Neural networks for identification and control of smart structures

The application of neural network technology for mathematical modeling and robust control of experimental smart structural system is studied. Four smart structure test articles were designed and fabricated to incorporate Nickel Titanium Naval Ordinance Lab (NiTiNOL) and Lead Zirconate Titanate (PZT) actuators and strain gauge and Poly Vinyledene Fluoride (PVDF) film sensors. A neural network based technique to directly identify a state space model of a structural system with direct state measurement has been developed. For a more general case where only the output measurements are available, a feedforward neural network has been incorporated with the Eigensystem Realization Algorithm (ERA) to obtain a discrete time state space model. An adaptive learning rate algorithm and a selective training scheme have been developed to significantly reduce the training time and improve the error performance of the networks with large numbers of neurons in the input and hidden layers. A single chip implementation of neural network based robust controllers for smart structures has been successfully demonstrated for the first time using Intel's Electronically Trainable Analog Neural Network (ETANN) chip and the analog delay line chip by Tanner Research. Custom interface hardware required for this implementation has been developed. Finally, a neural network based optimizing controller scheme based on the minimization of a Linear Quadratic (LQ) performance index which can directly incorporate structural nonlinearities, all the a priori information about the system, such as control effort and bandwidth limits, and adaptation to the time varying dynamics has been developed. Both simulation and experimental results have been included to demonstrate the effectiveness of neural networks as a good tool in the identification and robust control implementation for smart structural systems.