Adaptive feedback-feedforward control and sensor fault accommodation via neural networks for seismically loaded infrastructures

This dissertation studies a new combined feedback-feedforward control algorithm for structural engineering applications. The controller contains both feedback and feedforward components. The feedback component is assumed to be the same as that found from traditional LQR design. The feedforward component is obtained by estimating the external excitation as series of step functions at each time increment. This feedforward gain varies with the duration of the step function used for estimation and converges as the time duration increases. Thus, a finite number of pre-calculated gains can be used to represent the potential feedforward gain profiles. At any instant in time, the excitation is measured and by using the past measurements, the most effective feedforward gain for the recorded excitation values can be selected from the set of pre-calculated gains. This value is used as the feedforward gain for the current time step. Numerical examples are presented to show the effectiveness of this adaptive control scheme. The effects of varying the control objectives, the updating time for the feedforward gain, and the number and location of actuators are studied. Some practical issues such as sampled-data design, output feedback, actuator dynamics and time delays are also considered.; In order to address the possibility of sensor failures, Fault Detection Neural Networks (FDNNs) and Fault Accommodation Neural Networks (FANNs) have been developed in previous work. In the present work, alternative architectures are used to improve the performance of the neural networks. In particular, more highly integrated neural networks are examined for the FANNs. Once the networks have been trained, their effectiveness in an integrated control scheme is tested during simulations of seismic events using excitation data from actual earthquakes. The results of numerical studies show that the use of these improved neural networks for sensor fault accommodation makes improvements in the structure's controlled response possible.