Study On The Design And Optimization Of The Iterative Learning Control Algorithms

As a new subject presented in 1980's, the iterative learning control (ILC), with a great foreground in the control for the industry robot, numerical machine and other plants with repetition property, is feasible in dealing with the nonlinear systems, and un-modeling systems and so on. Certainly, as a new science branch, there are still many problems need to be improved and further studied in ILC.Designs of the iterative learning algorithms, the most important problems in the ILC, are also studied in this dissertation. At first, a kind of P-type causality iterative learning algorithm is presented and developed for linear discrete systems after analysis of the causality between the inputs and outputs of the dynamic systems and the deficiency of current iterative learning algorithm to this causality. The presented algorithm, in which the derivatives of the output errors are not needed, can reflect the causality between inputs and outputs rightly. The simulation also demonstrated that the developed algorithm has a better convergence trait compare with the current P-style iterative learning algorithms.Secondly, two kinds of optimal iterative learning algorithm (OILA) are investigated: 1) design of OILA based on the linear quadratic performance function in time domain, 2) design of OILA and guaranteed cost iterative learning algorithm based on the linear quadratic performance function in iteration domain for certain and uncertain systems respectively.To problem 1), the OILA and parameters estimate algorithms both for the certain and uncertain linear discrete systems are developed, and the conditions for the convergence also given in the text. Besides, the simulation results indicated that the OILA and parameters estimate algorithms are very effect, and an optimal performance or near optimal performance always achieved.To problem 2), a linear quadratic performance function in the iteration domain is defined in the text firstly, and then based on the new function an OILA for certain linear discrete systems and a guaranteed cost iterative learning algorithm for uncertain linear discrete systems are developed. One interesting property of these algorithms is that the iterative convergence speed can be controlled easily by adjusting the parameter matrices in the performance function. Meanwhile the guaranteed cost iterative learning algorithm is a LMI (linear matrix inequalities) based one, so it can be solved easily by using Matlab Toolbox.Comparing with its advantages, the deficiencies of ILC are evident too, especially that the structure and the parameters of the controlled plant should be repetitive strictly in alltrails. To boost up its robustness to the un-repetitive uncertainties, some robust ILC schemes for the uncertain rigid robot and constrained flexible links manipulators by combining the ILC with robust control methods are presented in this dissertation. At thesame time, some adaptive estimate methods for the bounded parameters of the un-repetitive uncertainties are used to reduce the conservation of the robust control, and the normal model of the controlled plant, which usually can obtain in practice, is also considered in the developed control to lessen the learning burden of ILC.The initial control inputs (the control inputs in the first trial) always choose as zeros or bound random vectors subjectively in the existing ILC. To avoid this unreasonableness, a "data-model" is established by using the control experience obtained in the control for other tasks. Then the initial control inputs can be gotten based on the data-model. In the text, the establishment of the "data-model", extraction of the relative data from this model and four methods, linearity weight, linear regression analysis, artificial neural networks approach and fuzzy decision, to get the initial control inputs are studied in detail. The systems can track the new desired trajectory with little error even at the first trial when the initial control inputs are choose by these methods, so the iterative learning times is reduced compare with the ILC in which the initial control inputs are choose subjectively under the same precision requirement.