 2(X2(k)) g 3 [ − x 3 ( k ) − Ts f 3 ( k ) + x 3d ( k + 1)] = ω *T3 θ3 ( X 3 ( k )) + ε 3 ( X 3 ( k )) ~ ( k ) = ω$ ( k ) − ω * . The update law for the where, ω 1 1 1 NN weights is: T 2 2 (24) Define X3(k) = [V(k),XT(k), x1d(k+1), x1d(k+2), x1d(k+3)]T. Similarly we can deduce that the uncertainty U3(k) is the function of X3(k) and it can be approximated by NN as: z1 ( k + 1) = x1 ( k ) + Ts [ f 1 ( k ) + g1 ( k ) x 2 ( k )] − x1d ( k + 1) g1 ( k ) c1 z1 ( k ) + ω~1T ( k )θ1 ( X 1 ( k )) − ε1 ( X 1 ( k )) (17) = g1 [ (23) Step 3: From (13): (16) Ts g1 ~ ( k ) = ω$ ( k ) − ω * . where, ω 2 2 2 where, 2 L1(k) = L11(k)+L12(k) 3932 (29) Res. J. Appl. Sci. Eng. Technol., 4(20): 3930-3936, 2012 2Ts g 1 z2 ( k ) z1 ( k + 1) ≤ Ts g 1 ρ14 z22 ( k ) + Ts g 1 ρ14−1 z12 ( k + 1) (38) L2(k) = L21(k)+L22(k) L3(k) = L31(k)+L32(k) Similar to Ge et al. (2008), we have the following inequality: with, ω~1T ( k + 1)ω~1 ( k + 1) ≤ (1 − δ1 )ω~1T ( k )ω~1 ( k ) L11(k) = z21(k) L12 ( k ) = ( + Ts g 1 ρ14 z22 ( k ) + λ1 ρ11ε M2 + δ1 ω *1 λ1 + δ1 (δ1 − 1 + λ1 ρ13 ) ω$ 1 ( k ) L21(k) = z22(k) L22 ( k ) = ω~2T ( k )ω~2 ( k ) λ2 2 (39) + (λ1 ρ11−1 + λ1δ1 ρ13−1 N 1 )z12 ( k + 1) ∆ L1 ( k ) ≤ − ω~3T ( k )ω~3 ( k ) λ3 + Define the first difference of L(k) as: + ∆ L( k ) = L( k + 1) − L( k ) = ∆ L1 ( k ) + ∆ L2 ( k ) + ∆ L3 ( k ) ω~1 ( k + 1) = ω~1 ( k ) − λ1θ1 ( X 1 ( k )) z1 ( k + 1) − δ1ω$ 1 λ1 ρ11ε 212 M + δ1 ω *1 λ1 2 + δ1 (δ1 + λ1 ρ13 − 1) ω$ 1 ( k ) λ1 Ts g 1 ρ24 z32 ( k ) λ2 (31) λ1 N 1 + c1 ρ12 + ρ11−1 + δ1 ρ13−1 N 1 − 1 ≤ 0 δ1 + λ1 ρ13 − 1 ≤ 0 (32) Then we have the conclusion of (41): ∆ L1 ( k ) ≤ − τ 11 L11 ( k ) − τ 12 L12 ( k ) + τ 13 + τ 14 L21 ( k ) where N1 is the number of the NN rules: From (17), we have: where, g z ( k + 1) 2 1 1 g1 ( k ) − c1 z1 ( k ) z1 ( k + 1) τ 11 = (33) + ε1 ( k ) z1 ( k + 1) − Ts g 1 z 2 ( k ) z1 ( k + 1) 2g1(k)z1(k+1)#D11 g21M+D-1 11Z21(k+1) (34) 2z1(k+1)z1(k)#D12z21(k+1)+D-1 12z21(k) (35) ~ T ( k )ω ~ ( k ) + ω$ ( k ) ~ T ( k )ω$ ( k ) = ω 2ω 1 1 1 1 1 2ω$ 1 ( k )θ1 ( X 1 ( k ))z1 ( k + 1) ≤ ρ13 ω$ 1 ( k ) 2 δ1 λ1 τ 12 = 1 − c1 ρ12−1 + 2 − ω *1 (36) (37) τ 14 = 3933 Ts g 1 ρ14 λ1 λ1 ρ11ε12M + δ1 ω *1 τ 13 = λ1 2 + ρ13−1 N 1 z12 ( k + 1) (40) 2 Provided the following inequalities hold: Noting the following facts: φ12 ( X 1 ( k )) ≤ N 1 ⎛ Ts g 1 ρ14 ⎞ 2 δ1 ~ T ⎟z ω A ( k )ω~ A ( k ) − ⎜⎜ 1 − c1 ρ12−1 + λ1 λ1 ⎟⎠ 1 ⎝ + (λ1 N 1 + c1 ρ12 + ρ11−1 + δ1 ρ13−1 N 1 − 1)z12 ( k + 1) (30) From (18) and ω~1 ( k ) = ω$ 1 ( k ) − ω *1 , we have: ω~1 ( k )θ1 ( X 1 ( k )) z1 ( k + 1) = 2 ⎛ ⎞ g1 + ⎜⎜ λ12 N 1 − 2λ1 + c1λ1 ρ12 ⎟⎟ z12 ( k + 1) g k ( ) ⎝ ⎠ 1 L31(k) = z23(k) L32 ( k ) = ) + c1λ1 ρ12−1 + Ts g 1 ρ14 z12 ( k ) ω~1T ( k )ω~1 ( k ) Ts g 1 ρ14 λ1 2 (41) Res. J. Appl. Sci. Eng. Technol., 4(20): 3930-3936, 2012 Similarly the following conclusions can be acquired: )L2(k)#!J21L21(k)J22L22(k)+J23J24L31 (42) )L3(k)#!J31L31(k)-J32L32(k)+J33 (43) Adaptive control for slow dynamics: There already exist some methods such as PID controller design and NN control on HFV velocity control. One way is to model the unknown function as follows: where τ 21 = δ2 3 λ2 Ts ρ24 λ ρ ε 2 + δ ω *2 τ 23 = 2 21 2 M 2 λ2 τ 31 = fV(k) = wTf(k)2f(k) (47) V(k+1) = V(k)+Ts[fV(k)+gV(k)uV(k)] λ2 = 2TV(k)TV(k) (48) where, TV(k) is composed state variables while 2V is the system parameters. For more detail about the formulation, refer to Fiorentini et al. (2009). However here we consider the discrete system so it should be with sampling period Ts. The control gain of uV(k) is denoted as B(k) and 2 Ts ρ24 T T T we have θV ( k ) = [θV ( k ), B( k )] λ2 δ3 λ3 τ 32 = 1 − c3 ρ (46) In this part, we try to model the uncertainty with parameter projection (Chen and Narendra, 2001): τ 22 = 1 − c2 ρ22−1 + τ 24 = g-1V(k) = wTg(k)2g(k)g!1V(k) The estimation error of V(k) is denoted as: eV ( k ) = V$ ( k ) − V ( k ) = θ$VT ( k − 1)ωV ( k − 1) − V ( k ) −1 32 λ3 ρ31ε 32M + δ 3 ω *3 τ 33 = λ3 (49) Define a(k): 2 ⎧⎪ 1, if eV ( k ) > ∆ a (k ) = ⎨ ⎪⎩ 0, otherwise Combining (41)-(43), we have the following inequality: )L(k)#!K1L1(k)!K2L2(k)-K3L3(k)+K4 The parameter projection method is selected as: (44) a ( k )eV ( k )ωV ( k − 1) (51) θ$V ( k ) = θ$V ( k − 1) − 2 1 + ωV ( k − 1) where the proper positive parameters Dij can be designed to have the following results: The controller is designed as: K1 = τ 11 > 0 ( K2 = min {τ 21 − τ 12 , τ 22 } > 0 K3 = min {τ 31 − τ 24 , ς 32 } > 0 (50) ) 1 uv ( k ) = $ Vd ( k + 1) − θVT ( k )ω ( k − 1) B( k ) (45) (52) SIMULATIONS K4 = τ 23 + τ 33 > 0 Reference commands are generated by the filter: Take rA = min{K1,K2,K3}, then )L(k)#!rAL(k)+K4. we know L(k) is bounded, so zi(k),i = 1, 2, 3 and Tj(k), j = 1, 2, 3, are semiglobal uniform ultimate bounded. This completes the proof. 3934 hd ω n1ω n22 = hc ( s + ω n1 )( s 2 + 2ε cω n 2 s + ω n22 ) (53) Res. J. Appl. Sci. Eng. Technol., 4(20): 3930-3936, 2012 Vd ωn3 = Vc ( S + ωn 3 ) -4 (54) 14 12 3 Attitude (ft) Yd γ 10 8 6 4 where, Tn1 = 0.2, Tn2 = 0.5, gc = 0.7, Tn3 = 10. 1.103 X 10 ×10 2 0 1.102 1.101 -2 -4 1.100 0 10 20 30 Time (sec) 40 50 60 1.099 0 10 20 30 40 50 Fig. 3: Flight path angle tracking (a) ω1 ω2 ω3 ω4 0.12 Time (sec) 4 ×10 Velocity (ft/s) 1.520 0.10 0.08 1.515 0.06 1.510 0.04 0.02 1.505 0 10 20 30 Time (sec) 40 50 60 0 0 10 20 (b) 30 Time (sec) 40 50 60 Fig. 4: Attack angle Fig. 1: Altitude and velocity response 0.08 0.06 10 rad Elevator deflection (deg) 0.10 20 0.04 0.02 0 0 -10 0 10 20 30 Time (sec) 40 50 -0.02 60 -0.04 0 (a) 10 20 30 Time (sec) 40 50 60 Fig. 5: Pitch rate Throttle setting 0.8 The parameters for the controller are selected as kh = 0.15, kI = 0.02, 81 = 0.05, 82 = 0.05, 83 = 0.05, *1 = 0.02, *2 = 0.02, *3 = 0.02, Ts = 0.02s, c1 = 0.9, c2 = 0.95, c3 = 0.87, cV = 0.9. 0.6 0.4 0.2 0 0 10 20 30 Time (sec) (b) Fig. 2: Control inputs 40 50 60 Figure 1 depicts the response performance that the altitude controller tracks the step change with magnitude 200 ft while the velocity steps from 15060 to 15160 ft/s. The control inputs of the elevator deflection and the throttle setting are shown in Fig. 2. Flight path angle tracks the reference command very well in Fig. 3 so that the system tracks the altitude step change. From the attack 3935 Res. J. Appl. Sci. Eng. Technol., 4(20): 3930-3936, 2012 angle in Fig. 4 and pitch rate response in Fig. 5, we know the related system states are bounded. CONCLUSION The altitude control of HFV is transformed into the flight path angle tracking. Considering the characteristics of HFV, the nominal part of the nonlinearity during each step is eliminated and the NN is taken to approximate the system uncertainty. Parameter estimation is designed for the linearized slow dynamics. Simulation result shows the effectiveness of this method. ACKNOWLEDGMENT This study was supported by the National Science Foundation of China (Grants No: 60625304, 90716021). The authors gratefully acknowledge Daoxiang Gao and Chenguang Yang for the support regarding the discrete controller design. REFERENCES Chen, L. and K. Narendra, 2001. Nonlinear adaptive control using neural networks and multiple models. Automatica, 37(8): 1245–1255. Fiorentini, L., A. Serrani, M. Bolender and D. Doman, 2009. Nonlinear robust adaptive control of flexible air-breathing hypersonic vehicles. J. Guid. Contr. Dyn., 32(2): 401-416. GAO, D., Z. Sun and T. Du, 2009. Discrete-time controller design for hypersonic vehicle via backstepping. Control Decision, 249(1), DOI: CNKI:SUN:KZYC.0.2009-03-027. Gao, D. and Z. Sun, 2011. Fuzzy tracking control design for hypersonic vehicles via ts model. Sci. China Inform. Sci., 54(3): 521-528. Ge, S., C. Yang, and T. Lee, 2008. Adaptive predictive control using neural network for a class of pure feedback systems in discrete time. IEEE T. Neural Network, 19(9): 1599-1614. Gibson, T., L. Crespo and A. Annaswamy, 2009. Adaptive control of hypersonic vehicles in the presence of modeling uncertainties. IEEE Am. Contr. Conf. ACC’09, pp: 3178-3183. Hu, Y., F. Sun and H. Liu, 2010. Neural network-based robust control for hypersonic flight vehicle with uncertainty modelling. Int. J. Modell. Identif. Contr., 11(1): 87-98. Hu, Y., Y. Yuan, H. Min and F. Sun, 2011. Multiobjective robust control based on fuzzy singularly perturbed models for hypersonic vehicles. Sci. China Inform. Sci., 54(3): 563-576. Kokotovic, P., 1991. The Joy of Feedback: Nonlinear and Adaptive: 1991 bode Prize Lecture. IEEE Contr. Syst. Magazine, 12: 7-17. Luo, X. and J. Li, 2011. Fuzzy dynamic characteristic model based attitude control of hypersonic vehicle in gliding phase. Sci. China Inform. Sci., 54(3): 448459. Xu, H., M. Mirmirani and P. Ioannou, 2004. Adaptive sliding mode control design for a hypersonic flight vehicle. J. Guid. Contr. Dyn., 27(5): 829-838. Xu, B., D. Gao and S. Wang, 2011a. Adaptive neural control based on hgo for hypersonic flight vehicles. Sci. China Inform. Sci., 54(3): 511-520. Xu, B., F. Sun, C. Yang, D. Gao and J. Ren, 2011b. Adaptive discrete-time controller design with neural network for hypersonic flight vehicle via backstepping. Int. J. Contr., 84(9): 1543-1552. Yang, C., S. Ge, C. Xiang, T. Chai and T. Lee, 2008. Output feedback nn control for two classes of discrete-time systems with unknown control directions in a unified approach. IEEE T. Neural Network, 19(11): 1873-1886. Wang, Q. and R. Stengel, 2000. Robust nonlinear control of a hypersonic aircraft. J. Guid. Contr. Dyn., 23(4): 577-585. 3936