This paper investigates a single parameter adaptive neural network control method for unknown nonlinear systems with bounded external disturbances. A smooth performance function is developed to achieve the transient and steady state of system tracking error that could be constrained in prescribed bounds. The difficulties in dealing with unknown system parameters and disturbances of nonlinear systems are resolved based on the single parameter adaptive neural network control which is proposed to effectively reduce the calculation amount. The theoretical analysis implies that the proposed control scheme makes the closed-loop system uniformly ultimately bounded. Simulation demonstrates that the proposed adaptive controller gives a favorable performance on tracking desired signal and constraining the bounds of tracking error which could be arbitrarily small with appropriate adaptive parameters. Both the theoretical analysis and simulations confirm the effectiveness of the control scheme.
National Natural Science Foundation of China617034021. Introduction
Neural networks (NNs) have been widely applied to estimate continuous unknown functions in complex and model-uncertain systems [1–5]. The primary advantage of NN techniques is that the controller does not require the accurate system information. During the recent researches, adaptive control was commonly used to combine with neural control [6–10], in particular with RBF NNs [11–13]. A direct adaptive controller [11] based on an improved RBF NN was proposed for an omnidirectional mobile robot. In [12], a model reference adaptive sliding mode using an RBF NN was proposed to control the single-phase active power filter. An adaptive fault-tolerant attitude control method [13] was presented of rigid body using RBF NN.
Although the stabilities of system are guaranteed by using adaptive RBF NN control design techniques and tracking errors of output could converge in the finite time, a desired performance of transient and steady state cannot be easily obtained. However, the transient and steady state performance plays a significant role when some control requirements are raised such as avoiding large overshoot. These requirements stimulate tracking-error-constrained developments such as the barrier Lyapunov function method [14–17], adaptive inverse of backlash [18], and the error transformation method [19–22]. In order to obtain the prescribed transient and steady state performance, we use a smooth function to regulate constraints in this study.
Recently, we have made some achievements on adaptive NN control [23, 24] and constrained control [25, 26]. In this study, we consider an adaptive NN control for unknown nonlinear systems with tracking error constraints. Our control objective is to develop the controller achieving the desired trajectory tracking and constraining the tracking error acting as our required performance. First, we use a smooth function which could regulate constraints such as the convergence speed, the maximum bounds of overshoot, and transient and steady state of the tracking error. Then, the constrained function is transformed to an unconstrained variable to facilitate the controller design. To proceed, an RBF NN, combined with the single parameter adaptive (SPA) control method, is designed to address the function-unknown issue of the system. The main contributions include the following three aspects:
The system tracking error could be constrained in the required bounds. The transient and steady state, convergence speed, and maximum bound of overshoot of the tracking error could be arbitrarily prescribed and regulated by the parameters of the proposed smooth performance function
Generally, NN weight vectors were used as the adaptive parameters in many literatures [27, 28]. To improve it and decrease the online computation, we design a single parameter adaptive (SPA) control method. We use only one parameter to estimate the norm of weight vector online, shortening the calculated time. Meanwhile, we do not have to consider the structure of weight vector but a single parameter instead, which also simplifies the design of the controller.
The research object is a nonlinear system, with unknown functions as well as external disturbances, which increases difficulties in designing the effective controller. Whereas the proposed controller is concise and the closed-loop system is guaranteed uniformly ultimately bounded
A coupled motor drive (CMD) system [29], generally applied in intelligent robots [30], is taken for simulation, which displays the favorable tracking performance of output under the designed controller. The effectiveness of the developed scheme is validated, and the convergence rate of the tracking error could be arbitrarily small through selecting the appropriate parameters.
The paper is organized as follows. Section 2 gives the model structure of the SISO nonlinear system, and then the performance function is also proposed. Section 3 presents an ideal controller. A single parameter adaptive neural network control method is introduced in Section 4, where the stability of the closed-loop system and the favorable performance of the control scheme are all guaranteed. Numerical results are shown in Section 5 to illustrate the effectiveness of the control scheme. Finally, the conclusion is presented in Section 6.
2. System Description and Preliminaries
Consider a class of unknown single-input single-output unknown nonlinear system as follows:
(1)ẋi=xi+1,i=1,2,…,n−1,ẋn=fx+bxu+dt,y=x1,where x=x1,x2,…,xnT∈Rn, u∈R, and y∈R are, respectively, the state, input, and output of the system; fx and bx denote unknown nonlinear smooth functions; and dt represents the unknown disturbance bounded as dt≤d1, where d1 is a positive constant. As all values of actual physical quantities in (1) are limited, x lies in a subset Φ∈Rn.
The control goal is designing the control input u such that the output y could track the desired trajectory yd and the tracking error et=y−yd acts as our required performance. For the second purpose, we first define a smooth function as
(2)μt=μ0−μ∞exp−lt+μ∞,where μ0, μ∞, and l are positive constants and μ∞<μ0. Meanwhile, an asymmetric hyperbolic tangent function Tξ is defined as follows:
(3)Tξ=expξ−σexp−ξexpξ+exp−ξ,ife0≥0,σexpξ−exp−ξexpξ+exp−ξ,ife0<0,where 0≤σ≤1 is a design constant and ξ is the transformed error variable. Obviously, Tξ possesses the following properties:
(4)−σ<Tξ<1,ife0>0,−1<Tξ<σ,ife0<0.
We define the following equation:
(5)et=μtTξ.
From (2), (3), and (5), we conclude that when e0≥0, it yields
(6)−σμt<et<μt,and when e0<0, we obtain
(7)−μt<et<σμt.
Remark 1.
The smooth function μt defined in (2) is designed as the guidance of the tracking error e such that the performance of e is constrained and complies with the performance of μt. To be specific, the constant μ∞ denotes the upper bound of e at the steady state, and the constant l represents the decreasing rate of μt, i.e., the bound of prescribed convergence speed of e, assuming that e0<μ0. In addition, σμ0 is the bound of the required maximum overshoot.
Remark 2.
Tξ is an error transforming function for obtaining an unconstrained tracking error ξ from the constrained system (5), and Tξ is a smooth and increasing function.
As the properties of Tξ, we could get the inverse function of Tξ as follows:
(8)ξ=T−1etμt,where ξ∈L∞ and the tracking error e is constrained in the set:
(9)E=e∈R:et<μ∞.
Thus, if we guarantee ξ bounded ∀t≥0, we can realize the fact that y tracks the desired trajectory yd and the tracking error e acts as our prescribed performance. In addition, the required performance of the system output is regulated by the appropriate choice of μt and σ.
Define the filtered error function s as
(10)s=c1ξ+c2ξ̇+⋯+cn−1ξn−2+ξn−1,where ci, i=1,2,…,n−1, is selected as ci>0 such that the polynomial λn−1+cn−1λn−2+⋯+c2λ+c1 is Hurwitz. Then, we could confirm that ξ is bounded if we guarantee the convergence of the filtered error function s [31]. Thus, the transient and steady state, convergence speed, and maximum bound of overshoot of the tracking error could be arbitrarily prescribed constrained and regulated by the parameters of the proposed smooth performance function.
Assumption 1.
The known desired trajectory signal yd∈Φd⊂Rn and its first to nth order time derivatives are continuous and bounded, where Φd is a connected subset of Φ.
From (1), (5), and (7), differentiating s with respect to time, we have
(11)ṡ=c1ξ̇+c2ξ¨+⋯+cn−1ξn−1+ξn=v+∂T−1∂et/μt1μtẋn=v+w+Fu+dft,where v is the term containing the following known variables: e,ė,…,en−1, μ,μ̇,…,μn, ydn, and the unknown term w=∂T−1/∂et/μt1/μtfx, F=∂T−1/∂et/μt1/μt1/μtbx, dft=∂T−1/∂et/μt1/μtdt.
Assumption 2.
The sign of bx is knowable and bx>0, ∀x∈Φ. It is assumed that bx>0. As the fact of ∂T−1/∂et/μt1/μt>0, we have F>0.
Assumption 3.
It could be found a smooth function F¯ such that F≤F¯, F¯≤F¯b∈R. The disturbance dft is bounded by a known positive constant d0, i.e., dft≤d0.
Remark 3.
The restrictions of Assumptions 2 and 3 are actually the same as the ones in the literature [27]. Many robotic systems, including CMD system in this paper, possess the properties as those of Assumptions 2 and 3.
3. Design of Ideal Controller
To design ideal controller u^, we assume that the unknown functions fx and bx are known and dt=0.
Theorem 1.
Given the system (1), Assumptions 1–3, the ideal controller is developed by
(12)u^=−1Fv+w+1δ+1δF−Ḟ2Fs,where δ is a positive parameter. Then, we have limt→∞s=0.
Proof 1.
Substituting u=u^ into (11), we get
(13)ṡ=−1δ+1δF−Ḟ2Fs.
Define a Lyapunov candidate as V′=1/2Fs2, then differentiating it with respect to time yields
(14)V̇′=1Fsṡ−Ḟ2F2s2=−1F1δ+1δF−Ḟ2Fs2−Ḟ2F2s2=−1δF+1δF2s2≤0.
According to the Lyapunov theorem, we obtain limt→∞s=0.
From (14), we can conclude that the smaller δ is, the bigger convergence rate will be. Owning to the unknown functions fx and bx, u^ is also unknown. Then, NN will be applied to develop the actual controller u in the following section.
4. Design of Actual Controller
We rewrite the ideal controller (12) to the following form, in which the ideal controller u^ is regarded as a function:
(15)u^=uψ,ψ=xTvsT∈Φψ⊂Rn+2,where the subset Φψ is expressed as
(16)Φψ=xTvs∣x∈Φ,yd∈Φd.
It is well known that continuous functions could be approximated by a linear combination of Gaussians. Because u^ is continuous on Φψ in (15), it has an ideal NN weight vector W¯ as follows:
(17)u^ψ=W¯Thψ+ϖε,where ϖε denotes the approximation error as ϖε≤ϖ0, ϖ0>0, and hψ denotes the radial-basis function vector as
(18)h=h1h2⋯hj⋯hmT,hj=exp−ψ−Aj2bj2,j=1,2,…,m,where ψ∈Rn+2 represents the input vector; n+2 denotes input neural nets number in the input layer; m is hidden neural nets number in the hidden layer; hj is commonly used Gaussian function; Aj=aj,aj,⋯,ajT∈Rn+2; and a∈Rm and bj∈R, respectively, denote the vector of the center of the receptive field and the width of Gaussian function.
Assumption 4.
There is a bound of ideal weight vector W¯ such that
(19)W¯F≤wmax,where wmax is a positive constant.
Then, a constant is introduced to improve the control scheme as
(20)ϕ=W¯F2.
Remark 4.
As in Assumption 4, ϕ is also bounded.
Remark 5.
ϕ is an unknown positive constant because of W¯ being unknown.
We define ϕ^ as the estimate of ϕ, then the estimate error is denoted as ϕ~=ϕ^−ϕ. Then, the SPA control technique will be introduced to effectively reduce the calculation amount. We use only one parameter ϕ^ to estimate the norm of weight vector W¯ online, instead of the estimate of a vector W¯, which could considerably decrease the online computation. Meanwhile, we do not have to consider the structure of W¯ but a single parameter ϕ^ instead, which also simplifies the design of the controller. It is one of the main contributions in our study.
First, we propose the controller as follows:
(21)u=−12δsϕ^hTh.
Substituting (21) into (11) yields
(22)ṡ=v+w+F−12δsϕ^hTh+dft.
From (17), we rewrite (22) by adding and subtracting Fu^ψ simultaneously as
(23)ṡ=v+w+F−12δsϕ^hTh−W¯Th−ϖε+Fu^ψ+dft,and substituting (12) into (23), we get
(24)ṡ=F−12δsϕ^hTh−W¯Th−ϖε−1δ+1δF−Ḟ2Fs+dft.
In order to validate all the states of system stabilizing in the compact subset Φ, we propose the following theorem.
Theorem 2.
On account of system (1) and Assumptions 1–4, with the SPA RBF NN controller (21), and the adaptive law:
(25)ϕ^̇=γ2δs2hTh−κγϕ^,where γ, κ>0, the tracking error et is uniformly ultimately bounded in the compact subset Φ for t→∞ and can be arbitrarily small by using the appropriate parameter μ∞. Moreover, the output y=x1 in system (1) will track yd acting as our required performance.
Proof 2.
A Lyapunov function candidate is defined as follows:
(26)V=12s2F+1γϕ~2.
Taking the time derivatives of (26), and from (24), we obtain
(27)V̇=sṡF−Ḟ2F2s2+1γϕ~ϕ^̇=sFF−12δsϕ^hTh−W¯Th−ϖε−Ḟ2F2s2−sF1δ+1δF−Ḟ2Fs−dft+1γϕ~ϕ^̇=−12δs2ϕ~+ϕhTh−sW¯Th−1δF+1δF2s2+dftFs−ϖεs+1γϕ~ϕ^̇.
Noting the following inequalities:
(28)s2ϕhTh2δ+δ2=s2W¯2hTh2δ+δ2≥−sW¯Th,(29)dftFs≤s2δF2+δ4dft2,(30)ϖεs≤s22δF+δ2ϖε2F,and from ϖε≤ϖ0, dft≤d0, and Assumption 3, we have
(31)V̇≤ϕ~−12δs2hTh+1γϕ^̇−1δF+1δF2s2+dftFs−ϖεs+δ2≤ϕ~−12δs2hTh+1γϕ^̇−s22δF+δ2ϖ02F¯b+δ4d02+δ2.
Considering (25) and (26), we get
(32)V̇≤−κϕ~ϕ^−s22δF+δ2ϖ02F¯b+δ4d02+δ2≤−κ2ϕ~2−ϕ2−s22δF+δ2ϖ02F¯b+δ4d02+δ2≤−κ2ϕ~2−s22δF+δ2ϖ02F¯b+δ4d02+δ2+κ2ϕ2.
Setting κ=η/γ, η>0, it yields
(33)V̇≤−η2γϕ~2−s22δF+δ2ϖ02F¯b+δ4d02+δ2+η2γϕ2≤−αV+β,where α=minη,1/δ, β=δ/2ϖ02F¯b+δ/4d02+δ/2+η/2γϕ2.
Lemma 1 (see [32]).
Let f, V:0,∞∈R. Then, V̇≤−α0V+f, ∀t≥t0≥0 implies that
(34)Vt≤e−α0t−t0Vt0+∫t0te−α0t−τfτdτ.
Solving the inequality (33) with Lemma 1, we obtain
(35)Vt≤e−αtV0+β×∫0te−αt−τdτ≤e−αtV0−βα+βα,∀t≥0.
From the definition of V, we obtain
(36)s≤2FV≤2F¯bV.
According to (30), and a+b≤a+ba>0,b>0, we have
(37)s≤2F¯be−αt/2V0+βα,∀t≥0.
From (37), we conclude that the filtered error function s is bounded such that limt→∞s≤2F¯bβ/α, ∀t≥0, i.e., bounded in the set S=s∈R:s≤2F¯bβ/α. According to (37), the upper bound of s can be arbitrarily small which depends on the appropriate parameters α and β. α and β are related to the parameters δ, η, and γ. Obviously, the increases in γ and η or decrease in δ will bring down the upper bound. Moreover, from the definitions of (2), (3), (5), and (10), we conclude that unconstrained tracking error ξ is bounded, indicating the tracking error et is uniformly ultimately bounded and constrained in the compact subset E=e∈R:et<μ∞ for t→∞. Hence, the two control objectives that y tracks the desired trajectory yd and the tracking error e=y−yd acts as our required performance are achieved. Furthermore, the suitable choice of μt and σ leads to a favorable performance of the system output.
5. Simulation Results
In this section, a CMD system is applied to demonstrate the effectiveness of the proposed technique. The schematic of the CMD system is displayed in Figure 1, and the dynamics is given by
(38)Jlθ¨2+c12θ̇2+kθ2−gr−1θ1=0,Jdθ¨1+c11θ̇1+kgr−1gr−1θ1−θ2=Td+d′t,where Jd and Jl denote the inertias of drive and load system, respectively; θ1 and θ2 represent the drive angle position and load angle position, respectively; c11 and c12 are the rotary damping on drive and load; Td denotes the input torque; d′t is the unknown disturbance on drive; gr=rlrpl/rp2rd is the gear ratio; and k=2klrl2 is the torsional spring constant.
Schematic of the CMD system.
We set x1=θ2 as the system output, and u=Td the input, then we could rewrite the dynamics (38) as follows:
(39)Jlx¨1+c12ẋ1+kx1=kgr−1θ1,Jdθ¨1+c11θ̇1+kgr−1gr−1θ1−x1=Td+d′t.
After elimination of the term θ1 in (39), and setting the time derivatives ẋ1=x2, ẋ2=x3, and ẋ3=x4, we could obtain the following form as system (1):
(40)ẋi=xi+1,i=1,2,…,n−1,ẋn=fx+bxu+dt,y=x1,where n=4, fx=−c11k/JlJd+c12k/JlJdgr−2x2−k/Jl+c11c12/JlJd+k/Jlgr−2x3−c12/Jl+c11/Jdx4, bx=k/JlJdgr, and dt=k/JlJdgrd′t.
As n=4, and from (10), we could select the parameters in (10) as c1=27, c2=27, and c3=9. The initial states of system and the adaptive law (25) are given by x0=0.5000T and ϕ^0=0. In simulation, the parameter values of system (38) are chosen as Jl=0.3575 kg∙m2, Jd=0.000425 kg∙m2, gr=4, k=8.45 N∙m/rad, c12=0.004 N∙m∙s/rad, c11=0.05 N∙m∙s/rad, and the external disturbance d′t=0.5sint. The desired trajectory is given by yd=sint.
In simulation, a three-layer network 6-9-1 is applied. We choose the input vector as ψ=x1x2x3x4vsT. In order to guarantee RBF vector h be effective, we should make the inputs of RBF be within the effective mapping of Gaussian membership function, which indicates that the center vector a and width value bj should be designed within the effective range of the inputs. In this practical system, we could choose a=10×−2−1.5−1−0.500.511.52 and b=10. The proposed scheme is exhibited in Figure 2.
The proposed control scheme.
The simulation results are displayed in Figures 3–5. We choose different parameters to illustrate the performance of proposed controller. In Figure 3, the parameters in the controller (21) and adaptive law (25) are chosen as δ=0.5, γ=10, and η=5. The prescribed steady state of the tracking error is given by μ∞=0.1, the parameter of convergence rate l=1 and the overshoot zero such that σ=0. It is observed that in Figure 3, the tracking error is constrained by the smooth function μt and along with it. In Figure 4, we increase the convergence rate such that l=1.5, and decrease the steady state of the tracking error such that μ∞=0.01, although the parameters δ, γ, and η are the same as the values in Figure 3. Then, we further change the parameters in Figure 5 as δ=0.05, γ=100, and η=50, but not regulate μ∞ and l. It can be seen that Figures 3–5 illustrate the different performances of tracking error. From the comparison, we validate that the smooth function μt can effectively constrain the tracking error to act as our required performance. The suitable choice of μt and σ leads to a favorable performance of output. Meanwhile, the smaller δ and larger γ and η will accelerate the convergence rate of the closed-loop system and the tracking error could be arbitrarily small. Simulations illustrate the effectiveness of the proposed scheme.
Position tracking and its tracking error of the closed-loop system with the parameters l=1, μ∞=0.1, δ=0.5, γ=10, and η=5: (a) load angle position x1 and desired trajectory yd; (b) tracking error e=x1−yd.
Position tracking and its tracking error of the closed-loop system with the parameters l=1.5, μ∞=0.01, δ=0.5, γ=10, and η=5: (a) load angle position x1 and desired trajectory yd; (b) tracking error e=x1−yd.
Position tracking and its tracking error of the closed-loop system with the parameters l=1.5, μ∞=0.01, δ=0.05, γ=100, and η=50: (a) load angle position x1 and desired trajectory yd; (b) tracking error e=x1−yd.
6. Conclusion
A tracking-error-constrained controller is proposed in this study for a class of unknown nonlinear systems with external disturbances. We develop the controller to achieve that the system tracking error acts as our required performance. The ideal controller with unknown smooth functions is designed to obtain the actual one using a single parameter RBF NN adaptive control method. The developed adaptive control law guarantees the stability of the closed-loop system via the theoretical analysis. The simulations illustrate a favorable tracking performance, and the tracking error could be arbitrarily small by selecting the appropriate parameters.
Data Availability
The data used to support the findings of this study are available from the corresponding author upon request.
Conflicts of Interest
The authors declare that they have no conflicts of interest.
Acknowledgments
This work was partially supported by the National Natural Science Foundation of China (Grant no. 61703402).
YangC.JiangY.LiZ.HeW.SuC.-Y.Neural control of bimanual robots with guaranteed global stability and motion precision20171331162117110.1109/TII.2016.26126462-s2.0-85019485116YangC.WangX.ChengL.MaH.Neural-learning-based telerobot control with guaranteed performance201747103148315910.1109/TCYB.2016.25738372-s2.0-8502871319428113610HeW.MengT.HeX.SunC.Iterative learning control for a flapping wing micro aerial vehicle under distributed disturbances201811210.1109/TCYB.2018.28083212-s2.0-85046336104MoallemP.RazmjooyN.AshourianM.Computer vision-based potato defect detection using neural networks and support vector machine201328213714510.2316/Journal.206.2013.2.206-37462-s2.0-84881096279YangC.LuoJ.PanY.LiuZ.SuC.-Y.Personalized variable gain control with tremor attenuation for robot teleoperation201711210.1109/TSMC.2017.26940202-s2.0-85018863536HeW.YanZ.SunC.ChenY.Adaptive neural network control of a flapping wing micro aerial vehicle with disturbance observer201747103452346510.1109/TCYB.2017.27208012-s2.0-8502922809228885146YangC.JiangY.HeW.NaJ.LiZ.XuB.Adaptive parameter estimation and control design for robot manipulators with finite-time convergence201865108112812310.1109/TIE.2018.28037732-s2.0-85041508667ZhangS.DongY.OuyangY.YinZ.PengK.Adaptive neural control for robotic manipulators with output constraints and uncertainties201811110.1109/TNNLS.2018.28038272-s2.0-85043472865HeW.DongY.Adaptive fuzzy neural network control for a constrained robot using impedance learning20182941174118610.1109/TNNLS.2017.26655812-s2.0-8501644960128362618YangC.WuH.LiZ.HeW.WangN.SuC.-Y.Mind control of a robotic arm with visual fusion technology2017110.1109/TII.2017.27854152-s2.0-85040069020FanJ.JiaS.LiX.Direct adaptive control based on improved RBF neural network for omni-directional mobile robotProceedings of the 2015 International Conference on Mechatronics, Electronic, Industrial and Control Engineering2015Shenyang, ChinaAtlantis PressFangY.FeiJ.MaK.Model reference adaptive sliding mode control using RBF neural network for active power filter20157324925810.1016/j.ijepes.2015.05.0092-s2.0-84929990687HuoB.XiaY.ChaiS.ShiP.Adaptive fault-tolerant control of rigid body using RBF neural networksProceeding of the 11th World Congress on Intelligent Control and Automation2014Shenyang, China11851190IEEE10.1109/WCICA.2014.70528872-s2.0-84932174797HeW.GeS. S.Cooperative control of a nonuniform gantry crane with constrained tension201666414615410.1016/j.automatica.2015.12.0262-s2.0-84959457456NiuB.ZhaoJ.Tracking control for output-constrained nonlinear switched systems with a barrier Lyapunov function201344597898510.1080/00207721.2011.6522222-s2.0-84876363075TeeK. P.GeS. S.TayE. H.Barrier Lyapunov functions for the control of output-constrained nonlinear systems200945491892710.1016/j.automatica.2008.11.0172-s2.0-61849185707RenB.GeS. S.TeeK. P.LeeT. H.Adaptive neural control for output feedback nonlinear systems using a barrier Lyapunov function20102181339134510.1109/TNN.2010.20471152-s2.0-77955516678HeW.HeX.ZouM.LiH.PDE model-based boundary control design for a flexible robotic manipulator with input backlash20181810.1109/TCST.2017.27800552-s2.0-85040044434KarayiannidisY.DoulgeriZ.Model-free robot joint position regulation and tracking with prescribed performance guarantees201260221422610.1016/j.robot.2011.10.0072-s2.0-84855189414KostarigkaA. K.DoulgeriZ.RovithakisG. A.Prescribed performance tracking for flexible joint robots with unknown dynamics and variable elasticity20134951137114710.1016/j.automatica.2013.01.0422-s2.0-84876685677HeW.MengT.HeX.GeS. S.Unified iterative learning control for flexible structures with input constraints20189632633610.1016/j.automatica.2018.06.051BechlioulisC. P.RovithakisG. A.Robust partial-state feedback prescribed performance control of cascade systems with unknown nonlinearities20115692224223010.1109/TAC.2011.2157399YangH.LiuJ.An adaptive RBF neural network control method for a class of nonlinear systems20185245746210.1109/JAS.2017.75108202-s2.0-85041926691YangH. J.TanM.Sliding mode control for flexible-link manipulators based on adaptive neural networks201815223924810.1007/s11633-018-1122-22-s2.0-85043391490YangH.LiuZ.Active control of an elastic beam based on state and input constraints201861106351064310.1109/ACCESS.2018.28058002-s2.0-85042687838YangH.LiuJ.Active vibration control for a flexible-link manipulator with input constraint based on a disturbance observer201810.1002/asjc.17932-s2.0-85044726716GeS. S.HangC. C.ZhangT.A direct method for robust adaptive nonlinear control with guaranteed transient performance199937527528410.1016/S0167-6911(99)00032-82-s2.0-0000640853LewisF. L.LiuK.YesildirekA.Neural net robot controller with guaranteed tracking performance19956370371510.1109/72.3779752-s2.0-002930463518263355KuljacaO.SwamyN.LewisF. L.KwanC. M.Design and implementation of industrial neural network controller using backstepping200350119320110.1109/TIE.2002.8076752-s2.0-0037318917HeW.LiZ.ChenC. L. P.A survey of human-centered intelligent robots: issues and challenges20174460260910.1109/JAS.2017.75106042-s2.0-85029820061KhalilH. K.GrizzleJ.20023Upper Saddle River, NJ, USAPrentice HallIoannouP. A.SunJ.2012Dover Publications