Research On Intelligent Control For Nonlinear Systems Via Deterministic Learning

Due to the highly developed brain,humans are capable of easily and accurately responding to a variety of complex events.The human brain is commonly considered to be the most complex and intelligent system in nature,which is mainly manifested in:“learning by doing”,that is,learning is completed in the process of work,and the learned knowledge can be further memorized/stored in the brain;“learning for utilizing”,that is,the stored knowledge can be used proudly as experience in new tasks.Motivated by this,researchers have proposed advanced intelligent control methods,such as neural network control,to deal with the control problems of complex systems in dynamic environments.One of its main characteristics is that it has a sufficient human-like control strategy,knowledge about the controlled plant and environment,and the ability to use the knowledge.However,in practical applications,the problem of learning in uncertain dynamic environments is recognized as the most difficult issue in the field of adaptive and learning control.Meanwhile,the intelligent control problem with autonomous learning ability is also the long-term difficulty of artificial intelligence research in dynamic environments.In addition,as the production working environment becomes more and more complex and diversified,and higher requirements have been put forward for control quality.These practical problems undoubtedly pose more challenges to the development of intelligent control technology.Based on the deterministic learning theory and existing relevant studies,this thesis conducts research on the following aspects for the learning,decisionmaking,and control problems of complex nonlinear systems under unknown dynamic environments:1.The problem of adaptive neural network(NN)dynamic surface control and learning for a class of representative pure-feedback nonlinear systems is investigated.First,by using a dynamic surface control(DSC)technique and deterministic learning(DL)theory,the complicated derivation of virtual control is avoided during the process of controller design,and the partial persistency excitation(PE)condition of the RBF network is established using a recursive design.The proposed scheme not only ensures the stability of the system and the convergence of the tracking errors,but also achieves the locallyaccurate identification of the unknown dynamics of the closed-loop system.Subsequently,the learned knowledge is used to construct an experience-based controller,which effectively improves the control performance since the tremendous repeated training process of the estimated parameters is avoided.This scheme partially solves the problem of deterministic learning control of high-order pure-feedback systems,and the effectiveness of the proposed scheme is verified using simulations.2.The problem of observer-based adaptive NN control and learning for pure-feedback nonlinear systems is investigated.First,a system transformation technique is introduced,with an observer,a simple adaptive NN control strategy is proposed using only one NN approximator,and the stability and convergence of the tracking errors are ensured.In addition,the proposed scheme avoids the recursive analysis process in the learning process,accurate approximation/learning of the unknown closed-loop dynamics of the transformed system is achieved easily via DL.Then,by using the learned knowledge,a novel learning control scheme is further proposed,superior control performance is achieved and the computation burden is greatly reduced.The continuous stirred tank reactor application system is used to illustrate the effectiveness of the proposed scheme.3.A pattern-based intelligent control scheme is proposed for pure-feedback nonlinear systems with multiple control situations.First,based on DL,the designed adaptive dynamic surface controller is used to accurately identify the unknown closed-loop system dynamics under different control situations(patterns),and a pattern-based experience controller bank is constructed for all the control situations by utilizing the learned knowledge.Second,under the action of the current normal controller,the dynamics of the subsystems under different control situations are accurately identified via DL,and the obtained knowledge is used to construct a series of dynamical models to represent the accurate classification of different control situations.Then,rapid and accurate recognition of the changed control situation can be realized based on the dynamical pattern recognition strategy,and the correct experience-based controller is selected from the constructed controller bank,thus stability and superior control performance can be obtained.The simulation results show that the proposed scheme not only realizes the acquisition of knowledge like humans,but also can use the experiential knowledge to realize fast decision-making and high-performance control.4.The problem of pattern-based learning and control for pure-feedback nonlinear systems with prescribed performance constraints is investigated.First,by using the system transformation and error transformation techniques,an adaptive NN control scheme with prescribed performance is proposed.The proposed scheme not only achieves stability and tracking control with prescribed performance,that is,the system performance on both transient and steady-state stages are ensured,but also realizes the accurate identification/learning of the closed-loop dynamics of the transformed system during the stable adaptive control process.By utilizing the knowledge learned in different control situations,a pattern-based experience controller bank is constructed.Then,with the observer,under the action of the current normal experience-based controller,accurate identification of the transformed system dynamics under different control situations is obtained.Specifically,the proposed scheme avoids the learning/training of all subsystems of the original system,and the number of NN units to be trained is greatly reduced.Subsequently,by using the modeling knowledge,one set of dynamical estimators is constructed to accurately classify these different patterns.When the control situation changes suddenly,a series of recognition residuals are obtained by comparing the monitored system with the constructed estimators,and accurate recognition of the changed control situation is realized by the principle of minimum residuals.Based on the recognition result,the experience-based controller corresponding to the current control situation is selected to achieve stability and high-performance control,while the prescribed performance constraint is still guaranteed.Simulation results are given to verify the effectiveness of the proposed scheme.5.The problem of adaptive NN control and learning for a class of sampled-data affine nonlinear systems is investigated.The objective is to extend the deterministic learning control theory from continuous-time systems to sampled-data systems.To achieve learning,a novel adaptive NN control scheme is first designed,and the closed-loop stability and convergence of the tracking errors in a finite number of steps are guaranteed by the Lyapunov theory.In addition,a state transformation method is introduced to solve the learning difficulty caused by unknown affine terms.Locally-accurate approximation/learning of the sampling dynamics of the closed-loop system is achieved during the stable adaptive control process.Subsequently,using the acquired knowledge,a new sampling learning control scheme is further proposed,and superior control performance with faster convergence rate and smaller tracking error is achieved.Finally,simulations on a pendulum balancing system and a two-link manipulator system are implemented to illustrate the effectiveness of the proposed scheme.