Adaptive Control for Nonlinear Pure-Feedback Systems With High-Order Sliding Mode Observer

Abstract

Most of the available control schemes for pure-feedback systems are derived based on the backstepping technique. On the contrary, this paper presents a novel adaptive control design for nonlinear pure-feedback systems without using backstepping. By introducing a set of alternative state variables and the corresponding transform, state-feedback control of the pure-feedback system can be viewed as output-feedback control of a canonical system. Consequently, backstepping is not necessary and the previously encountered explosion of complexity and circular issue are also circumvented. To estimate unknown states of the newly derived canonical system, a high-order sliding mode observer is adopted, for which finite-time observer error convergence is guaranteed. Two adaptive neural controllers are then proposed to achieve tracking control. In the first scheme, a robust term is introduced to account for the neural approximation error. In the second scheme, a novel neural network with only a scalar weight updated online is constructed to further reduce the computational costs. The closed-loop stability and the convergence of the tracking error to a small compact set around zero are all proved. Comparative simulation and practical experiments on a servo motor system are included to verify the reliability and effectiveness.

Command filter‐based adaptive neural network controller design for uncertain nonsmooth nonlinear systems with output constraint

“…$$$ {\varrho}_1 $$$ and $$$ {\varrho}_2 $$$ are positive constants. According to References 40 and 41 , we have the following Lemmas .…”

Adaptive super‐twisting control of underwater intervention system considering dynamic couplings and uncertainties

“…To drive the estimation error to the origin, the adaptive law of the RBFNN proposed in Equation (51) is designed as: $$$$\begin{equation} \dot{\hat{\bm {W}}}=-\bm {\Gamma }\bm {\Lambda }|\bm {\zeta }_{i}|\hat{\bm {W}}, \end{equation}$$$$where $$\bm {\Gamma }=diag(\Gamma _{1},\Gamma _{2},\ldots,\Gamma _{n}), \bm {\Lambda }=diag(\Lambda _{1},\Lambda _{2},\ldots,\Lambda _{n})$$ are positive definite diagonal matrices, representing the gain matrix and a matrix with small positive constants, respectively. Lemma [ 52 ] Considering the adaptive law Equation (52), the estimation of RBFNN weight $$\hat{\bm {W}}$$ is bounded by $$\Vert \hat{\bm {W}}_{i}\Vert \le \frac{k_{\Phi i}}{\Lambda _{i}}$$, where $$k_{\Phi i}$$ denotes the boundary of the Gaussian function vector, that is, $$\Vert \Phi _{i}(\bm {h})\Vert \le k_{\Phi i}$$ …”

Double integral‐enhanced Zeroing neural network with linear noise rejection for time‐varying matrix inverse

“…Differing from the numerical methods, the neural network methods can effectively implement parallel computing and has hardware realisability [14][15][16]. So they have been extensively studied and proposed for mathematical applications [17][18][19], such as static matrix inversion problems [20].…”