Elsevier

Neurocomputing

Volume 117, 6 October 2013, Pages 138-149
Neurocomputing

Self-structuring fuzzy-neural backstepping control with a B-spline-based compensator

https://doi.org/10.1016/j.neucom.2013.01.017Get rights and content

Abstract

To relieve the burdens of network controller design and approximation error bound determination, a self-structuring fuzzy-neural backstepping control system (SSFNBS) with a B-spline-based compensator is proposed. In this paper, a network-identification-based control is represented where the self-structuring fuzzy neural network-based (SSFNN) is used as the observer to approximate the controlled system dynamics. To balance the tradeoff between the structure efficiency and the identification accuracy, a structure learning mechanism of the node-adding process and the node-pruning process is introduced. On the other hand, the B-spline-based compensator is introduced to dispel the effect of approximation error. With the adoption of B-spline functions, the compensation controller can be given in a recurrent way based on the introduction of knot vector and the drawbacks of the conventional compensation controllers can be freed. With the introduction of the B-spline function, the proposed SSFNBS features the following advantages: (1) the capability of network-based controller is improved, (2) the design of the compensation controller can be easily established based on the characteristics of the B-spline function, (3) the stability of closed-loop control system is guaranteed by the means of Lyapunov function with the tuning law of the B-spline-based compensator. To investigate the capabilities of the proposed approach, the SSFNBS is applied to the nonlinear system, chaotic system, and wing rock motion control problems. Through the simulation results the advantages of the proposed SSFNBS can be observed.

Introduction

Over the past two decades, neural networks-based (NNs) control has attracted growing popularity as an adaptive control approach for engineering problems [1], [2], [3]. With the property of approximating nonlinearity to arbitrary accuracy by sufficient hidden neurons and learning, the NN-based controller design has been applied to cope with the increasing demands for handling complexity, uncertainty, or unknown dynamics. Based on the interactions within layers, the NN-based method has widely been recognized as a powerful tool for industrial applications [1], [4], [5], [6]. The characteristics of approximation, parallelism, and adaptation also imply their potentials for real-time machine intelligence realization [4], [5], [6], [7]. Recently, the alternative of fuzzy neural network (FNN) aroused much interest for the enhancement of network performance. With the proved ability of universal approximation [8], [9], FNN is regarded as a complementary approach and possesses the advantages from low-level learning and computation to high-level human-like reasoning. The property of linguistic expression also enables the control strategy being formed in a rule-based manner which is derived from observation and expert knowledge. With the fuzzified nodes embedded in the layers, the learning abilities of network can be upgraded [10], [11], [12].

Although it is widely accepted that the parameterized neural network is capable of approximating linear or nonlinear mapping by adequately choosing network structures and training methods, however, a challenging problem for designers is to select an appropriate structure for balancing the number of rules and the approximation accuracy. If the network size is chosen too small, it is impossible to assure the approximation to converge to an acceptable level due to the limited nodes. On the other hand, if over-determined nodes are given, the computational burden is huge and the waste of computation resource implies its impracticality for real-world applications. To tackle the problem of structure optimization, the self-organizing (or self-learning) methods have aroused much interest [12], [13], [14], [15], [16], [17], [18], [19], [20], [21]. For those approaches, the increasing rule construction concept was the first proposed [13], [14], [15], [16], where the new node is generated when the current structure fails to cover the input space. Park et al. [17] presented a self-structuring neural network, where the over-activated nodes in the parameter learning will be split into two separated nodes to cover the input space as evenly as possible. The similar concept was also adopted in [12]. However, the increasing node evolution implies the redundancy since it does not remove the ineffective nodes. In order to find the optimal or sub-optimal rule structure through rule adding and rule pruning procedures, Wu et al. proposed a dynamic mechanism for adjusting the nodes dynamically for giving the appropriate input space interpretation [18], [19] and other similar approaches were also proposed [20], [21]. However, the heavy computational load for node significance determination is unavoidable. In [13], [14], [15], [16], [21], the discussion of control stability during system evolution was not provided.

From the control point of view, if the exact model of the controlled system is well-defined there will be an ideal control scheme to achieve favorable performance by canceling the system dynamics. However, the ideal model is difficult to obtain with the involvements of internal parameters variation and external disturbances for real applications. To compensate for the approximation imperfection, the compensation mechanism such as the chattering control and robust compensation are introduced. However, the chattering phenomenon of the sliding mode control strategy is unfavorable. To reduce the adverse effect, the saturation function was proposed but it loses the guarantee of system stability. In the past decades, an approximation error bound estimation mechanism was proposed to estimate the bound of approximation error so that the chattering phenomenon of the control effort can be reduced [1]. However, the adaptive law for estimating the error bound tends to make the bound going to infinity if the main controller is ill-designed. For the NN-based control collaborates with the robust compensation control [10], the control performance is subject to the predetermined attenuation level. If an inappropriate attenuation level is given, the control effort may lead to undesirable large signal which is hard for realization.

Backstepping control is known as a systematic and recursive design methodology with the flexibility to avoid cancellations of useful nonlinearities [22], [23], [24], [25], [26]. It provides an alternative feedback design in terms of pseudo-control, where the appropriate state and virtual control are derived from the state variables in a recursive way to form smaller subsystems. Based on this manipulation, the final Lyapunov functions are represented by these subsystems associated with each individual pseudo-control stage [27]. The rapid development of backstepping design was also witnessed where the backstepping method is combined with other methods such as sliding mode control, robust control, fuzzy control, neural network, and so on [28], [29], [30], [31]. For these approaches, the backstepping-based control method provides the flexibility to accommodate unmodeled nonlinearity, where the asymptotic stability and robustness to some unmatched uncertainties can be guaranteed.

In this paper, a self-structuring fuzzy-neural backstepping control system (SSFNBS) is proposed. The SSFNBS comprises two elements: the self-structuring fuzzy neural network-based (SSFNN) backstepping controller and the B-spline-based compensation controller. Based on the backstepping control theory, the SSFNN-based observer is exploited as the observer to approximate the controlled system dynamics and the B-spline-based compensation controller is used to compensate for the approximation error caused by SSFNN. To facilitate the network design, a structure learning mechanism is introduced, where the node-adding and node-pruning are involved to network structure determination. In this paper, a new B-spline-based compensator is also introduced. With the recurrence operation of the given knot vector and control points, the generated B-spline functions are exploited to generate the compensation signal. With the adoption of B-spline concept and control point adjusting mechanism, the drawbacks of the conventional compensation controllers can be freed. Based on the evolution mechanism, the structure learning and the parameter estimation can be conducted simultaneously. In the SSFNBS, the parameters are evolved by the means of the Lyapunov function to ensure the system stability. Thus, an intelligent control approach is presented. To validate the capabilities of the proposed approach, the SSFNBS is applied to several control problems. Through the simulation results the advantages of the proposed SSFNBS can be observed.

Section snippets

Problem formulation

For simplicity, the general form of a second order nonlinear system is considered in this paper:ẍ=f(x)+u,where x=[xẋ]T is the state vector of the system, which is assumed to be available for measurement, f(x) is the unknown continuous functions, and u is the input. The control objective is to find a control law so that the state trajectory x can track a trajectory command xc closely. To achieve successful backstepping control, some appropriate functions of stale variables are selected as

Fuzzy neural network observer

The fuzzy neural network-based observer is shown in Fig. 1. Layer 1 accepts the input variables. Layer 2 is used to calculate the membership values, where the Gaussian function is embedded in this layer as the membership function. The nodes of layer 3 are regarded as the fuzzy rules. The links before layer 3 represent the premise. Layer 4 is the output layer, where the links of layer 4 represent the consequence and the node in this layer is the output of the fuzzy neural network. Based on the

Simulation results

In this section, the proposed SSFNBS is applied to nonlinear system, chaotic system and wing rock motion control problems to verify its effectiveness. It should be emphasized that the development of the SSFNBS does not need to know the system dynamics of the controlled system. For practical implementation, the structure and parameters of SSFNN can be online tuned by the proposed methods to estimate the controlled system dynamics. In the simulation, the parameters for SSFNBS are selected with n

Conclusions

The determination of the network size is difficult because the designed structure varies from the experience of designers. The consequential approximation imperfection of an ill-designed controller also increases the designing difficulty of the compensation controller. In this paper, the developed SSFNN observer is capable of dealing with the tradeoff between the approximation accuracy and computation complexity. With the proposed B-spline-based compensator, the disadvantages of the traditional

Acknowledgment

The author appreciates the financial support from the National Science Council of Republic of China under grant NSC 100–2917-I-564-017. The authors would like to express their gratitude to the Reviewers for their valuable comments and suggestions.

Kuo-Hsiang Cheng was born in Taipei, Taiwan, ROC, in 1978. He received B.S. degree in automatic control engineering from the Feng Chia University and the Ph.D. degree in electrical engineering from the Chang Gung University, Taiwan, in 2000 and 2006, respectively. During 2006–2011, he was a researcher of the Mechanical and Systems Research Laboratories (MSL), Industrial Technology Research Institute (ITRI), Hsinchu, Taiwan, ROC. In 2011, he received the grant from National Science Council (NSC)

References (44)

  • S. Labiod et al.

    Adaptive fuzzy control of a class of SISO nonaffine nonlinear systems

    Fuzzy Sets Syst.

    (2007)
  • C.M. Lin et al.

    Neural-network-based adaptive control for induction servomotor drive system

    IEEE Trans. Ind. Electron.

    (2002)
  • O. Omidvar et al.

    Neural Systems for Control

    (1997)
  • Z. Man et al.

    An adaptive tracking controller using neural networks for a class of nonlinear systems

    IEEE Trans. Neural Networks

    (1998)
  • M.U. Polycarpou et al.

    Stable adaptive tracking of uncertain systems using nonlinearly parameterized on-line approximators

    Int. J. Control

    (1998)
  • F.-J. Lin et al.

    Robust control using neural network uncertainty observer for linear induction motor servo drive

    IEEE Trans. Ind. Electron.

    (2002)
  • L.X. Wang et al.

    Fuzzy basis functions, universal approximation, and orthogonal least squares learning

    IEEE Trans. Neural Networks

    (1992)
  • W.Y. Wang et al.

    H tracking-based sliding mode control for uncertain nonlinear systems via an adaptive fuzzy-neural approach

    IEEE Trans. Syst. Man Cybern. B: Cybern.

    (2002)
  • C. Li et al.

    Pseudo-error-based self-organizing neuro-fuzzy system

    IEEE Trans. Fuzzy Syst.

    (2004)
  • C.F. Juang et al.

    An on-line self-constructing neural fuzzy inference network and its applications

    IEEE Trans. Fuzzy Syst.

    (1998)
  • F.J. Lin et al.

    Self-constructing fuzzy neural network speed controller for permanent-magnet synchronous motor drive

    IEEE Trans. Fuzzy Syst.

    (2001)
  • F.J. Lin et al.

    A permanent-magnet synchronous motor servo drive using self-constructing fuzzy neural network controller

    IEEE Trans. Energy Convers.

    (2004)
  • Cited by (11)

    • Intelligent dynamic sliding-mode neural control using recurrent perturbation fuzzy neural networks

      2016, Neurocomputing
      Citation Excerpt :

      But, the control gain is not usually constant in many physical systems. Type-1 fuzzy neural network (T1FNN) has been studied extensively for over 30 years [15–20]. However, T1FNN is unable to directly handle the rule uncertainties due to the fact that the value of the membership function is a crisp number.

    • Adaptive backstepping Elman-based neural control for unknown nonlinear systems

      2014, Neurocomputing
      Citation Excerpt :

      Based on the terms of appropriate state and virtual control, the final Lyapunov function is represented by these subsystems associated with each individual pseudo-control stage. Recently, many intelligent backstepping controllers have been proposed to control nonlinear systems [27–30]. In [27,29], an adaptive fuzzy backstepping output feedback control approach is proposed.

    • Neural network based adaptive dynamic surface control for cooperative path following of marine surface vehicles via state and output feedback

      2014, Neurocomputing
      Citation Excerpt :

      However, such a technique has not been explored for cooperative path following problem of marine surface vehicles. It is worthwhile to mention that the cooperative path following controllers developed in [2,9,11,13,14,20] are designed based on a backstepping technique [30–33,36–39]. A major drawback in the traditional backstepping design is the problem of computational complexity, which is caused by the repeated differentiations of virtual controls.

    • Intelligent control of chaotic systems via self-organizing Hermite-polynomial-based neural network

      2014, Neurocomputing
      Citation Excerpt :

      But, this method often suffers from a heavy computation loading. In [15–18], a self-structuring approach is applied into a fuzzy neural network. When an input signal is too far from the current clusters, a new membership function and a new fuzzy rule will be generated automatically.

    View all citing articles on Scopus

    Kuo-Hsiang Cheng was born in Taipei, Taiwan, ROC, in 1978. He received B.S. degree in automatic control engineering from the Feng Chia University and the Ph.D. degree in electrical engineering from the Chang Gung University, Taiwan, in 2000 and 2006, respectively. During 2006–2011, he was a researcher of the Mechanical and Systems Research Laboratories (MSL), Industrial Technology Research Institute (ITRI), Hsinchu, Taiwan, ROC. In 2011, he received the grant from National Science Council (NSC) and currently he is conducting his postdoctoral research in the Department of Mechanical Engineering, University of California, Berkeley. His research interests include fuzzy logic, neural networks, intelligent systems and control, and intelligent vehicles.

    View full text