An iterative learning control technique has been proposed as an alternative for controlling plants that are difficult to model and control due to uncertainties including variable loads, nonlinearities and so forth. With less a priori information about the target plant, the learning controller performs better as it carries out the same task repetitively. So far, the method has been successfully applied for robotic repetitive tasks such as painting, pick-and-place, etc.
In this thesis, it is reported that a good example which reveals merits of the learning controller may be found in the control of optical beam to be focused on the desired track of the disk surface in the Magneto Optical Disk Drive (MODD) system. The contour of the surface of the rotating disk becomes the reference input in the MODD system. The objective of control is to make the system to effectively follow the repetitive reference generated by unflatnesses of the disk surface and axial runout of the rotating motor.
The conventional iterative learning algorithm utilizes the derivative of error, as is usually the case, which is rather sensitive to external disturbances. Also, the conventional iterative learning algorithm is not adequate for the system where the initial condition exists and/or the final value of current iteration becomes the initial condition of next iteration. Therefore a new learning control algorithm with feedback loop is suggested in this thesis.
The suggested learning control algorithm utilizes not only the control input and the error obtained at previous iteration but also the error of current iteration. Since it composes a closed loop, it complements the weakness of being too sensitive to the disturbances. By using the learning filter, it can reject high frequency terms in the control input which may cause unstable oscillation.
The designed learning controller is applied for focusing control of the MODD system, and it is found that the error is effectively reduced. In our experimental result, the peak-to-peak value of the focusing error is shown to be half of that achieved with an conventional PID controller.
Also, the method proves better performance compared with the conventional type controller even when the gain of the error detector is changed from the nominal model value. This means that the designed learning controller is robust to the variation of gain of the error detector.