Font Size: a A A

Study On Pattern-Based NN Control Of Nonlinear Systems

Posted on:2016-04-29Degree:DoctorType:Dissertation
Country:ChinaCandidate:F F YangFull Text:PDF
GTID:1108330479495115Subject:Control theory and control engineering
Abstract/Summary:PDF Full Text Request
The combination of pattern recognition and control has drawn a lot of interests recently. Studies have shown that the research on pattern-based learning and control will be useful to many areas, such as motion control of a robot, or security assessment and control of power systems. For a long time, scientists have been intrigued by the complexity of human movements and the ease with which people move. It is observed that with su?cient practice human can learn many highly complicated control tasks, and these tasks can be performed by a proficient individual with little cognitive e?ort. A recent interesting development in this field is to study human movement via a dynamic systems approach, which exhibits features of pattern-forming dynamical systems. It is shown by experiments that the control and coordination of human movements at all levels is associated with dynamic patterns.It is thus suggested that mechanisms of pattern-based learning and control may be responsible for the proficiency of complicated human control skills. This pattern-based learning and control abilities, however, have been less studied by the control community.Such abilities require a rigorous definition of dynamic patterns, and solutions to problems of e?ective representation, rapid recognition and classification of dynamical patterns.These problems, nevertheless, are di?cult to be solved in pattern recognition area. For example, since time could not be e?ciently considered as a static attribute, the problem of how to appropriately represent the time-varying patterns in a time-independent way is a very di?cult and fundamental problem. Recently, a newly developed deterministic learning(DL) theory was proposed to cope with representation of time-varying dynamical pattern in a time-invariant and spatially-distributed manner by using the locally-accurate radial basis function(RBF) approximation of system dynamics underlying the dynamical pattern. Similarity of dynamical patterns is characterized by comparison of the system dynamics inherently within these dynamical patterns. A mechanism for rapid recognition of dynamical patterns is presented, by which a test dynamical pattern is recognized as similar to a training dynamical pattern if state estimation or synchronization is achieved according to a kind of internal and dynamical matching on system dynamics.It is seen that the above structure of pattern-based control is closely related to multiple model adaptive control(MMAC). In the last two decades, multiple models switching control approach has been regarded as one of the powerful methods to improve the performance of linear and nonlinear systems. The objective in multiple models controller is to determine the most appropriate model at any instant, using a suitable performance criterion based on the identification error, and consequently to activate the corresponding controller. It is shown that multiple models can be used to detect changes in the environment and initiate appropriate action, also, to combine the advantages of di?erent controllers and achieve both stability and improved performance. However, classical MMAC methods have su?ered some drawbacks, such as, studies restricted to linear system, requiring exponentially high number of fixed models, frequent switching avoided by a admissible switching sequence, switching resulting in discontinuous control signals. In this thesis, based on DL theory, we study a novel multiple model control method for nonlinear systems, i.e.pattern-based NN control method, which is used to deal with control in multiple environments.This thesis contains three parts. The first part investigates the NN controller design and stability analysis of space manipulator systems based on reference patterns(patterns in command space). The second part extends pattern-based control of simple closed-loop nonlinear systems using the average dwell time(ADT) method. The third parts studies pattern-based control of a more general class of nonlinear systems(patterns in parameter space).First, we investigate control of space manipulator systems based on reference patterns. By defining a tracking control task as a reference pattern, identification of which is achieved in a local region via DL. Likewise, identification of the local controlled manipulator system dynamics corresponding to each reference pattern is also realized. Then a set of pattern-based constant NN controllers are constructed accordingly by using the obtained manipulator system dynamics. When tracking control task begins to change, rapid recognition of reference pattern is naturally implemented due to the internal matching on reference system dynamics, then the corresponding NN controller with learned experience is selected and activated. Research results show that the pattern-based NN controllers can guarantee the stability of the pattern-based space manipulator systems and improve the control performance.Second, we investigate control of a class of simple nonlinear system based on closedloop patterns. It is shown that by adopting a class of switching signals with average dwell time(ADT) property, the NN learning controller can achieve small tracking errors and fast convergence rate with small control gains. The minimal ADT is obtained by multiple lyapunov function method.Third, we investigate pattern-based control of a more general class of nonlinear systems in the Brunovsky form, in which the a?ne term is an unknown function of system states. The class of systems considered is subjected to possible large and abrupt parameter or dynamics changes which may yield multiple Brunovsky systems. The existence of the unknown a?ne term makes the recognition of the closed-loop control situations as well as stability analysis of patter-based control di?cult tasks. To overcome the di?culties,firstly, in the identification phase, when the system is in di?erent control situations(either in normal or abnormal conditions), adaptive NN controllers are designed to achieve closed-loop stability and tracking performance, and the closed-loop control system dynamics are locally-accurately identified via deterministic learning. The identified control system dynamics are stored by constant RBF networks, and a set of constant NN controllers are constructed by using the obtained constant RBF networks. Secondly, still in the phase of identification but for di?erent control situations, the plant is controlled by the normal constant NN controller. When the system is operated under di?erent or abnormal conditions, the underlying system dynamics are identified via deterministic learning.The identified abnormal system dynamics under normal control are stored in constant RBF networks. Thirdly, in the phase of recognition of di?erent control situations, a set of estimators are constructed by using the identified abnormal system dynamics under normal control. When one identified control situation(i.e. a test closed-loop dynamical pattern) recurs, by using the constructed estimators, the recurred control situation will be rapidly recognized via dynamical pattern recognition. Finally, in the phase of patternbased control, select the corresponding constant NN controller based on the result of rapid recognition. A common lyapunov function is used as an e?cient analysis tool for pattern-based control, and both stability and improved performance can be achieved.Compared with MMAC for nonlinear systems, the proposed pattern-based control method has the following features:(i) locally-accurate identification of the multiple models of closed-loop dynamics is achieved, this reduces the number of models unnecessarily used in MMAC;(ii) the rapid recognition of a recurred control situation yields the selection of the correct(better than the closest) model to the plant, since the recognition of control situations is based on the matching of the underlying system dynamics;(iii)switching to the correct model leads naturally to guaranteed stability and improved control performance.The result presented in this thesis show that pattern-based control will provide a new approach for fast decision and control in dynamic environments.
Keywords/Search Tags:Pattern-based control, Nonlinear system, Deterministic learning, Rapid recognition, RBF network
PDF Full Text Request
Related items