Nonlinear system identification and learning control of flexible arm with one degree of freedom using neural networks

This paper studies control methods of a flexible arm with one degree of freedom in order to suppress both vibration due to step movement and residual vibration due to signal noise. When using the conventional PD control methods, these two problems become adverse problems. Namely, the residual vibration becomes large when high gain feedback is used to improve response during suppressing the vibration of the step movement. Moreover, this system becomes a nonlinear system because the deformation of the flexible arm is large by use of soft elastic material. Neural Network (NN) is used as both a simulator and a controller with a view of nonlinear and learning ability. In this study, On-line learning of NN is impossible as the flexible arm breaks by fatigue on many repetitions of learning. Therefore, Off-line learning is employed and the NN simulator is used because of the nonlinear mapping ability for Neuro-controller training. As to the Neuro-controller training, repetitive learning method is proposed, which gives excellent performance of the Neuro-controller by improving both the simulator and controller in turn. This proposed method can give excellent results for the adverse problem of suppressing both the vibration of due to the step movement and the residual vibration due to the signal noise.