چكيده به لاتين
In this thesis, constrained optimal sliding-mode control for a class of single input-single output dynamic nonlinear systems has been designed. The proposed control algorithm has been realized through decomposing the system dynamics into tow linear and nonlinear parts using projection- theory-based recurrent neural network. This approach has been developed with the aim of chattering avoidance, singularity avoidance, considering actuator limitations, minimizing control effort and closed-loop system robustness with respect to the model uncertainties. At the first step, the performance index has been defined based on the stabilization of the sliding surface. Then, minimization of the control effort and has been converted into quadratic programming problem in the presence of actuator limitations, where the weighting coefficients of the performance index and the sliding surface have been considered as the designing parameters. Next, the dynamic and algebraic models of the recurrent neural network have been determined. This network obtains the control signal online as the optimization variable, by solving a constrained quadratic programming problem. The stability analysis of the neural network and the closed-loop system have been performed using Lyapunov theorem. Suitable conditions have been obtained for designing parameters to insure the stability. The proposed approach satisfied the considered performance goals. However, the time varying changes of the control parameters may cause to undesired variations on the state variables and control signal for some systems. According to this and based on the presented principles for designing the constrained optimal sliding-mode control, an robust control strategy has been extended to determine the designing parameters offline. This has been done by formulating the dynamic of the closed-loop error and obtaining the robust error feedback gain. This approach, despite its advantages such as offline adjustment of the designing parameters, chattering-free responses, and obtaining constrained optimal control signal, is more complicated, has state-dependent constraint, and may exhibit feasibility solution problems for high-order dynamic systems. In order to improve the performance specifications and eliminating the sliding condition as a dynamic constraint, a new performance index is proposed based on the exponential reaching law. Moreover, an offline self-tuning process has been developed along with transient and steady-state analysis. This strategy despite having dominant aspects of the time-varying gain based method and the based adjustments, is only suitable for tracking step reference inputs and the complete state controllability condition for the controlled system. At the end, in order to generalize and adapt to the digital implementation, a new method based on singular value approach has been developed for stability analysis of discrete-time constrained optimal sliding-mode control using the recurrent neural network. In order to investigate the performance quality of the presented strategy, nonlinear systems such as inverse pendulum, continues stirred tank reactor, magnetic levitation system, and electro-hydraulic serve actuator have been used. Furthermore, the obtained results have been compared with newest advanced sliding-mode methods in recently developed method in literature. Simulating results show very good performance of the proposed approaches.
