A new learning algorithm for neural network state estimation in active vibration control

Abstract

The authors present a numerical study related to the active control of the vibrations of a cantilevered beam with piezoelectric actuators using the modified independent modal space control (MIMSC) algorithm coupled with a neural network (NN) for state estimation. Among other control strategies the MIMSC algorithm has been shown to have excellent closed-loop structural damping. Such an algorithm requires as input data the modal displacements and velocities of the vibrating beam. A neural network, trained with an extended backpropagation learning algorithm, is proposed as an alternative approach to classic estimation structure for the on-line estimation of the modal parameters. This particular modification of the backpropagation (BP) algorithm introduces heterogeneous processing elements for the hidden and output layers. The results of a numerical study using the standard and extended BP algorithms are presented and discussed.