A self-similar neural network for distributed vibration control

A self-similar neural network has been investigated for the control of flexible structures with unknown dynamics. The system model is approximated by a partially recurrent neural network. The network predicts the future state based on the current control command and a window of the past state. The controller, which is part of the system model, holds a copy of the system model. Thus the controller can compute the optimal control by using the estimation of all the future states within a look-ahead horizon. This ability to look ahead enables the controller to successfully attenuate vibrations generated by impulse, high frequency sine waves and random excitation. When an unfamiliar situation is encountered, the controller switches to exploration mode. During exploration, the learning rate is increased, and a low level of control is applied to stimulate the system without driving it out of bounds. Test cases show that the network resumed effective control after only a few seconds of exploration. This quick learning ability makes the controller suitable for systems having a wide range of operating conditions