Loading [MathJax]/jax/element/mml/optable/BasicLatin.js
A journal of IEEE and CAA , publishes high-quality papers in English on original theoretical/experimental research and development in all areas of automation
Volume 6 Issue 1
Jan.  2019

IEEE/CAA Journal of Automatica Sinica

  • JCR Impact Factor: 15.3, Top 1 (SCI Q1)
    CiteScore: 23.5, Top 2% (Q1)
    Google Scholar h5-index: 77, TOP 5
Song Ling, Huanqing Wang and Peter X. Liu, "Adaptive Fuzzy Dynamic Surface Control of Flexible-Joint Robot Systems With Input Saturation," IEEE/CAA J. Autom. Sinica, vol. 6, no. 1, pp. 97-107, Jan. 2019. doi: 10.1109/JAS.2019.1911330
Citation: Song Ling, Huanqing Wang and Peter X. Liu, "Adaptive Fuzzy Dynamic Surface Control of Flexible-Joint Robot Systems With Input Saturation," IEEE/CAA J. Autom. Sinica, vol. 6, no. 1, pp. 97-107, Jan. 2019. doi: 10.1109/JAS.2019.1911330

Adaptive Fuzzy Dynamic Surface Control of Flexible-Joint Robot Systems With Input Saturation

doi: 10.1109/JAS.2019.1911330
Funds:

the National Natural Science Foundation of China 61773051

the National Natural Science Foundation of China 61773072

the National Natural Science Foundation of China 61761166011

the Fundamental Research Fund for the Central Universities 2016RC021

the Fundamental Research Fund for the Central Universities 2017JBZ003

More Information
  • In this paper, we propose an adaptive fuzzy dynamic surface control (DSC) scheme for single-link flexible-joint robotic systems with input saturation. A smooth function is utilized with the mean-value theorem to deal with the difficulties associated with input saturation. An adaptive DSC design with an auxiliary first-order filter is used to solve the "explosion of complexity" problem. It is proved that all the signals in the closed-loop system are semi-globally uniformly ultimately bounded, and the tracking error eventually converges to a small neighborhood around zero. The main advantage of the proposed method is that only one adaptation parameter needs to be updated, which reduces the computational burden significantly. Simulation results demonstrate the feasibility of the proposed scheme and the comparison results show that the improved DSC method can reduce the computational burden by almost two thirds in comparison with the standard DSC method.

     

  • In recent years, considerable attention has been paid in tracking control of flexible-joint (FJ) robot systems with robot dynamics and actuator dynamics. A number of different control strategies have been reported for FJ robot tracking control; Chang et al. [1] provided an excellent review of these methods including feedback linearisation control [2], passivity control [3], singular perturbation control [4], adaptive sliding mode control [5], proportional-derivative control [6], observer-based control [2] and adaptive backstepping control [7]-[9]. The backstepping technique is one of the popular techniques to control nonlinear systems with parameter uncertainty during the past two decades. However, the backstepping technique has a drawback called "explosion of complexity". To solve the problem, Swaroop et al. [10] proposed an adaptive dynamic surface control (DSC) technique to first solve the problem by introducing a first-order filter at each step of the traditional backstepping design. Many other important research results combined with DSC have also been presented. For example, the authors in [11] considered the problem of a class of strict-feedback nonlinear systems with input saturation using adaptive neural dynamic surface control. Su [12] proposed an output-feedback tracking controller for a class of nonlinear non-minimum phase systems.

    Neural networks or fuzzy logic systems (FLSs) have been often used to approximate unknown continuous functions because of their universal approximation properties [13]. Since Polycarpou [14] first developed a neural-based adaptive controller for a class of nonlinear systems and Wang et al. [15] first presented a class of fuzzy systems to approximate a controller for a nonlinear ball and beam system, adaptive neural control and fuzzy control of nonlinear systems have received extensive attention in theoretical and practical applications [16]-[28]. It is worth mentioning that some of these articles have improved the structure of neural networks or fuzzy systems by combining network techniques. The successful application of these technologies will greatly enhance the control of nonlinear systems in the future. On the other hand, many researchers have used fuzzy systems to address control design of nonlinear systems in recent years. For instance, an adaptive fuzzy output feedback control method for single-input-single-output (SISO) nonlinear systems are presented to solve the problems associated with unmodeled dynamics and dynamic disturbances [18]. The FLSs are extended to control FJ robots [19], and an adaptive type-2 fuzzy logic controller (FLC) is proposed for FJ robotic manipulators with dynamical uncertainties [20].

    As it is well known, input saturation may lead to control failure due to limited input amplitude of the actuator. Significant research efforts have been made to solve input saturation in the past few years [29]-[35]. The authors ([30] and [31]) solved the problem of uncertain systems with input saturation using fuzzy output control, and the problem of saturation is solved by approximation with a smooth function. Combining fuzzy logic systems and a small-gain approach, Zhou et al. [32] considered the problem of input saturation for single-input and single-output nonlinear systems with the nonstrict-feedback form. To solve the input saturation issue in MIMO nonlinear time-delay systems, a fuzzy sliding-mode control is presented [33]. While all of the above studies have made progress and achieved good results, the problem of an "explosion of complexity" needs to be further investigated. Moreover, the tracking control of FJ robot systems with input saturation also needs to be taken into account in practice.

    Inspired by the above research, we present an adaptive fuzzy tracking control system for single-link FJ robots by combining the backstepping technique with DSC design. In order to solve the issue of input saturation that is nonsmooth in the system, we choose a smooth function to approximate it such that the backstepping technique can be utilized to design the controller. FLSs are applied to approximate unknown continuous functions. Adaptive DSC design is used to overcome the problem of the "explosion of complexity" by introducing auxiliary first-order filters. Lyapunov stability analysis shows that the uniform boundedness of all signals within the closed-loop system is guaranteed. The main contributions of this paper could be summarized in two parts:

    1) An adaptive fuzzy DSC control strategy is presented for single-link flexible-joint robot systems with input saturation;

    2) Only one adaptive law is designed by estimating the maximum value of the square of the weight vector norm. In this way, the structure of the designed controller is simple and the computational burden is significantly reduced.

    It is shown that the proposed control scheme can guarantee the semi-global uniform ultimate boundedness of all signals in the closed-loop system. A comparison is conducted to show that the improved DSC method can reduce the computational burden by almost two thirds in comparison with the standard DSC method.

    The remainder of this paper is organized as follows. Problem formulation and preliminaries are given in Section Ⅱ. Adaptive fuzzy control design using DSC and stability analysis is presented in Section Ⅲ. Section Ⅳ provides simulation results to show the effectiveness of the proposed scheme. The conclusions are included in Section Ⅴ.

    Throughout this paper, ~()=()^(), where ^() is the estimate of (), ()T denotes transposition of a vector, denotes the two-norm.

    According to the Euler-Lagrangian equations, the single-link FJ robot dynamics and actuator dynamics can be obtained as follows [5], [36]:

    ML2¨q+MgLsinq+F(˙q)+Kq=Kqm (1)
    J¨qm+B˙qm+K(qmq)=u(v(t)) (2)

    where q, ˙q, ¨qR denote the link position, velocity, and acceleration, respectively. M, g, LR are the mass of the link, gravity acceleration, and the length of the link, respectively. qm, ˙qm, ¨qmR denote the rotor angular position, velocity, and acceleration, respectively. K, J, BR represent the coefficients of joint stiffness, the joint flexibility, and the natural damping, respectively, where B is unknown. F(˙q)R represents friction. u(v(t)) denotes the characteristic of input saturation. v(t)R is the control input to be designed of the input saturation. u() denotes the plant input which is described by the following equation

    u(v(t))=sat(v(t))={sign(v(t))uM, |v(t)|uMv(t),|v(t)|<uM (3)

    where uM denotes the bound of the plant input u(t). It can be observed that the plant input u(t) is not smooth; as the result, the backstepping technique can not be utilized directly. In such a situation, we adopt a smooth function to approximate the saturation characteristic as below:

    g(v)=uMtanh(vuM)=uMevuMevuMevuM+evuM (4)

    with ρ(v) being the approximation error. Then, the u(v(t)) in (3) can be written as

    u(v(t))=g(v)+ρ(v)=uMtanh(vuM)+ρ(v). (5)

    Since the sat function in (3) and tanh function in (4) are both bounded functions, ρ(v) in (5) is also a bounded function and its bound is

    |ρ(v)|=|sat(v(t))g(v)|uM(1tanh(1))=D1. (6)

    For the convenience of controller design, we utilize the mean-value theorem to tackle the function g(v).

    g(v)=g(v0)+g(v)v|v=vμ(vv0) (7)

    where vμ=μv+(1μ)v0 with 0<μ<1.

    By choosing v0=0, one can obtain

    g(v)=g(0)+g(v)v|v=μvv. (8)

    By initializing g(0)=0 and using the above equation, we can have the following result

    g(v)=uMtanh(vuM)=g(v)v|v=μvv. (9)

    Letting h(v)=g(v)v|v=μv, the following equation satisfies

    g(v)=h(v)v=(uM(1tanh2(vuM))1uM)|v=μvv=(1tanh2(vuM))|v=μvv. (10)

    Since the tanh function and control input v in (10) are both bounded, h(v) in (10) is also a bounded function and its bound is

    0<hmh(v)1. (11)

    To facilitate design, the following assumptions are needed.

    Assumption 1 [37]: The desired signal yd is continuous and available, and [yd,˙yd,¨yd]TΩd with the known compact set Ωd={[yd,˙yd,¨yd]T:y2d+˙y2d+¨y2dB0}R3, where B0 is a known positive constant.

    Assumption 2: The minimum value of h(v) in (11) is known. That is, the constant hm is known and we can use it in the process of design.

    Remark 1: Since the control input v we will design later is always bounded in the actual system, there exists a minimum value for h(v) and Assumption 2 is reasonable in the actual system. On the other hand, the boundary value of control input v can be obtained when the input saturation is not considered into. According to hm=1tanh2(ˉvmax, the constant h_m can be obtained in the simulation, where \bar{v}_{\max} denotes the maximum value of v in the absence of input saturation, and u_M denotes the bound of the plant input u(t). The detailed explanation about h_m is illustrated in appendix.

    Based on the above discussion and analysis, we transfer the system (1) and (2) with new state variables. Let x_1=q, \ x_2=\dot{q}, \ x_3=q_m and x_4=\dot{q}_m. Then, the robot dynamics and actuator dynamics of (1) and (2) can be described as follows:

    \begin{align*} \left\{ \begin{array}{lcl} \dot{x}_1=x_2\\ \dot{x}_2=\frac{1}{ML^2}\big(Kx_3-Kx_1-F(x_2)-MgL\sin x_1\big)\\ \dot{x}_3=x_4\\ \dot{x}_4=J^{-1}\big(h(v)v+\rho(v)-Bx_4-K(x_3-x_1)\big)\\ \ y\, =x_1. \end{array}\right. \end{align*} (12)

    The control objective is to design an adaptive fuzzy controller v for the single-link FJ robot system to guarantee that the system output y can track the desired trajectory y_d, and all the signals in the closed-loop system are bounded.

    In this paper, we will utilize FLSs to approximate the unknown function f(x) and the results of approximation can be represented as

    \begin{equation*} f(x)=W^{*T}\Psi(x) \end{equation*} (13)

    where W^*=[w^*_1, \dots, w^*_N]^T, \Psi(x)=[\psi_1(x), \dots, \psi_N(x)]^T.

    Only one adaptive parameter needs to be updated in this paper, which reduces the computational burden significantly. The parameter \theta is defined as

    \begin{equation*} \theta={\rm max}\left\{\left\rVert W^*_i\right\rVert^2\right\}, \ \ i=1, 2. \end{equation*} (14)

    Lemma 1 [31]: For a continuous unknown function f(x) defined on a compact set \Omega, there exists a fuzzy system W^T\Psi(x) such that for any given constant \varepsilon>0, the following inequality holds

    \begin{equation*} \sup\limits_{x\in\Omega}\left\vert f(x)-W^T\Psi(x)\right\vert\le\varepsilon. \end{equation*} (15)

    In this section, we will attempt to design an adaptive DSC controller with FLSs for the single-link FJ system. The design of adaptive DSC is based on the following change of coordinates:

    \begin{align*} \left\{ \begin{array}{lcl} s_1=x_1-y_d\\ s_i=x_i-\omega_i, \ i=2, 3, 4 \end{array}\right. \end{align*} (16)

    where x_1, x_i, y_d, \omega_i\in \mathbb{R}, i=2, 3, 4, \omega_i is the output of the first-order filter with \alpha_{i-1} as the input, and \alpha_{i-1} is an intermediate control. The first-order filter with time design constant \tau_i is defined as

    \begin{equation*} \tau_i\dot{\omega}_i+\omega_i=\alpha_{i-1}, \ \omega_i(0)=\alpha_{i-1}(0), \ i=2, 3, 4 \end{equation*} (17)

    with the filter errors

    \begin{equation*} y_i=\omega_i-\alpha_{i-1}, \ i=2, 3, 4. \end{equation*} (18)

    The recursive design procedure contains 4 steps. At each recursive step i, the FLSs may be used to approximate unknown system dynamics, and then the virtual control law \alpha_i and the first-order filter are designed.

    Step 1: Considering the first equation of system (12) using (16) and (18) gives

    \begin{equation*} \dot{s}_1=\dot{x}_1-\dot{y}_d=x_2-\dot{y}_d=s_2+y_2+\alpha_1-\dot{y}_d. \end{equation*} (19)

    Choose a Lyapunov function candidate as

    \begin{align*} V_{s_1}=\frac{1}{2}s^2_1. \end{align*} (20)

    Subsequently, taking the time derivative of V_{s_1} and using (19) yields

    \begin{align*} \dot{V}_{s_1}=s_1\dot{s}_1=s_1(s_2+y_2+\alpha_1-\dot{y}_d). \end{align*} (21)

    Next, construct the virtual control law \alpha_1 as

    \begin{equation*} \alpha_1=-k_1s_1+\dot{y}_d \end{equation*} (22)

    where k_1 is a positive design parameter. Substituting (22) into (21) yields

    \begin{align*} \dot{V}_{s_1}=-k_1s^2_1+s_1s_2+s_1y_2. \end{align*} (23)

    Using Young's inequality, the following inequalities hold

    s_1s_2\le s^2_1+\frac{1}{4}s^2_2 (24)
    s_1y_2\le s^2_1+\frac{1}{4}y^2_2. (25)

    Then, substituting (24) and (25) into (23) produces

    \begin{align*} \dot{V}_{s_1}\le -(k_1-2)s^2_1+\frac{1}{4}s^2_2+\frac{1}{4}y^2_2. \end{align*} (26)

    Let \alpha_1 pass through a first-order filter with the filter time design constant \tau_2 to introduce a new filter variable \omega_2 as

    \begin{align*} \tau_2\dot{\omega}_2+\omega_2=\alpha_1, \ \omega_2(0)=\alpha_1(0). \end{align*} (27)

    Based on y_2=\omega_2-\alpha_1, \dot{\omega}_2=\frac{\alpha_1-\omega_2}{\tau_2}=-\frac{y_2}{\tau_2} is satisfied. And then the time derivative of y_2 can be obtained as

    \begin{align*} \dot{y}_2&=\dot{\omega}_2-\dot{\alpha}_1=-\frac{y_2}{\tau_2}+(k_1\dot{s}_1-\ddot{y}_d). \end{align*} (28)

    Define \vert k_1\dot{s}_1-\ddot{y}_d\vert=\vert -\dot{\alpha}_1\vert\le\zeta_2(\dot{s}_1, \ddot{y}_d), where the continuous function \zeta_2(\dot{s}_1, \ddot{y}_d)=\zeta_2(s_1, s_2, y_2, y_d, \dot{y}_d, \ddot{y}_d). Further, we have

    \begin{align*} \left\vert \dot{y}_2+\frac{y_2}{\tau_2}\right\vert\le \zeta_2(s_1, s_2, y_2, y_d, \dot{y}_d, \ddot{y}_d). \end{align*} (29)

    For clarity of design, \zeta_2(s_1, s_2, y_2, y_d, \dot{y}_d, \ddot{y}_d) is abbreviated as \zeta_2 in the following design. By combining (28) with (29) and using the Young's inequality, one has

    \begin{align*} y_2\dot{y}_2&\le-\frac{y^2_2}{\tau_2}+\vert y_2\vert\zeta_2\le -\frac{y^2_2}{\tau_2}+y^2_2+\frac{1}{4}\zeta^2_2. \end{align*} (30)

    Next, we construct the following Lyapunov function as

    \begin{align*} V_1=V_{s_1}+\frac{1}{2}y^2_2. \end{align*} (31)

    Combining (26) with (30), the time derivative of V_1 can be gained as

    \begin{align*} \dot{V_1}\le-(k_1-2)s^2_1+\frac{1}{4}s^2_2-\Big(\frac{1}{\tau_2}-\frac{5}{4}\Big)y^2_2+\frac{1}{4}\zeta^2_2. \end{align*} (32)

    Step 2: The time derivative of s_2 is

    \begin{align*} \dot{s}_2&=\frac{1}{ML^2}(Kx_3-Kx_1-F(x_2)-MgL\sin x_1)-\dot{\omega}_2\\ &=\frac{Kx_3}{ML^2}+\frac{-Kx_1-F(x_2)-MgL\sin x_1}{ML^2}-\dot{\omega}_2. \end{align*} (33)

    In view of (16) and (18), (33) is rewritten as

    \begin{align*} \dot{s}_2&=\frac{K}{ML^2}(s_3+y_3+\alpha_2)+\frac{1}{ML^2}(-Kx_1-F(x_2)\\ &\quad\, -MgL\sin x_1)-\dot{\omega}_2. \end{align*} (34)

    Define

    \begin{equation*} f_1=\frac{1}{ML^2}(-Kx_1-F(x_2)-MgL\sin x_1)-\dot{\omega}_2 \end{equation*} (35)

    where f_1 denotes a unknown continuous function, which can be approximated by using FLSs later. Then, (34) can be rewritten as

    \begin{equation*} \dot{s}_2=\frac{K}{ML^2}(s_3+y_3)+\frac{K}{ML^2}\alpha_2+f_1. \end{equation*} (36)

    Construct the following Lyapunov candidate function based on the virtual error and the adaptive parameter

    \begin{align*} V_{s_2}=\frac{1}{2}s^2_2+\frac{1}{2\gamma}\tilde{\theta}^2 \end{align*} (37)

    where \tilde{\theta}=\theta-\hat{\theta}, \hat{\theta} is the estimate of \theta which is defined in (14) and \gamma is a positive design constant. Taking the time derivative of V_{s_2} yields

    \begin{align*} \dot{V}_{s_2}=s_2\dot{s}_2-\frac{1}{\gamma}\tilde{\theta}\dot{\hat{\theta}}. \end{align*} (38)

    Substituting (36) into (38) produces

    \begin{align*} \dot{V}_{s_2}&=s_2\frac{K}{ML^2}(s_3+y_3)+s_2\frac{K}{ML^2}\alpha_2+s_2f_1-\frac{1}{\gamma}\tilde{\theta}\dot{\hat{\theta}}. \end{align*} (39)

    Next, a fuzzy system W^T_1\Psi_1(X_1) is employed to approximate f_1 such that for any given constant \varepsilon_1>0, f_1 is written as

    \begin{align*} f_1=W^T_1\Psi_1(X_1)+\delta_1(X_1), \ \vert \delta_1(X_1)\vert\le\varepsilon_1 \end{align*} (40)

    where \delta_1(X_1) is the approximation error with \varepsilon_1 being the bound of \delta_1(X_1). Combining (40) with (39) gives

    \begin{align*} \dot{V}_{s_2}&=s_2\frac{K}{ML^2}(s_3+y_3)+s_2\frac{K}{ML^2}\alpha_2+s_2W^T_1\Psi_1(X_1)\\ &\quad\, +s_2\delta_1(X_1)-\frac{1}{\gamma}\tilde{\theta}\dot{\hat{\theta}}. \end{align*} (41)

    Using the Young's inequality, the following inequalities hold

    s_2W^T_1\Psi_1(X_1)\le\frac{s^2_2\theta\Psi^T_1(X_1)\Psi_1(X_1)}{2a^2_1}+\frac{a^2_1}{2} (42)
    s_2\delta_1(X_1)\le\frac{s^2_2}{2}+\frac{\varepsilon^2_1}{2}. (43)

    Substituting (42) and (43) into (41) yields

    \begin{align*} \dot{V}_{s_2}&=s_2\frac{K}{ML^2}(s_3+y_3)+\frac{s^2_2\theta\Psi^T_1(X_1)\Psi_1(X_1)}{2a^2_1}\\ &\quad\, +s_2\frac{K}{ML^2}\alpha_2+\frac{a^2_1}{2}+\frac{s^2_2}{2}+\frac{\varepsilon^2_1}{2}-\frac{1}{\gamma}\tilde{\theta}\dot{\hat{\theta}}. \end{align*} (44)

    Now, a virtual control signal can be chosen as

    \begin{align*} \alpha_2=\frac{ML^2}{K}\left(-k_2s_2-\frac{s_2\hat{\theta}\Psi^T_1(X_1)\Psi_1(X_1)}{2a^2_1}\right) \end{align*} (45)

    where \hat{\theta} is the estimate of \theta. The adaptive law will be designed in the last step. Combining (45) with (44) gives

    \begin{align*} \dot{V}_{s_2}&=\frac{s_2K(s_3+y_3)}{ML^2}+\frac{s^2_2\theta\Psi^T_1(X_1)\Psi_1(X_1)}{2a^2_1}+\frac{a^2_1}{2}+\frac{s^2_2}{2}\\ &\quad\, +\frac{\varepsilon^2_1}{2}-k_2s^2_2-\frac{s^2_2\hat{\theta}\Psi^T_1(X_1)\Psi_1(X_1)}{2a^2_1}-\frac{1}{\gamma}\tilde{\theta}\dot{\hat{\theta}}\\ &=s_2\frac{K}{ML^2}(s_3+y_3)-k_2s^2_2+\frac{a^2_1}{2}+\frac{\varepsilon^2_1}{2}+\frac{s^2_2}{2}\\ &\quad\, +\frac{1}{\gamma}\tilde{\theta}\left(\frac{\gamma s^2_2\Psi^T_1(X_1)\Psi_1(X_1)}{2a^2_1}-\dot{\hat{\theta}}\right). \end{align*} (46)

    By using the Young's inequality, we have

    \frac{s_2Ks_3}{ML^2}\le\frac{s^2_2K^2}{2ML^2}+\frac{s^2_3}{2ML^2} (47)
    \frac{s_2Ky_3}{ML^2}\le\frac{s^2_2K^2}{2ML^2}+\frac{y^2_3}{2ML^2}. (48)

    Combining (47) and (48) with (46) yields

    \begin{align*} \dot{V}_{s_2}&\le-k_2s^2_2+\frac{s^2_2K^2}{ML^2} +\frac{s^2_2}{2}\\ &\quad\, +\frac{1}{\gamma}\tilde{\theta}\left(\frac{\gamma s^2_2\Psi^T_1(X_1)\Psi_1(X_1)}{2a^2_1}-\dot{\hat{\theta}}\right)\\ &\quad\, +\frac{s^2_3}{2ML^2}+\frac{y^2_3}{2ML^2}+\frac{a^2_1}{2}+\frac{\varepsilon^2_1}{2}\\ &=-\left(k_2-\frac{K^2}{ML^2}-\frac{1}{2}\right)s^2_2+\frac{s^2_3}{2ML^2}+\frac{y^2_3}{2ML^2}\\ &\quad\, +\frac{1}{\gamma}\tilde{\theta}\left(\frac{\gamma s^2_2\Psi^T_1(X_1)\Psi_1(X_1)}{2a^2_1}-\dot{\hat{\theta}}\right)+\frac{a^2_1}{2}+\frac{\varepsilon^2_1}{2}. \end{align*} (49)

    Next, we let \alpha_2 pass through a first-order filter with the filter time design constant \tau_3 to introduce a new filter variable \omega_3 as

    \begin{align*} \tau_3\dot{\omega}_3+\omega_3=\alpha_2, \ \omega_3(0)=\alpha_2(0). \end{align*} (50)

    Based on y_3=\omega_3-\alpha_2, \dot{\omega}_3=\frac{\alpha_2-\omega_3}{\tau_3}=-\frac{y_3}{\tau_3} is satisfied. And then the time derivative of y_3 can be obtained as

    \begin{align*} \dot{y}_3&=-\frac{y_3}{\tau_3}+\frac{ML^2}{K}\Big(k_2\dot{s}_2+\frac{\dot{s}_2\hat{\theta}\Psi^T_1(X_1)\Psi_1(X_1)}{2a^2_1}\\ &\quad\, +\frac{s_2\dot{\hat{{\theta}}}\Psi^T_1(X_1)\Psi_1(X_1)}{2a^2_1}+\frac{s_2\hat{\theta}\Psi^T_1(X_1)\dot{\Psi}_1(X_1)}{a^2_1}\Big). \end{align*} (51)

    By following the similar process to step 1, we define \vert -\dot{\alpha}_2\vert\le\zeta_3(s_1, s_2, s_3, y_2, y_3, \hat{\theta}, y_d, \dot{y}_d, \ddot{y}_d). Further, one has

    \begin{align*} \left\vert \dot{y}_3+\frac{y_3}{\tau_3}\right\vert\le \zeta_3(s_1, s_2, s_3, y_2, y_3, \hat{\theta}, y_d, \dot{y}_d, \ddot{y}_d). \end{align*} (52)

    Let \zeta_3 represent \zeta_3(s_1, s_2, s_3, y_2, y_3, \hat{\theta}, y_d, \dot{y}_d, \ddot{y}_d) for clarity. Combining (51) with (52) and using Young's inequality yields

    \begin{align*} y_3\dot{y}_3&\le-\frac{y^2_3}{\tau_3}+\vert y_3\vert\zeta_3\le -\frac{y^2_3}{\tau_3}+y^2_3+\frac{1}{4}\zeta^2_3. \end{align*} (53)

    Next, we construct the following Lyapunov function as

    \begin{align*} V_2=V_{s_2}+\frac{1}{2}y^2_3. \end{align*} (54)

    Using (49) and (53), the time derivative of V_2 can be obtained as

    \begin{align*} \dot{V_2}&\le\dot{V}_{s_2}-\frac{y^2_3}{\tau_3}+y^2_3+\frac{1}{4}\zeta^2_3\\ &\le-\left(k_2-\frac{K^2}{ML^2}-\frac{1}{2}\right)s^2_2+\frac{s^2_3}{ML^2}\\ &\quad\, +\frac{1}{\gamma}\tilde{\theta}\left(\frac{\gamma s^2_2\Psi^T_1(X_1)\Psi_1(X_1)}{2a^2_1}-\dot{\hat{\theta}}\right)\\ &\quad\, -\left(\frac{1}{\tau_3}-1-\frac{1}{2ML^2}\right)y^2_3+\frac{a^2_1}{2}+\frac{\varepsilon^2_1}{2}+\frac{1}{4}\zeta^2_3. \end{align*} (55)

    Step 3: Combining (16) with (18), the time derivative of s_3 can be obtained as

    \begin{equation*} \dot{s}_3=x_4-\dot{\omega}_3=s_4+y_4+\alpha_3-\dot{\omega}_3. \end{equation*} (56)

    Design the following Lyapunov candidate function as

    \begin{align*} V_{s_3}=\frac{1}{2}s^2_3. \end{align*} (57)

    By taking the time derivative of V_{s_3}, we get the following result

    \begin{align*} \dot{V}_{s_3}=s_3\dot{s}_3=s_3(s_4+y_4+\alpha_3-\dot{\omega}_3). \end{align*} (58)

    We then choose the virtual control law \alpha_3 as

    \begin{equation*} \alpha_3=-k_3s_3+\dot{\omega}_3 \end{equation*} (59)

    where k_3 is a positive design parameter. Substituting (59) into (58) yields

    \begin{align*} \dot{V}_{s_3}=-k_3s^2_3+s_3s_4+s_3y_4. \end{align*} (60)

    Using Young's inequality, the following two inequalities hold

    s_3s_4\le s^2_3+\frac{1}{4}s^2_4 (61)
    s_3y_4\le s^2_3+\frac{1}{4}y^2_4. (62)

    By substituting (61) and (62) into (60), we have

    \begin{align*} \dot{V}_{s_3}\le -(k_3-2)s^2_3+\frac{1}{4}s^2_4+\frac{1}{4}y^2_4. \end{align*} (63)

    Next, let \alpha_3 pass through a first-order filter with the design constant \tau_4 to introduce a new filter variable \omega_4 as

    \begin{align*} \tau_4\dot{\omega}_4+\omega_4=\alpha_3, \ \omega_4(0)=\alpha_3(0). \end{align*} (64)

    Repeating steps similar to (50) and (51) yields

    \begin{align*} \dot{y}_4&=\dot{\omega}_4-\dot{\alpha}_3=-\frac{y_4}{\tau_4}+(k_3\dot{s}_3-\ddot{\omega}_3). \end{align*} (65)

    We define \vert k_3\dot{s}_3-\ddot{\omega}_3\vert=\vert -\dot{\alpha}_3\vert\le\zeta_4(\bar{s}_4, \bar{y}_4, \hat{\theta}, y_d, \dot{y}_d, \ddot{y}_d), where \bar{s}_4=[s_1, s_2, s_3, s_4], \bar{y}_4=[y_2, y_3, y_4]. Based on (65), one has

    \begin{align*} \left\vert \dot{y}_4+\frac{y_4}{\tau_4}\right\vert\le \zeta_4(\bar{s}_4, \bar{y}_4, \hat{\theta}, y_d, \dot{y}_d, \ddot{y}_d). \end{align*} (66)

    Simplify \zeta_4(\bar{s}_4, \bar{y}_4, \hat{\theta}, y_d, \dot{y}_d, \ddot{y}_d) as \zeta_4 for the convenience of design. Combining (65) with (66) and using Young's inequality gives

    \begin{align*} y_4\dot{y}_4&\le-\frac{y^2_4}{\tau_4}+\vert y_4\vert\zeta_4\le -\frac{y^2_4}{\tau_4}+y^2_4+\frac{1}{4}\zeta^2_4. \end{align*} (67)

    Next, we construct the following Lyapunov function as

    \begin{align*} V_3=V_{s_3}+\frac{1}{2}y^2_4. \end{align*} (68)

    Combining (63) with (67), the time derivative of V_3 can be obtained as

    \begin{align*} \dot{V}_3\le-(k_3-2)s^2_3+\frac{1}{4}s^2_4-\left(\frac{1}{\tau_4}-\frac{5}{4}\right)y^2_4+\frac{1}{4}\zeta^2_4. \end{align*} (69)

    Step 4: The time derivative of s_4 is

    \begin{align*} \dot{s}_4&=J^{-1}\Big(h(v)v+\rho(v)-Bx_4-K(x_3-x_1)\Big)-\dot{\omega}_4. \end{align*} (70)

    Choose a Lyapunov function as

    \begin{align*} V_4=\frac{1}{2}s^2_4. \end{align*} (71)

    Then, combining (70) with (71) and using the Young's inequality, the time derivative of V_4 can be obtained as

    \begin{align*} \dot{V}_4&=s_4\Big(J^{-1}\Big(h(v)v+\rho(v)-Bx_4-K(x_3-x_1)\Big)-\dot{\omega}_4\Big)\\ &\le s_4J^{-1}h(v)v+\frac{s^2_4}{2}+\frac{D^2_1}{2J^2}+s_4f_2 \end{align*} (72)

    where f_2=J^{-1}\Big(-Bx_4-K(x_3-x_1)\Big)-\dot{\omega}_4 denotes a continuous function. Then, a fuzzy system W^T_2\Psi_2(X_2) is employed to approximate f_2 such that for any given \varepsilon_2>0, f_2 is written as

    \begin{align*} f_2=W^T_2\Psi_2(X_2)+\delta_2(X_2), \ \vert \delta_2(X_2)\vert\le\varepsilon_2 \end{align*} (73)

    where \delta_2(X_2) is the approximation error with \varepsilon_2 being the bound of \delta_2(X_2). Substituting (73) into (72), using Young's inequality, and repeating the steps similar to (42) and (43) yields

    \begin{align*} \dot{V}_4&\le s_4J^{-1}h(v)v+\frac{s^2_4}{2}+\frac{D^2_1}{2J^2}+s_4W^T_2\Psi_2(X_2)+s_4\delta_2(X_2)\\ &\le s_4J^{-1}h(v)v+\frac{s^2_4\theta\Psi^T_2(X_2)\Psi_2(X_2)}{2a^2_2}\\ &\quad\, +\frac{D^2_1}{2J^2}+\frac{a^2_2}{2}+\frac{\varepsilon^2_2}{2}+s^2_4. \end{align*} (74)

    Now, we are ready to design the actual controller v as

    \begin{align*} v=\frac{J}{h_m}\left(-k_4s_4-\frac{s_4\hat{\theta}\Psi^T_2(X_2)\Psi_2(X_2)}{2a^2_2}\right). \end{align*} (75)

    Remark 2: It is worth noting that we utilize FLSs to approximate the unknown continuous function f_1 and known continuous function f_2 in Step 2 and this step, respectively. It is better not to use fuzzy logic systems to estimate the known continuous function f_2 in this step because of the existence of approximation error. However, the impact of this processing on system performance is negligible for because FLSs can make very accurate approximations. The main reason for this is to make the design similar to that in Step 2 and to make the form of the controller simple.

    Substituting (75) into (74) yields

    \begin{align*} \dot{V}_4&\le s_4J^{-1}h(v)\Big(\frac{J}{h_m}\Big(-k_4s_4-\frac{s_4\hat{\theta}\Psi^T_2(X_2)\Psi_2(X_2)}{2a^2_2}\Big)\Big)\\ &\quad\, +\frac{s^2_4\theta\Psi^T_2(X_2)\Psi_2(X_2)}{2a^2_2}+\frac{D^2_1}{2J^2}+\frac{a^2_2}{2}+\frac{\varepsilon^2_2}{2}+s^2_4\\ &=-k_4\frac{h(v)}{h_m}s^2_4-\frac{h(v)}{h_m}\frac{s^2_4\hat{\theta}\Psi^T_2(X_2)\Psi_2(X_2)}{2a^2_2}\\ &\quad\, +\frac{s^2_4\theta\Psi^T_2(X_2)\Psi_2(X_2)}{2a^2_2}+\frac{D^2_1}{2J^2}+\frac{a^2_2}{2}+\frac{\varepsilon^2_2}{2}+s^2_4\\ &\le-\left(k_4-1\right)s^2_4+\frac{s^2_4\tilde{\theta}\Psi^T_2(X_2)\Psi_2(X_2)}{2a^2_2}\\ &\quad\, +\frac{a^2_2}{2}+\frac{\varepsilon^2_2}{2}+\frac{D^2_1}{2J^2}. \end{align*} (76)

    In order to facilitate stability analysis, consider the following Lyapunov function as

    \begin{align*} V&=V_1+V_2+V_3+V_4\\ &=\frac{s^2_1}{2}+\frac{s^2_2}{2}+\frac{s^2_3}{2}+\frac{s^2_4}{2}+\frac{\tilde{\theta}^2}{2\gamma}+\frac{y^2_2}{2}+\frac{y^2_3}{2}+\frac{y^2_4}{2}. \end{align*} (77)

    With the help of (32), (55), (69) and (76), the time derivative of V can be obtained as

    \begin{align*} \dot{V}&\le\frac{\tilde{\theta}}{\gamma}\Big(\frac{\gamma s^2_2\Psi^T_1(X_1)\Psi_1(X_1)}{2a^2_1}+\frac{\gamma s^2_4\Psi^T_2(X_2)\Psi_2(X_2)}{2a^2_2}-\dot{\hat{\theta}}\Big)\\ &\quad\, -\bar{k}_1s^2_1-\bar{k}_2s^2_2-\bar{k}_3s^2_3-\bar{k}_4s^2_4-\bar{\tau}_2y^2_2-\bar{\tau}_3y^2_3-\bar{\tau}_4y^2_4\\ &\quad\, +\frac{a^2_1}{2}+\frac{a^2_2}{2}+\frac{\varepsilon^2_1}{2}+\frac{\varepsilon^2_2}{2}+\frac{\zeta^2_2}{4}+\frac{\zeta^2_3}{4}+\frac{\zeta^2_4}{4}+\frac{D^2_1}{2J^2} \end{align*} (78)

    where \bar{k}_1=k_1-2, \bar{k}_2=k_2-\frac{K^2}{ML^2}-\frac{3}{4}, \bar{k}_3=k_3-2-\frac{1}{2ML^2}, \bar{k}_4=k_4-\frac{5}{4}, \bar{\tau}_2=\frac{1}{\tau_2}-\frac{5}{4}, \bar{\tau}_3=\frac{1}{\tau_3}-1-\frac{1}{2ML^2}, \bar{\tau}_4=\frac{1}{\tau_4}-\frac{5}{4}.

    For any B_0>0 and p>0, the sets \Omega_d=\{(y_d, \dot{y}_d, \ddot{y}_d):y^2_d+\dot{y}^2_d+\ddot{y}^2_d\le B_0\} and \Omega_i=\{[\bar{s}^T_i, \bar{y}^T_i, \hat{\theta}]^T:\sum^i_{j=1}V_j\le p\}, i=1, 2, 3, 4 are compact in \mathbb{R}^3 and \mathbb{R}^{2i}, respectively. Therefore, \zeta_j has a maximum T_j on \Omega_j\times\Omega_i, j=2, 3, 4.

    Subsequently, design the following adaptation law as

    \begin{align*} \dot{\hat{\theta}}\!=\!\gamma \bigg(\frac{s^2_2\Psi^T_1(X_1)\Psi_1(X_1)}{2a^2_1}\!+\!\frac{ s^2_4\Psi^T_2(X_2)\Psi_2(X_2)}{2a^2_2}\!-\!\hat{\theta}\bigg). \end{align*} (79)

    Now, the fuzzy DSC design procedure is complete. The main results are summarized as follows.

    Theorem 1: Consider the closed-loop flexible-joint robot system consisting of (1) and (2) with Assumptions 1 and 2, the virtual control signals (22), (45), and (59), the actual controller (75), and the adaptive law (79); Moreover, assume that the unknown continuous function f_1 and f_2 can be approximated by FLSs when the approximation errors are bounded. For any given bounded initial conditions, all the signals in the closed-loop system are bounded, and the output of the system can track the desired trajectory with the tracking error converging to a small domain around the origin.

    Proof: By using the relationship of \hat{\theta}=\theta-\tilde{\theta} and Young's inequality, one has

    \begin{equation*} \tilde{\theta}\hat{\theta}=\tilde{\theta}(\theta-\tilde{\theta})=-\tilde{\theta}^2+\tilde{\theta}\theta\le-\frac{1}{2}\tilde{\theta}^2+\frac{1}{2}\theta^2. \end{equation*} (80)

    Substituting (79) into (78) and using (80), the following design parameters are chosen to guarantee that (78) is stable:

    \begin{equation} \left\{ \begin{aligned} \alpha_0&\le\gamma\\ k_1&\ge2+\frac{\alpha_0}{2}\\ k_2&\ge\frac{3}{4}+\frac{K^2}{ML^2}+\frac{1}{2}\alpha_0\\ k_3&\ge 2+\frac{1}{2ML^2}+\frac{\alpha_0}{2}\\ k_4&\ge\frac{5}{4}+\frac{\alpha_0}{2}\\ \frac{1}{\tau_2}&\ge\frac{5}{4}+\frac{\alpha_0}{2}\\ \frac{1}{\tau_3}&\ge 1+\frac{1}{2ML^2}+\frac{\alpha_0}{2}\\ \frac{1}{\tau_4}&\ge\frac{5}{4}+\frac{\alpha_0}{2}. \end{aligned} \right. \end{equation} (81)

    Then, by substituting (81) into (78) and using (80), one can obtain the following result:

    \begin{equation*} \dot{V}\le-\alpha_0V+C \end{equation*} (82)

    where \alpha_0 is a positive design parameter, and C is a bounded constant which is defined as

    \begin{equation*} C=\frac{a^2_1}{2}+\frac{a^2_2}{2}+\frac{\varepsilon^2_1}{2}+\frac{\varepsilon^2_2}{2}+\frac{T^2_2}{4}+\frac{T^2_3}{4}+\frac{T^2_4}{4}+\frac{D^2_1}{2J^2}+\frac{\theta^2}{2}. \end{equation*} (83)

    Remark 3: The choice of design parameters follows the principles of the subsystem. For example, for the design of parameter k_1, equation (82) is the final form we expect. By combining equation (77) with (78), we have -\bar{k}_1s^2_1\le -\alpha_0\frac{s^2_1}{2}, and eventually we can obtain k_1\ge 2+\frac{\alpha_0}{2}. The other design parameters are the same as this. During the simulation process, we found that the design parameter k_2 has greatest influence on system performance and better performance can be obtained by selecting larger k_2. Furthermore, design parameters \tau_i~(i=2, 3, 4) need to be carefully chosen. If the selected parameter \tau_i is too large, it will cause the designed controller to oscillate. Conversely, if the selected parameter \tau_i is too small, it will result in a large controller output. Other design parameters such as a_1, a_2, h_m also need to be adjusted accordingly.

    Let \alpha_0>C/p, where a detailed description of p is made following (78). Then \dot{V}\le0 on V=p. That is, V\le p is a invariant set, i.e., if V(0)\le p, then V(t)\le p, for all t\ge0. By solving the inequality of (82), we can obtain

    \begin{equation*} 0\le V(t)\le \frac{C}{\alpha_0}+\Big[V(0)-\frac{C}{\alpha_0}\Big]e^{-\alpha_0t} \end{equation*} (84)

    which means that

    \begin{equation*} \lim \limits_{t\rightarrow\infty}V(t)\le \frac{C}{\alpha_0}. \end{equation*} (85)

    Therefore, all signals in the closed-loop system, i.e., s_i, y_j, \tilde{\theta} (i=1, 2, 3, 4, j=2, 3, 4) are semi-globally uniformly ultimately bounded. Since x_1=s_1+y_d, x_1 is also bounded from the boundedness of s_1 and y_d. The boundedness of x_2 follows from boundedness of \alpha_1(\dot{y}_d, s_1) and the fact that x_2=s_2+y_2+\alpha_1. Similarly, the boundedness of x_3 and x_4 can be obtained. The boundedness of \hat{\theta} can be ensured from the boundedness of \tilde{\theta} and \theta. From (75), because v is a function of s_4, \hat{\theta} and \Psi_2(X_2), which are all bounded, v is also bounded. Moreover, by appropriately choosing the design parameters, such as increasing \gamma or decreasing a_1 , a_2, we ensure that the tracking error converges to a small neighborhood around 0 such that the output of the system can track the desired trajectory very well.

    To illustrate the effectiveness of the proposed approach, consider the following dynamics of the single-link FJ robot system:

    \begin{align*} \left\{ \begin{array}{lcl} ML^2\ddot{q}+MgL\sin q+F(\dot{q})+Kq=Kq_m\\ J\ddot{q}_m+B\dot{q}_m+K(q_m-q)=u(v(t)). \end{array}\right. \end{align*} (86)

    The friction term is selected as F(\dot{q})=B_1\cos\dot{q}. Table Ⅰ gives the system parameters and the desired reference trajectory is given as y_d=0.5\big(\sin(t)+\sin(0.5t)\big).

    Table  Ⅰ.  PARAMETERS OF THE SINGLE-LINK FJ ROBOT SYSTEM
    ParametersDescriptionValuesUnit
    Mthe mass of link0.25kg
    Lthe length of link1m
    ggravity acceleration9.8 \text{m}/\text{s}^2
    Kjoint stiffness2N\cdot \text{m}/\text{rad}
    Bdamping coefficient0.6none
    B_1friction coefficient0.01none
    Jjoint flexibility0.001 \text{m}/\text{s}^2
     | Show Table
    DownLoad: CSV

    For the FLSs, the number of the fuzzy rules N is selected as N=7 and the Gaussian function \psi_i(x) is chosen as \psi_i(x)=\exp[-(x-\mu_i)^T(x-\mu_i)/(\eta^2_i)], where the center vector \mu_i=[-3\ -2\ -1\ 0\ 1\ 2\ 3] and the width of the Gaussian function \eta_i=1, i=1, 2, \ldots, N.

    We choose design parameters as \alpha_0=\gamma=0.5, h_m=0.03, \tau_2=\tau_3=\tau_4=0.002, k_1=2.3, k_2=17, k_3=k_4=6, a_1=0.01, a_2=25, initial state values q=0.1, \dot{q}=0.2, q_m=0.5, \dot{q}_m=0.5, the initial values of the first-order filters are \omega_2=\omega_3=\omega_4=0.5, the boundary of the plant input is selected as u_M=2 and the initial value of the square for the weight vector norm is chosen as \theta=5.

    The results of the simulation are shown in Figs. 1-5. It can be seen from Figs. 1 and 2 that the system output q can track the desired trajectory y_d well. Fig. 3 shows the control input v and plant input u, respectively. We can observe that the plant input u satisfies \vert u\vert\le u_M, which indicates the effectiveness of our proposed approach. The response curve of adaptive parameter is illustrated in Fig. 4. Fig. 5 shows that the state signals are bounded.

    Figure  1.  Trajectories of the link position q and the desired trajectory y_d.
    Figure  2.  Tracking error for the single-link robot.
    Figure  3.  Control input v and plant input u of the single-link robot.
    Figure  4.  Estimation of the maximum value of the square for the vector norm of the weight.
    Figure  5.  Trajectories of link velocity \dot{q}, rotor angular position q_m and rotor angular velocity \dot{q}_m

    To further demonstrate the main contribution of our proposed control scheme, we apply the standard DSC method [37] which updates the weight vector to the same single-link FJ robot system in (86).

    Similar to the design procedure in Section Ⅲ, in order to save space, we only give the final design results. The virtual controller and actual controller are chosen as

    \begin{equation} \left\{ \begin{aligned} \alpha_1&=-k_1s_1+\dot{y}_d\\ \alpha_2&=-k_2s_2-\frac{ML^2}{K}\hat{W}^T_1\Psi_1(X_1)\\ \alpha_3&=-k_3s_3+\dot{\omega}_3\\ v&=-k_4s_4-\frac{Js_4}{h^2_mu_M}-\hat{W}^T_2\Psi_2(X_2) \end{aligned} \right. \end{equation} (87)

    and the adaptive law is designed as

    \begin{equation} \left\{ \begin{aligned} \dot{\hat{W}}_1&=\Gamma_1[s_2\frac{ML^2}{K}\Psi_1(X_1)-\sigma_1\hat{W}_1]\\ \dot{\hat{W}}_2&=\Gamma_2[s_4\Psi_2(X_2)-\sigma_2\hat{W}_2]. \end{aligned} \right. \end{equation} (88)

    For fairness of comparison, we use the same system parameters and the same initial conditions in simulation. FLSs and the desired reference signal are the same as before. Design parameters are chosen as \Gamma_1=\Gamma_2=\text{diag}(0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5), \sigma_1=\sigma_2=0.1, k_1=2.5, k_2=5, k_3=2.5, k_4=3, and \tau_2=\tau_3=\tau_4=0.02. Other design parameters are consistent with the above design parameters. The initial values of the weight vector are chosen as W_1=W_2=[0.5;0.5;0.5;0.5;0.5;0.5;0.5].

    The results of the simulation using standard DSC which updates the weight vector are shown in Figs. 6 and 7. Considering the running time of the program, the simulation time is set to 20 s since long simulation time will consume more time than the improved DSC scheme. Trajectories of the link position q and the desired trajectory y_d using the standard DSC method are shown in Figs. 6 and 7, which illustrate the control input v and plant input u of the single-link robot. It is obvious that the control performance of our proposed improved DSC scheme is better than the standard DSC method comparing Figs. 1-3 with Figs. 6 and 7.

    Figure  6.  Trajectories of the link position q and the desired trajectory y_d by using standard DSC method.
    Figure  7.  Control input v and plant input u of the single-link robot by using standard DSC method.

    Remark 4: It should be noted that the objective of comparison experiments is to illustrate the fact that only one adaptation parameter can reduce the computational burden significantly. The time complexity of the two control schemes rather than the control performance is the main concern here.

    The solving time and total running time of the improved DSC scheme, which updates the maximum value of weight vector norm, and standard DSC method, which updates the weight vector, are shown in Table Ⅱ and Table Ⅲ.

    Table  Ⅱ.  STANDARD DSC TIME CONSUMING VS. IMPROVED DSC TIME CONSUMING: SOLVING TIME (UNIT: S)
    Simulation timeImproved DSCStandard DSC
    5010.729
    408.622.6
    306.516.9
    204.511.5
    102.25.8
    51.12.9
     | Show Table
    DownLoad: CSV
    Table  Ⅲ.  IMPROVED DSC TIME CONSUMING VS. STANDARD DSC TIME CONSUMING: TOTAL RUNNING TIME (UNIT: S)
    Simulation timeImproved DSCStandard DSC
    5011 \backslash
    408.7 \backslash
    306.7 \backslash
    204.5452
    102.3107
    51.226
     | Show Table
    DownLoad: CSV

    Remark 5: The solving time and total running time are obtained under the condition that the program is run three times and averaged. The program runs on MATLAB R2014a, which is installed on a Win 10 OS equipped with 4 GB RAM and an Intel Core i5-6300HQ processor. The solving time refers to the time it takes for the algorithm to solve the problem, which represents the computational complexity of the algorithm. Total running time includes solving time and subsequent processing time. In fact, we expect to shorten the solving time to reduce the computational burden, and we prefer to reduce the total running time to quickly select design parameters to optimize control performance.

    It can be observed from Table Ⅱ that the improved DSC method is almost three times faster than the standard DSC method in terms of simulation time. We can see that the total running time of the improved DSC method is much less than standard DSC method from Table Ⅲ. With the increase of simulation time, the time consumed also increases exponentially for the standard DSC method. Moreover, the advantage of the improved method is more obvious as system order and fuzzy rules increase. From the results of the comparison study, it can be concluded that the proposed scheme using only one adaptive parameter can reduce the computational burden significantly.

    This paper considers the tracking control problem of single-link flexible-joint robot system with input saturation using an adaptive fuzzy DSC, where the saturation problem is solved by a smooth function with the help of the mean-value theorem. The problem of an "explosion of complexity" is overcome by the adaptive DSC. The developed adaptive fuzzy control scheme can guarantee that all signals in the closed-loop system are uniformly ultimately bounded. The simulations on a single-link robot system show that the system output tracks the desired trajectory well under the input saturation. Comparison studies show that the proposed scheme can reduce the computational burden by nearly two-thirds in comparison with the standard method.

    It should be pointed out, however, that our proposed DSC scheme does not consider the errors caused by the introduction of filters, which may limit performance of the system. Future research may investigate this potential problem associated with the DSC approach.

    In (10), h(v) is defined as

    \begin{align*} g(v)=h(v)v&=\left(u_M\left(1-\tanh^2\left(\frac{v}{u_M}\right)\right)\frac{1}{u_M}\right)\Big\vert_{v=\mu v}v\\ &=\left(1-\tanh^2\left(\frac{v}{u_M}\right)\right)\Big\vert_{v=\mu v}v \end{align*} (A1)

    and

    \begin{align*} h(v)=\left(1-\tanh^2\left(\frac{v}{u_M}\right)\right)\Big\vert_{v=\mu v}. \end{align*} (A2)

    Next, we will discuss how to select the minimum value h_m of h(v) in simulation. The following two cases should be considered.

    Case 1: When v>0, there are still two cases should be considered. That is, when 0<v\le u_M, h_m can be determined as

    \begin{align*} h(v)&=\left(1-\tanh^2\left(\frac{v}{u_M}\right)\right)\Big\vert_{v=\mu v}\\ &\ge \left(1-\tanh^2\left(\frac{v}{u_M}\right)\right)\\ &\ge \left(1-\tanh^2\left(\frac{u_M}{u_M}\right)\right)\\ &=h_{m1} \end{align*} (A3)

    where h_{m1}=0.42. When v>u_M, h_m can be determined as

    \begin{align*} h(v)&=\frac{g(v)}{v}\ge \frac{g(v)_{\min}}{v^+_{\max}}=\frac{0.7616u_M}{v^+_{\max}}=h_{m2} \end{align*} (A4)

    where v^+_{\max} denotes the maximum value of input v under the condition that v is greater than zero in the simulation. In the actual controller, the input v is always bounded. Therefore, we can choose a bigger v_{\max} to determine h_{m2}.

    Case 2: When v<0, the results are similar to case 1. That is, when 0<v\le u_M, h_m can be determined as h_{m1}. When v>u_M, h_m can be determined as

    \begin{align*} h(v)&=\frac{\vert g(v)\vert}{\vert v\vert}\ge \frac{\vert g(v)\vert_{\min}}{\vert v^-\vert_{\max}}=\frac{0.7616u_M}{\vert v^-\vert_{\max}}=h_{m3} \end{align*} (A5)

    where \vert v^-\vert_{\max} denotes the maximum value of input v in the case of v less than zero in the simulation.

    It should be noted that the input v we discussed is carried out without input saturation. By combining with those two cases, we finally obtain

    \begin{align*} h_m=\min\{h_{m1}, h_{m2}, h_{m3}\}. \end{align*} (A6)
  • [1]
    T. K. Chang, A. Spowage, and K. Y. Chan, "Review of control and sensor system of flexible manipulator, " Journal of Intelligent & Robotic Systems, vol. 77, no. 1, pp. 187-213, 2015. http://www.wanfangdata.com.cn/details/detail.do?_type=perio&id=fb5cfc55d19d36e3412e9634fd4625fd
    [2]
    K. Melhem and W. Wang, "Global output tracking control of flexible joint robots via factorization of the manipulator mass matrix, " IEEE Transactions on Robotics, vol. 25, no. 2, pp. 428-437, 2009. doi: 10.1109/TRO.2009.2012016
    [3]
    J. R. Forbes and C. J. Damaren, "Design of optimal strictly positive real controllers using numerical optimization for the control of flexible robotic systems, " Journal of the Franklin Institute, vol. 348, no. 8, pp. 2191-2215, 2011. doi: 10.1016/j.jfranklin.2011.06.013
    [4]
    J. Kim and E. A. Croft, "Full-state tracking control for flexible joint robots with singular perturbation techniques, " IEEE Transactions on Control Systems Technology, doi: 10.1109/TCST.2017.2756962.
    [5]
    A. C. Huang and Y. C. Chen, "Adaptive sliding control for single-link flexible-joint robot with mismatched uncertainties, " IEEE Transactions on Control Systems Technology, vol. 12, no. 5, pp. 770-775, 2004. doi: 10.1109/TCST.2004.826968
    [6]
    A. D. Luca, B. Siciliano, and L. Zollo, "Pd control with on-line gravity compensation for robots with elastic joints: Theory and experiments, " Automatica, vol. 41, no. 10, pp. 1809-1819, 2005. doi: 10.1016/j.automatica.2005.05.009
    [7]
    J. S. Bang, H. Shim, K. P. Sang, and H. S.Jin, "Robust tracking and vibration suppression for a two-inertia system by combining backstepping approach with disturbance observer, " IEEE Transactions on Industrial Electronics, vol. 57, no. 9, pp. 3197-3206, 2010. doi: 10.1109/TIE.2009.2038398
    [8]
    S. K. Min and S. L. Jin, "Adaptive tracking control of flexible-joint manipulators without overparametrization, " Journal of Robotic Systems, vol. 21, no. 7, pp. 369-379, 2004. doi: 10.1002/(ISSN)1097-4563
    [9]
    W. Chang, Y. Li, and S. Tong, "Adaptive fuzzy backstepping tracking control for flexible robotic manipulator, " IEEE/CAA J. Autom. Sinica, doi: 10.1109/JAS.2017.7510886, 2017.
    [10]
    D. Swaroop, J. K. Hedrick, P. P. Yip, and J. C. Gerdes, "Dynamic surface control for a class of nonlinear systems, " IEEE Transactions on Automatic Control, vol. 45, no. 10, pp. 1893-1899, 2002. http://d.old.wanfangdata.com.cn/Periodical/kzyjc201202019
    [11]
    W. Si and X. Dong, "Adaptive neural dsc for nonlinear switched systems with prescribed performance and input saturation, " IEEE/CAA J. Autom. Sinica, doi: 10.1109/JAS.2017.7510661, 2017.
    [12]
    S. Su, "Output-feedback dynamic surface control for a class of nonlinear non-minimum phase systems, " IEEE/CAA J. Autom. Sinica, vol. 3, no. 1, pp. 96-104, 2016. doi: 10.1109/JAS.2016.7373767
    [13]
    H. Wang, P. X. Liu, and B. Niu, "Robust fuzzy adaptive tracking control for nonaffine stochastic nonlinear switching systems, " IEEE Transactions on Cybernetics, vol. 48, no. 8, pp. 2462-2471, 2018. doi: 10.1109/TCYB.2017.2740841
    [14]
    M. M. Polycarpou, "Stable adaptive neural control scheme for nonlinear systems, " IEEE Transactions on Automatic Control, vol. 41, no. 3, pp. 447-451, 2002. http://www.wanfangdata.com.cn/details/detail.do?_type=perio&id=79932ab111d6e2a3dbe6f3874f279c87
    [15]
    L. X. Wang and J. M. Mendel, "Fuzzy basis functions, universal approximation, and orthogonal least-squares learning, " IEEE Transactions on Neural Networks, vol. 3, no. 5, pp. 807-814, 1992. doi: 10.1109/72.159070
    [16]
    L. Jin, S. Li, H. M. La, and X. Luo, "Manipulability optimization of redundant manipulators using dynamic neural networks, " IEEE Transactions on Industrial Electronics, vol. 64, no. 6, pp. 4710-4720, 2017. doi: 10.1109/TIE.2017.2674624
    [17]
    Y. Zhang, S. Li, J. Gui, and X. Luo, "Velocity-level control with compliance to acceleration-level constraints: A novel scheme for manipulator redundancy resolution, " IEEE Transactions on Industrial Informatics, vol. 14, no. 3, pp. 921-930, 2018. doi: 10.1109/TII.2017.2737363
    [18]
    S. C. Tong, X. L. He, and H. G. Zhang, "A combined backstepping and small-gain approach to robust adaptive fuzzy output feedback control, " IEEE Transactions on Fuzzy Systems, vol. 17, no. 5, pp. 1059-1069, 2009. doi: 10.1109/TFUZZ.2009.2021648
    [19]
    M. Mirshekaran, F. Piltan, Z. Esmaeili, T. Khajeaian, and M. Kazeminasab, "Design sliding mode modified fuzzy linear controller with application to flexible robot manipulator, " International Journal of Modern Education & Computer Science, vol. 5, no. 10, pp. 53-63, 2013. http://cn.bing.com/academic/profile?id=907ce5894f0ad5df8127d450044cb20b&encoded=0&v=paper_preview&mkt=zh-cn
    [20]
    H. Chaoui, W. Gueaieb, M. Biglarbegian, and M. C. E. Yagoub, "Computationally efficient adaptive type-2 fuzzy control of flexible-joint manipulators, " Robotics, vol. 2, no. 2, pp. 66-91, 2013. doi: 10.3390/robotics2020066
    [21]
    X. Luo, M. Zhou, S. Li, Y. Xia, Z. H. You, Q. Zhu, and H. Leung, "Incorporation of efficient second-order solvers into latent factor models for accurate prediction of missing QoS data, " IEEE Transactions on Cybernetics, vol. 48, no. 4, pp. 1216-1228, 2018. doi: 10.1109/TCYB.2017.2685521
    [22]
    S. Li, Z. H. You, H. Guo, X. Luo, and Z. Q. Zhao, "Inverse-free extreme learning machine with optimal information updating, " IEEE Transactions on Cybernetics, vol. 46, no. 5, pp. 1229-1241, 2016. doi: 10.1109/TCYB.2015.2434841
    [23]
    P. Sun and Z. Yu, "Tracking control for a cushion robot based on fuzzy path planning with safe angular velocity, " IEEE/CAA J. Autom. Sinica, vol. 4, no. 4, pp. 610-619, 2017. doi: 10.1109/JAS.2017.7510607
    [24]
    X. Luo, J. Sun, Z. Wang, S. Li, and M. Shang, "Symmetric and nonnegative latent factor models for undirected, high dimensional and sparse networks in industrial applications, " IEEE Transactions on Industrial Informatics, vol. 13, no. 6, pp. 3098-3107, 2017. doi: 10.1109/TII.2017.2724769
    [25]
    X. Luo, M. Zhou, S. Li, Z. You, Y. Xia, and Q. Zhu, "A nonnegative latent factor model for large-scale sparse matrices in recommender systems via alternating direction method, " IEEE Transactions on Neural Networks & Learning Systems, vol. 27, no. 3, pp. 579-592, 2016. http://www.wanfangdata.com.cn/details/detail.do?_type=perio&id=1b3a26299d2a8a0d45fbd4e4d51f20d1
    [26]
    S. Li, M. C. Zhou, X. Luo, and Z. H. You, "Distributed winner-takeall in dynamic networks, " IEEE Transactions on Automatic Control, vol. 62, no. 2, pp. 577-589, 2017. doi: 10.1109/TAC.2016.2578645
    [27]
    H. Yang and J. Liu, "An adaptive rbf neural network control method for a class of nonlinear systems, " IEEE/CAA J. Autom. Sinica, vol. 5, no. 2, pp. 457-462, 2018. doi: 10.1109/JAS.2017.7510820
    [28]
    X. Luo, M. Zhou, Y. Xia, Q. Zhu, A. C. Ammari, and A. Alabdulwahab, "Generating highly accurate predictions for missing qos data via aggregating nonnegative latent factor models, " IEEE Transactions on Neural Networks & Learning Systems, vol. 27, no. 3, pp. 524-537, 2015. http://www.wanfangdata.com.cn/details/detail.do?_type=perio&id=9ea020b4a53845c6ba1596b8192872c5
    [29]
    Y. Yang, J. Tan, and D. Yue, "Prescribed performance control of onedof link manipulator with uncertainties and input saturation constraint, " IEEE/CAA J. Autom. Sinica, doi: 10.1109/JAS.2018.7511099, 2018.
    [30]
    Y. Li, S. Tong, and T. Li, "Hybrid fuzzy adaptive output feedback control design for uncertain mimo nonlinear systems with time-varying delays and input saturation, " IEEE Transactions on Fuzzy Systems, vol. 24, no. 4, pp. 841-853, 2016. doi: 10.1109/TFUZZ.2015.2486811
    [31]
    Y. Li, S. Tong, and T. Li, "Composite adaptive fuzzy output feedback control design for uncertain nonlinear strict-feedback systems with input saturation, " IEEE Transactions on Cybernetics, vol. 45, no. 10, pp. 2299-2308, 2015. doi: 10.1109/TCYB.2014.2370645
    [32]
    Q. Zhou, H. Li, C. Wu, L. Wang, and C. K. Ahn, "Adaptive fuzzy control of nonlinear systems with unmodeled dynamics and input saturation using small-gain approach, " IEEE Transactions on Systems Man & Cybernetics Systems, vol. 47, no. 8, pp. 1979-1989, 2017. http://ieeexplore.ieee.org/document/7572123/
    [33]
    X. Zhao, H. Yang, W. Xia, and X. Wang, "Adaptive fuzzy hierarchical sliding-mode control for a class of mimo nonlinear time-delay systems with input saturation, " IEEE Transactions on Fuzzy Systems, vol. 25, no. 5, pp. 1062-1077, 2017. doi: 10.1109/TFUZZ.2016.2594273
    [34]
    R. Bai, "Adaptive sliding-mode control of automotive electronic throttle in the presence of input saturation constraint, " IEEE/CAA J. Autom. Sinica, vol. 5, no. 4, pp. 878-884, 2018. doi: 10.1109/JAS.2018.7511147
    [35]
    Z. Fu, W. Xie, S. Rakheja, and J. Na, "Observer-based adaptive optimal control for unknown singularly perturbed nonlinear systems with input constraints, " IEEE/CAA J. Autom. Sinica, vol. 4, no. 1, pp. 48-57, 2017. doi: 10.1109/JAS.2017.7510322
    [36]
    S. J. Yoo, J. B. Park, and Y. H. Choi, Adaptive dynamic surface control of flexible-joint robots using self-recurrent wavelet neural networks, " IEEE Transactions on Systems Man & Cybernetics, Part B Cybernetics, vol. 36, no. 3, pp. 1342-1355, 2006. http://europepmc.org/abstract/MED/17186810
    [37]
    B. Ren, P. P. San, S. S. Ge, and H. L. Tong, "Adaptive dynamic surface control for a class of strict-feedback nonlinear systems with unknown backlash-like hysteresis, " in Proc. 2009 American Control Conference, St. Louis, MO, USA, 2009, pp. 4482-4487.
  • Related Articles

    [1]Zheng Chen, Shizhao Zhou, Chong Shen, Litong Lyu, Junhui Zhang, Bin Yao. Observer-Based Adaptive Robust Precision Motion Control of a Multi-Joint Hydraulic Manipulator[J]. IEEE/CAA Journal of Automatica Sinica, 2024, 11(5): 1213-1226. doi: 10.1109/JAS.2024.124209
    [2]Shengnan Gao, Zhouhua Peng, Haoliang Wang, Lu Liu, Dan Wang. Long Duration Coverage Control of Multiple Robotic Surface Vehicles Under Battery Energy Constraints[J]. IEEE/CAA Journal of Automatica Sinica, 2024, 11(7): 1695-1698. doi: 10.1109/JAS.2023.123438
    [3]Xiongbo Wan, Chaoling Zhang, Fan Wei, Chuan-Ke Zhang, Min Wu. Hybrid Dynamic Variables-Dependent Event-Triggered Fuzzy Model Predictive Control[J]. IEEE/CAA Journal of Automatica Sinica, 2024, 11(3): 723-733. doi: 10.1109/JAS.2023.123957
    [4]Dahui Luo, Yujuan Wang, Yongduan Song. Practical Prescribed Time Tracking Control With Bounded Time-Varying Gain Under Non-Vanishing Uncertainties[J]. IEEE/CAA Journal of Automatica Sinica, 2024, 11(1): 219-230. doi: 10.1109/JAS.2023.123738
    [5]Xiangqian Yao, Hao Sun, Zhijia Zhao, Yu Liu. Event-Triggered Bipartite Consensus Tracking and Vibration Control of Flexible Timoshenko Manipulators Under Time-Varying Actuator Faults[J]. IEEE/CAA Journal of Automatica Sinica, 2024, 11(5): 1190-1201. doi: 10.1109/JAS.2024.124266
    [6]Wei Qian, Yanmin Wu, Bo Shen. Novel Adaptive Memory Event-Triggered-Based Fuzzy Robust Control for Nonlinear Networked Systems via the Differential Evolution Algorithm[J]. IEEE/CAA Journal of Automatica Sinica, 2024, 11(8): 1836-1848. doi: 10.1109/JAS.2024.124419
    [7]Yunjun Zheng, Jinchuan Zheng, Ke Shao, Han Zhao, Hao Xie, Hai Wang. Adaptive Trajectory Tracking Control for Nonholonomic Wheeled Mobile Robots: A Barrier Function Sliding Mode Approach[J]. IEEE/CAA Journal of Automatica Sinica, 2024, 11(4): 1007-1021. doi: 10.1109/JAS.2023.124002
    [8]Ashish Kumar Jain, Shubhendu Bhasin. Tracking Control of Uncertain Nonlinear Systems With Unknown Constant Input Delay[J]. IEEE/CAA Journal of Automatica Sinica, 2020, 7(2): 420-425. doi: 10.1109/JAS.2019.1911807
    [9]Jingwei Lu, Qinglai Wei, Fei-Yue Wang. Parallel Control for Optimal Tracking via Adaptive Dynamic Programming[J]. IEEE/CAA Journal of Automatica Sinica, 2020, 7(6): 1662-1674. doi: 10.1109/JAS.2020.1003426
    [10]Khatir Khettab, Samir Ladaci, Yassine Bensafia. Fuzzy Adaptive Control of a Fractional Order Chaotic System With Unknown Control Gain Sign Using a Fractional Order Nussbaum Gain[J]. IEEE/CAA Journal of Automatica Sinica, 2019, 6(3): 816-823. doi: 10.1109/JAS.2016.7510169
    [11]Ping Sun, Zhuang Yu. Tracking Control for a Cushion Robot Based on Fuzzy Path Planning With Safe Angular Velocity[J]. IEEE/CAA Journal of Automatica Sinica, 2017, 4(4): 610-619. doi: 10.1109/JAS.2017.7510607
    [12]Longsheng Chen, Qi Wang. Adaptive Robust Control for a Class of Uncertain MIMO Non-affine Nonlinear Systems[J]. IEEE/CAA Journal of Automatica Sinica, 2016, 3(1): 105-112.
    [13]Shanwei Su. Output-feedback Dynamic Surface Control for a Class of Nonlinear Non-minimum Phase Systems[J]. IEEE/CAA Journal of Automatica Sinica, 2016, 3(1): 96-104.
    [14]Bin Xian, Jianchuan Guo, Yao Zhang. Adaptive Backstepping Tracking Control of a 6-DOF Unmanned Helicopter[J]. IEEE/CAA Journal of Automatica Sinica, 2015, 2(1): 19-24.
    [15]Delong Hou, Qing Wang, Chaoyang Dong. Output Feedback Dynamic Surface Controller Design for Airbreathing Hypersonic Flight Vehicle[J]. IEEE/CAA Journal of Automatica Sinica, 2015, 2(2): 186-197.
    [16]Yanlong Zhou, Mou Chen, Changsheng Jiang. Robust Tracking Control of Uncertain MIMO Nonlinear Systems with Application to UAVs[J]. IEEE/CAA Journal of Automatica Sinica, 2015, 2(1): 25-32.
    [17]Jing Na, Guido Herrmann. Online Adaptive Approximate Optimal Tracking Control with Simplified Dual Approximation Structure for Continuous-time Unknown Nonlinear Systems[J]. IEEE/CAA Journal of Automatica Sinica, 2014, 1(4): 412-422.
    [18]Mark J. Balas, Susan A. Frost. Robust Adaptive Model Tracking for Distributed Parameter Control of Linear Infinite-dimensional Systems in Hilbert Space[J]. IEEE/CAA Journal of Automatica Sinica, 2014, 1(3): 294-301.
    [19]Xiaoming Sun, Shuzhi Sam Ge. Adaptive Neural Region Tracking Control of Multi-fully Actuated Ocean Surface Vessels[J]. IEEE/CAA Journal of Automatica Sinica, 2014, 1(1): 77-83.
    [20]Ruimin Zhang, Lu Dong, Changyin Sun. Adaptive Nonsingular Terminal Sliding Mode Control Design for Near Space Hypersonic Vehicles[J]. IEEE/CAA Journal of Automatica Sinica, 2014, 1(2): 155-161.

Catalog

    通讯作者: 陈斌, bchen63@163.com
    • 1. 

      沈阳化工大学材料科学与工程学院 沈阳 110142

    1. 本站搜索
    2. 百度学术搜索
    3. 万方数据库搜索
    4. CNKI搜索

    Figures(7)  / Tables(3)

    Article Metrics

    Article views (2142) PDF downloads(118) Cited by()

    /

    DownLoad:  Full-Size Img  PowerPoint
    Return
    Return